From 4692052ee353c7f4618ca701f8dbf9c1672e2488 Mon Sep 17 00:00:00 2001 From: lin-bot23 Date: Thu, 5 Nov 2026 18:06:50 +0800 Subject: [PATCH 1/7] docs: cover the Sol-H3-VAE decode trade and the Lightning LoRA on image edits --- es/tutorials/image/qwen/qwen-image-2-1.mdx | 8 ++++-- .../video/minimax/minimax-h3-native.mdx | 27 ++++++++++++++++--- ja/tutorials/image/qwen/qwen-image-2-1.mdx | 8 ++++-- .../video/minimax/minimax-h3-native.mdx | 27 ++++++++++++++++--- ko/tutorials/image/qwen/qwen-image-2-1.mdx | 8 ++++-- .../video/minimax/minimax-h3-native.mdx | 27 ++++++++++++++++--- tutorials/image/qwen/qwen-image-2-1.mdx | 4 +++ tutorials/video/minimax/minimax-h3-native.mdx | 21 ++++++++++++++- zh/tutorials/image/qwen/qwen-image-2-1.mdx | 8 ++++-- .../video/minimax/minimax-h3-native.mdx | 27 ++++++++++++++++--- 10 files changed, 140 insertions(+), 25 deletions(-) diff --git a/es/tutorials/image/qwen/qwen-image-2-1.mdx b/es/tutorials/image/qwen/qwen-image-2-1.mdx index 9f52abd799..1ade8841e3 100644 --- a/es/tutorials/image/qwen/qwen-image-2-1.mdx +++ b/es/tutorials/image/qwen/qwen-image-2-1.mdx @@ -2,7 +2,7 @@ title: "Qwen-Image-2.1 en ComfyUI: texto a imagen, edición y salida 2K" description: "Un solo checkpoint de Qwen-Image-2.1 cubre texto a imagen y edición por instrucciones en ComfyUI, con salida 2K nativa, canal alfa y hasta 16 imágenes de referencia." sidebarTitle: "Qwen-Image-2.1" -translationSourceHash: 13a6434f +translationSourceHash: 73760f36 translationFrom: tutorials/image/qwen/qwen-image-2-1.mdx translationBlockHashes: "_intro": 1aac4055 @@ -11,7 +11,7 @@ translationBlockHashes: "Workflow settings": 1f64eacd "VAE decoding artifacts": 8c94f042 "Faster sampling with the Turbo LoRA": 7c22e331 - "Faster sampling with the Lightning LoRA": fa73c3a1 + "Faster sampling with the Lightning LoRA": 3fc46534 "Alpha channel output": d52c962c --- import UpdateReminder from "/snippets/es/tutorials/update-reminder.mdx" @@ -281,6 +281,10 @@ El LoRA funciona con `cfg` 1 y una programación de cuatro pasos. Su tarjeta lis Un nodo `ManualSigmas` está marcado como experimental y pasa sus números sin cambios, así que una programación escrita a mano tiene que partir de los valores desplazados. ComfyUI ajusta su desplazamiento a 1024x1024, por lo que la coincidencia se pierde en otros tamaños de salida. +El LoRA también sirve para editar imágenes. Cárgalo con `LoraLoaderModelOnly` en la línea de modelo de la plantilla de edición, igual que en texto a imagen, y ejecuta el flujo de trabajo con `steps` 4 y `cfg` 1, con el sampler `euler` y el scheduler `simple` de la plantilla. + +Incluso con el lienzo ajustado a `image_1`, una edición de cuatro pasos puede desplazar el resultado hacia un lado. Baja la fuerza del LoRA a unos 0.5 para fijar el encuadre. El número de pasos sigue en cuatro, así que la velocidad no cambia. + ## Salida del canal alpha El VAE de Qwen-Image-2.1 decodifica cuatro canales, por lo que una imagen decodificada incluye alpha además de RGB. El flujo de trabajo de eliminación de fondo depende de esto: el fondo transparente se genera en el canal alpha en lugar de recortarse después. diff --git a/es/tutorials/video/minimax/minimax-h3-native.mdx b/es/tutorials/video/minimax/minimax-h3-native.mdx index dcc0f6cb27..2dbf3f969b 100644 --- a/es/tutorials/video/minimax/minimax-h3-native.mdx +++ b/es/tutorials/video/minimax/minimax-h3-native.mdx @@ -2,16 +2,16 @@ title: "Flujos de trabajo de MiniMax H3 en ComfyUI: texto a video, imagen a video y referencia a video" description: "Ejemplos paso a paso de flujos nativos de MiniMax H3 en ComfyUI, con anclaje de guías y máscaras de ruido latente para un control más fino del video." sidebarTitle: "Flujos nativos" -translationSourceHash: 9e8d385c +translationSourceHash: b66253f0 translationFrom: tutorials/video/minimax/minimax-h3-native.mdx translationBlockHashes: - "_intro": 4f5274f5 + "_intro": dde651e5 "MiniMax H3 Text to Video (T2V)": 8807c872 "MiniMax H3 Image to Video (I2V)": c59554b0 "MiniMax H3 Reference to Video (R2V)": 961829df - "Advanced workflows with the native nodes": 37164e51 + "Advanced workflows with the native nodes": 8d994444 --- -Esta página cubre las tres plantillas base de MiniMax H3 y las técnicas avanzadas con nodos nativos. Para conocer las capacidades, el ajuste y las mejoras de velocidad, consulta la [descripción general de MiniMax H3](/es/tutorials/video/minimax/minimax-h3). +Esta página cubre las tres plantillas base de MiniMax H3 y las técnicas avanzadas con nodos nativos. Para conocer las capacidades y la licencia, consulta la [descripción general de MiniMax H3](/es/tutorials/video/minimax/minimax-h3). ## MiniMax H3 Texto a Video (T2V) @@ -257,3 +257,22 @@ Consulta la plantilla [Multiframe Reference](/es/tutorials/video/minimax/minimax Las [máscaras de ruido por token](https://github.com/Comfy-Org/ComfyUI/pull/15375) en los latentes de video y audio te permiten regenerar parte de un clip. Conecta una máscara a `denoise_mask` del sampler: **0** = mantener, **1** = regenerar. Las máscaras de video se ajustan a la cuadrícula de parches latentes de 2×2; las de audio, a frames latentes completos. Úsalas para relleno inteligente, eliminación o extensión mientras mantienes el resto estable. + +### Decodificación más rápida con Sol-H3-VAE (#17411) + +El VAE de video estándar decodifica en mosaicos de 256 px con fragmentos temporales solapados. [Sol-H3-VAE](https://huggingface.co/Efficient-Large-Model/Sol-H3-VAE) es un decodificador de 24 capas destilado de ese decodificador que decodifica el clip completo en una sola pasada. Comfy-Org lo reempaquetó junto con el codificador de H3 como [minimax_h3_sol_vae_int8_convrot.safetensors](https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/vae/minimax_h3_sol_vae_int8_convrot.safetensors), así que reemplaza al VAE de video en el nodo `Load VAE` que alimenta la decodificación de video. Deja intacto el cargador del VAE de audio. Requiere ComfyUI **0.44.0** o posterior. + +Decodificación de un clip de 1344x768 con 124 frames en una RTX 5090: + +| Decodificador | Decodificación | VRAM máxima | +| --- | --- | --- | +| H3 VAE int8 | 4.61 s | 3.2 GiB | +| Sol-H3-VAE int8 | 1.70 s | 4.3 GiB | +| H3 VAE fp16 | 10.7 s | 5.5 GiB | +| Sol-H3-VAE fp16 | 3.28 s | 5.9 GiB | + +Ambos decodificadores también existen en fp16: [minimax_h3_sol_vae_fp16.safetensors](https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/vae/minimax_h3_sol_vae_fp16.safetensors) y [minimax_h3_video_vae_fp16.safetensors](https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/vae/minimax_h3_video_vae_fp16.safetensors), ambos en `ComfyUI/models/vae/`. + +El coste en calidad está en la textura. El decodificador destilado mide unos 39 dB PSNR frente al estándar en un clip de 768p generado: visualmente idéntico, con las diferencias en el detalle fino, así que decodifica un clip con textura y compara antes de cambiarlo. + +La memoria va en el otro sentido. El decodificador destilado alcanza 1.1 GiB más que el int8 estándar en el clip anterior y no puede usar mosaicos, así que `VAEDecodeTiled` decodifica el clip completo con él. Las activaciones crecen unos 4.3 KB por token latente: 3.5 GiB con 362 frames en el mismo lienzo y 5.4 GiB a 2048x1152. Además, la decodificación es una parte pequeña de un render completo, así que la ganancia de extremo a extremo queda muy por debajo de la tabla. diff --git a/ja/tutorials/image/qwen/qwen-image-2-1.mdx b/ja/tutorials/image/qwen/qwen-image-2-1.mdx index 28b3dea2af..a8f74d8876 100644 --- a/ja/tutorials/image/qwen/qwen-image-2-1.mdx +++ b/ja/tutorials/image/qwen/qwen-image-2-1.mdx @@ -2,7 +2,7 @@ title: "Qwen-Image-2.1: ComfyUI テキストからの画像生成・編集と 2K 出力" description: "Qwen-Image-2.1 は 1 つのチェックポイントでテキストからの画像生成と指示ベースの編集をこなし、ComfyUI でネイティブ 2K 出力、アルファチャンネル、最大 16 枚の参照画像に対応します。" sidebarTitle: "Qwen-Image-2.1" -translationSourceHash: 13a6434f +translationSourceHash: 73760f36 translationFrom: tutorials/image/qwen/qwen-image-2-1.mdx translationBlockHashes: "_intro": 1aac4055 @@ -11,7 +11,7 @@ translationBlockHashes: "Workflow settings": 1f64eacd "VAE decoding artifacts": 8c94f042 "Faster sampling with the Turbo LoRA": 7c22e331 - "Faster sampling with the Lightning LoRA": fa73c3a1 + "Faster sampling with the Lightning LoRA": 3fc46534 "Alpha channel output": d52c962c --- import UpdateReminder from "/snippets/ja/tutorials/update-reminder.mdx" @@ -281,6 +281,10 @@ Qwen-Image-2.1 標準の VAE は、ディテールの多い写実的な出力に `ManualSigmas` ノードは実験的とされており、入力した数値をそのまま渡すため、手書きのスケジュールにはシフト後の値を最初から入れる必要があります。ComfyUI のシフトは 1024x1024 に合わせてあるので、出力サイズが離れるほど一致度は下がります。 +この LoRA は画像編集にも使えます。テキストから画像の場合と同じく `LoraLoaderModelOnly` で編集テンプレートのモデル線に接続し、ワークフローで `steps` を 4、`cfg` を 1 にし、サンプラーとスケジューラーはテンプレートの `euler` と `simple` のままにしてください。 + +キャンバスを `image_1` に合わせていても、4 ステップの編集では結果が横方向にずれることがあります。LoRA の強度を 0.5 程度まで下げると構図が固定されます。ステップ数は 4 のままなので速度は変わりません。 + ## アルファチャンネル出力 Qwen-Image-2.1 の VAE は 4 チャンネルをデコードするため、デコード後の画像は RGB に加えてアルファを持ちます。背景除去ワークフローはこれに依存しています。透明な背景は後から切り抜くのではなく、アルファチャンネルに生成されます。 diff --git a/ja/tutorials/video/minimax/minimax-h3-native.mdx b/ja/tutorials/video/minimax/minimax-h3-native.mdx index 3c775ec48a..e2ab0094ce 100644 --- a/ja/tutorials/video/minimax/minimax-h3-native.mdx +++ b/ja/tutorials/video/minimax/minimax-h3-native.mdx @@ -2,17 +2,17 @@ title: "ComfyUI MiniMax H3 テキストから動画・画像から動画・参照から動画のワークフロー" description: "ComfyUI における MiniMax H3 ネイティブワークフローの手順例。ガイドのアンカー固定や latent ノイズマスクで生成動画をより細かく制御できます。" sidebarTitle: "ネイティブワークフロー" -translationSourceHash: 9e8d385c +translationSourceHash: b66253f0 translationFrom: tutorials/video/minimax/minimax-h3-native.mdx translationBlockHashes: - "_intro": 4f5274f5 + "_intro": dde651e5 "MiniMax H3 Text to Video (T2V)": 8807c872 "MiniMax H3 Image to Video (I2V)": c59554b0 "MiniMax H3 Reference to Video (R2V)": 961829df - "Advanced workflows with the native nodes": 37164e51 + "Advanced workflows with the native nodes": 8d994444 --- -このページは 3 つの基本 MiniMax H3 テンプレートとネイティブノードの高度なテクニックを扱います。機能、チューニング、高速化は [MiniMax H3 概要](/ja/tutorials/video/minimax/minimax-h3) を参照してください。 +このページは 3 つの基本 MiniMax H3 テンプレートとネイティブノードの高度なテクニックを扱います。機能とライセンスは [MiniMax H3 概要](/ja/tutorials/video/minimax/minimax-h3) を参照してください。 ## MiniMax H3 テキストからビデオ (T2V) @@ -258,3 +258,22 @@ H3 ノードの `positive` と latent を `MiniMaxH3AddGuide` に接続: [トークン単位マスク](https://github.com/Comfy-Org/ComfyUI/pull/15375) で動画・音声 latent の一部を再生成。 サンプラー `denoise_mask` にマスク: **0** 保持、**1** 再生成。動画は 2×2 latent パッチ、音声は latent フレーム単位。インペイント、削除、拡張に使用。 + +### Sol-H3-VAE によるデコードの高速化 (#17411) + +標準の動画 VAE は 256px タイルと時間方向に重なるチャンクでデコードします。[Sol-H3-VAE](https://huggingface.co/Efficient-Large-Model/Sol-H3-VAE) はそのデコーダーから蒸留した 24 層のデコーダーで、クリップ全体を 1 回でデコードします。Comfy-Org は H3 のエンコーダーと合わせて [minimax_h3_sol_vae_int8_convrot.safetensors](https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/vae/minimax_h3_sol_vae_int8_convrot.safetensors) として再パックしているため、動画のデコードを担う `Load VAE` ノードの動画 VAE と差し替えます。音声 VAE のローダーはそのままにしてください。ComfyUI **0.44.0** 以降が必要です。 + +1344x768、124 フレームのクリップを RTX 5090 でデコードした時間: + +| デコーダー | デコード | ピーク VRAM | +| --- | --- | --- | +| H3 VAE int8 | 4.61 s | 3.2 GiB | +| Sol-H3-VAE int8 | 1.70 s | 4.3 GiB | +| H3 VAE fp16 | 10.7 s | 5.5 GiB | +| Sol-H3-VAE fp16 | 3.28 s | 5.9 GiB | + +どちらのデコーダーにも fp16 版があります。[minimax_h3_sol_vae_fp16.safetensors](https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/vae/minimax_h3_sol_vae_fp16.safetensors) と [minimax_h3_video_vae_fp16.safetensors](https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/vae/minimax_h3_video_vae_fp16.safetensors) で、いずれも `ComfyUI/models/vae/` に置きます。 + +画質の代償はテクスチャです。蒸留デコーダーは生成した 768p のクリップで標準デコーダーに対して約 39 dB PSNR となり、見た目は一致し、差は細部に出ます。切り替える前にテクスチャの多いクリップで比較してください。 + +メモリは逆方向です。上のクリップで、蒸留デコーダーのピークは標準 int8 デコーダーより 1.1 GiB 大きくなります。さらにタイル化できないため、`VAEDecodeTiled` はクリップ全体をデコードします。アクティベーションは latent トークンあたり約 4.3 KB 増え、同じキャンバスで 362 フレームなら約 3.5 GiB、2048x1152 では約 5.4 GiB になります。デコードはレンダリング全体のごく一部なので、エンドツーエンドの短縮は表の数字よりかなり小さくなります。 diff --git a/ko/tutorials/image/qwen/qwen-image-2-1.mdx b/ko/tutorials/image/qwen/qwen-image-2-1.mdx index ea39738435..a68627632b 100644 --- a/ko/tutorials/image/qwen/qwen-image-2-1.mdx +++ b/ko/tutorials/image/qwen/qwen-image-2-1.mdx @@ -2,7 +2,7 @@ title: "Qwen-Image-2.1: ComfyUI 텍스트-이미지 생성, 이미지 편집 및 2K 출력" description: "Qwen-Image-2.1은 하나의 체크포인트로 텍스트-이미지 생성과 지시 기반 편집을 처리하며, ComfyUI에서 네이티브 2K 출력, 알파 채널, 최대 16장의 참조 이미지를 지원합니다." sidebarTitle: "Qwen-Image-2.1" -translationSourceHash: 13a6434f +translationSourceHash: 73760f36 translationFrom: tutorials/image/qwen/qwen-image-2-1.mdx translationBlockHashes: "_intro": 1aac4055 @@ -11,7 +11,7 @@ translationBlockHashes: "Workflow settings": 1f64eacd "VAE decoding artifacts": 8c94f042 "Faster sampling with the Turbo LoRA": 7c22e331 - "Faster sampling with the Lightning LoRA": fa73c3a1 + "Faster sampling with the Lightning LoRA": 3fc46534 "Alpha channel output": d52c962c --- import UpdateReminder from "/snippets/ko/tutorials/update-reminder.mdx" @@ -281,6 +281,10 @@ Qwen-Image-2.1 기본 VAE는 디테일이 많은 사실적인 출력에 미세 `ManualSigmas` 노드는 실험적이며 입력한 숫자를 그대로 전달하므로, 손으로 만든 스케줄에는 시프트 후 값을 처음부터 넣어야 합니다. ComfyUI의 시프트는 1024x1024 기준으로 맞춰져 있어 출력 크기가 멀어질수록 일치도가 떨어집니다. +이 LoRA는 이미지 편집에도 쓸 수 있습니다. 텍스트→이미지와 마찬가지로 `LoraLoaderModelOnly`로 편집 템플릿의 모델 라인에 연결한 뒤, 워크플로에서 `steps`를 4, `cfg`를 1로 두고 샘플러와 스케줄러는 템플릿의 `euler`와 `simple`을 그대로 사용하세요. + +캔버스를 `image_1`에 맞춰도 4스텝 편집에서는 결과가 옆으로 밀릴 수 있습니다. LoRA 강도를 0.5 정도로 낮추면 구도가 고정됩니다. 스텝 수는 4로 유지되므로 속도는 달라지지 않습니다. + ## 알파 채널 출력 Qwen-Image-2.1 VAE는 4개 채널을 디코딩하므로 디코딩된 이미지는 RGB와 함께 알파를 가집니다. 배경 제거 워크플로가 여기에 의존합니다. 투명한 배경은 나중에 잘라내는 것이 아니라 알파 채널에 생성됩니다. diff --git a/ko/tutorials/video/minimax/minimax-h3-native.mdx b/ko/tutorials/video/minimax/minimax-h3-native.mdx index 27da4717e1..41964bdd40 100644 --- a/ko/tutorials/video/minimax/minimax-h3-native.mdx +++ b/ko/tutorials/video/minimax/minimax-h3-native.mdx @@ -2,17 +2,17 @@ title: "ComfyUI MiniMax H3 텍스트·이미지·레퍼런스 기반 비디오 워크플로" description: "ComfyUI의 MiniMax H3 네이티브 워크플로 단계별 예제입니다. 가이드 고정과 잠재 노이즈 마스크로 생성 비디오를 더 세밀하게 제어할 수 있습니다." sidebarTitle: "네이티브 워크플로" -translationSourceHash: 9e8d385c +translationSourceHash: b66253f0 translationFrom: tutorials/video/minimax/minimax-h3-native.mdx translationBlockHashes: - "_intro": 4f5274f5 + "_intro": dde651e5 "MiniMax H3 Text to Video (T2V)": 8807c872 "MiniMax H3 Image to Video (I2V)": c59554b0 "MiniMax H3 Reference to Video (R2V)": 961829df - "Advanced workflows with the native nodes": 37164e51 + "Advanced workflows with the native nodes": 8d994444 --- -이 페이지는 세 가지 기본 MiniMax H3 템플릿과 네이티브 노드 고급 기법을 다룹니다. 기능, 튜닝, 가속은 [MiniMax H3 개요](/ko/tutorials/video/minimax/minimax-h3)를 참조하세요. +이 페이지는 세 가지 기본 MiniMax H3 템플릿과 네이티브 노드 고급 기법을 다룹니다. 기능과 라이선스는 [MiniMax H3 개요](/ko/tutorials/video/minimax/minimax-h3)를 참조하세요. ## MiniMax H3 텍스트 기반 비디오 생성 (T2V) @@ -258,3 +258,22 @@ H3 노드 `positive`·latent를 `MiniMaxH3AddGuide`에 연결: [토큰 단위 마스크](https://github.com/Comfy-Org/ComfyUI/pull/15375)로 비디오·오디오 latent 일부 재생성. 샘플러 `denoise_mask`: **0** 유지, **1** 재생성. 비디오 2×2 latent 패치, 오디오 latent 프레임 단위. 인페인팅, 제거, 확장. + +### Sol-H3-VAE로 디코딩 속도 높이기 (#17411) + +기본 비디오 VAE는 256px 타일과 시간 방향으로 겹치는 청크로 디코딩합니다. [Sol-H3-VAE](https://huggingface.co/Efficient-Large-Model/Sol-H3-VAE)는 그 디코더에서 증류한 24층 디코더로, 클립 전체를 한 번에 디코딩합니다. Comfy-Org가 H3 인코더와 함께 [minimax_h3_sol_vae_int8_convrot.safetensors](https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/vae/minimax_h3_sol_vae_int8_convrot.safetensors)로 다시 패키징했으므로 비디오 디코딩을 담당하는 `Load VAE` 노드의 비디오 VAE를 대체합니다. 오디오 VAE 로더는 그대로 두세요. ComfyUI **0.44.0** 이상이 필요합니다. + +1344x768, 124프레임 클립을 RTX 5090에서 디코딩한 시간: + +| 디코더 | 디코딩 | 최대 VRAM | +| --- | --- | --- | +| H3 VAE int8 | 4.61 s | 3.2 GiB | +| Sol-H3-VAE int8 | 1.70 s | 4.3 GiB | +| H3 VAE fp16 | 10.7 s | 5.5 GiB | +| Sol-H3-VAE fp16 | 3.28 s | 5.9 GiB | + +두 디코더 모두 fp16 버전이 있습니다. [minimax_h3_sol_vae_fp16.safetensors](https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/vae/minimax_h3_sol_vae_fp16.safetensors)와 [minimax_h3_video_vae_fp16.safetensors](https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/vae/minimax_h3_video_vae_fp16.safetensors)이며, 모두 `ComfyUI/models/vae/`에 넣습니다. + +화질의 대가는 텍스처입니다. 증류 디코더는 생성한 768p 클립에서 기본 디코더 대비 약 39 dB PSNR로, 눈으로 보기에는 같고 차이는 세부에 나타납니다. 바꾸기 전에 텍스처가 많은 클립으로 비교하세요. + +메모리는 반대입니다. 위 클립에서 증류 디코더의 최대 사용량은 기본 int8 디코더보다 1.1 GiB 높고, 타일링도 할 수 없어 `VAEDecodeTiled`는 클립 전체를 디코딩합니다. 활성값은 latent 토큰당 약 4.3 KB씩 늘어 같은 캔버스에서 362프레임이면 약 3.5 GiB, 2048x1152에서는 약 5.4 GiB가 됩니다. 디코딩은 전체 렌더링에서 작은 비중이므로 실제 단축 시간은 표보다 훨씬 작습니다. diff --git a/tutorials/image/qwen/qwen-image-2-1.mdx b/tutorials/image/qwen/qwen-image-2-1.mdx index f398b23ad1..518aa65525 100644 --- a/tutorials/image/qwen/qwen-image-2-1.mdx +++ b/tutorials/image/qwen/qwen-image-2-1.mdx @@ -271,6 +271,10 @@ The LoRA runs at `cfg` 1 with a four-step schedule. Its card lists a `1.0, 0.75, A `ManualSigmas` node is marked experimental and passes its numbers through unchanged, so a hand-written schedule has to start from the shifted values. ComfyUI fits its shift at 1024x1024, so the match loosens at other output sizes. +The LoRA covers image editing too. Load it with `LoraLoaderModelOnly` on the edit template's model line, the same way as for text to image, then run the workflow with `steps` 4 and `cfg` 1 on the template's `euler` and `simple`. + +Even with the canvas matched to `image_1`, a four-step edit can slide the result sideways. Lower the LoRA strength to about 0.5 to hold the frame in place; the step count stays at four, so the speed does not change. + ## Alpha channel output The Qwen-Image-2.1 VAE decodes four channels, so a decoded image carries alpha as well as RGB. The background removal workflow depends on this: the transparent background is generated into the alpha channel rather than cut out afterwards. diff --git a/tutorials/video/minimax/minimax-h3-native.mdx b/tutorials/video/minimax/minimax-h3-native.mdx index ac0d6102db..cdeb1b691d 100644 --- a/tutorials/video/minimax/minimax-h3-native.mdx +++ b/tutorials/video/minimax/minimax-h3-native.mdx @@ -4,7 +4,7 @@ description: "Step-by-step MiniMax H3 native workflow examples in ComfyUI, inclu sidebarTitle: "Native workflows" --- -This page covers the three base MiniMax H3 templates plus advanced native-node techniques. For capabilities, tuning, and speedups, see the [MiniMax H3 overview](/tutorials/video/minimax/minimax-h3). +This page covers the three base MiniMax H3 templates plus advanced native-node techniques. For capabilities and licensing, see the [MiniMax H3 overview](/tutorials/video/minimax/minimax-h3). ## MiniMax H3 Text to Video (T2V) @@ -250,3 +250,22 @@ See the [Multiframe Reference](/tutorials/video/minimax/minimax-h3-multiframe) t [Per-token noise masks](https://github.com/Comfy-Org/ComfyUI/pull/15375) on video and audio latents let you regenerate part of a clip. Connect a mask to the sampler `denoise_mask`: **0** = keep, **1** = regenerate. Video masks snap to the 2×2 latent patch grid; audio masks to whole latent frames. Use for inpainting, removal, or extension while holding the rest stable. + +### Faster decoding with the Sol-H3-VAE (#17411) + +The stock video VAE decodes in 256px tiles with overlapping temporal chunks. [Sol-H3-VAE](https://huggingface.co/Efficient-Large-Model/Sol-H3-VAE) is a 24-layer decoder distilled from that decoder that decodes the whole clip in one pass. Comfy-Org repacked it with the H3 encoder as [minimax_h3_sol_vae_int8_convrot.safetensors](https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/vae/minimax_h3_sol_vae_int8_convrot.safetensors), so it replaces the video VAE on the `Load VAE` node that feeds the video decode. Leave the audio VAE loader alone. It needs ComfyUI **0.44.0** or later. + +Decode of a 1344x768 clip at 124 frames on an RTX 5090: + +| Decoder | Decode | Peak VRAM | +| --- | --- | --- | +| H3 VAE int8 | 4.61 s | 3.2 GiB | +| Sol-H3-VAE int8 | 1.70 s | 4.3 GiB | +| H3 VAE fp16 | 10.7 s | 5.5 GiB | +| Sol-H3-VAE fp16 | 3.28 s | 5.9 GiB | + +Both decoders also ship at fp16: [minimax_h3_sol_vae_fp16.safetensors](https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/vae/minimax_h3_sol_vae_fp16.safetensors) and [minimax_h3_video_vae_fp16.safetensors](https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/vae/minimax_h3_video_vae_fp16.safetensors), both placed in `ComfyUI/models/vae/`. + +The quality cost is texture. The distilled decoder measures about 39 dB PSNR against the stock decoder on a generated 768p clip: visually identical, with the differences in fine detail, so decode a textured clip and compare before switching. + +Memory goes the other way. The distilled decoder peaks 1.1 GiB above the stock int8 decoder on the clip above, and it cannot tile, so `VAEDecodeTiled` decodes the whole clip on it. Activations grow about 4.3 KB per latent token, which is 3.5 GiB at 362 frames on the same canvas and 5.4 GiB at 2048x1152. Decode is also a small share of a full render, so the end-to-end gain is well below the table. diff --git a/zh/tutorials/image/qwen/qwen-image-2-1.mdx b/zh/tutorials/image/qwen/qwen-image-2-1.mdx index be42467517..f9f3a677e4 100644 --- a/zh/tutorials/image/qwen/qwen-image-2-1.mdx +++ b/zh/tutorials/image/qwen/qwen-image-2-1.mdx @@ -2,7 +2,7 @@ title: "Qwen-Image-2.1:ComfyUI 文生图、图像编辑与 2K 输出" description: "Qwen-Image-2.1 用同一套权重完成文生图与指令式图像编辑,在 ComfyUI 中支持原生 2K 输出、Alpha 透明度通道,以及最多 16 张参考图。" sidebarTitle: "Qwen-Image-2.1" -translationSourceHash: 13a6434f +translationSourceHash: 73760f36 translationFrom: tutorials/image/qwen/qwen-image-2-1.mdx translationBlockHashes: "_intro": 1aac4055 @@ -11,7 +11,7 @@ translationBlockHashes: "Workflow settings": 1f64eacd "VAE decoding artifacts": 8c94f042 "Faster sampling with the Turbo LoRA": 7c22e331 - "Faster sampling with the Lightning LoRA": fa73c3a1 + "Faster sampling with the Lightning LoRA": 3fc46534 "Alpha channel output": d52c962c --- import UpdateReminder from "/snippets/zh/tutorials/update-reminder.mdx" @@ -281,6 +281,10 @@ Qwen-Image-2.1 自带的 VAE 会在细节丰富的写实类输出上留下细密 `ManualSigmas` 节点标记为实验性,会把填入的数值原样传递,所以手写调度必须直接填平移之后的数值。ComfyUI 的平移按 1024x1024 拟合,输出尺寸偏离该分辨率后两者的贴合度会下降。 +这个 LoRA 同样覆盖图像编辑。像文生图那样,用 `LoraLoaderModelOnly` 把它接在编辑模板的模型线上,然后在工作流里把 `steps` 设为 4、`cfg` 设为 1,采样器和调度器沿用模板的 `euler` 与 `simple`。 + +即使画布已经对齐 `image_1`,四步编辑的结果仍可能整体横向偏移。把 LoRA 强度降到 0.5 左右可以固定住画面。步数仍是四步,速度不变。 + ## Alpha 通道输出 Qwen-Image-2.1 的 VAE 解码出四个通道,所以解码后的图像除 RGB 之外还带 alpha。背景移除工作流就依赖这一点:透明背景是生成进 alpha 通道的,而不是事后抠出来的。 diff --git a/zh/tutorials/video/minimax/minimax-h3-native.mdx b/zh/tutorials/video/minimax/minimax-h3-native.mdx index 2162165410..74880b0524 100644 --- a/zh/tutorials/video/minimax/minimax-h3-native.mdx +++ b/zh/tutorials/video/minimax/minimax-h3-native.mdx @@ -2,16 +2,16 @@ title: "ComfyUI MiniMax H3 文生视频、图生视频与参考生视频工作流" description: "ComfyUI 中 MiniMax H3 原生工作流的逐步示例,含帧锚定与潜在噪声蒙版技巧,帮助你对生成视频做更精细的控制。" sidebarTitle: "原生工作流" -translationSourceHash: 9e8d385c +translationSourceHash: b66253f0 translationFrom: tutorials/video/minimax/minimax-h3-native.mdx translationBlockHashes: - "_intro": 4f5274f5 + "_intro": dde651e5 "MiniMax H3 Text to Video (T2V)": 8807c872 "MiniMax H3 Image to Video (I2V)": c59554b0 "MiniMax H3 Reference to Video (R2V)": 961829df - "Advanced workflows with the native nodes": 37164e51 + "Advanced workflows with the native nodes": 8d994444 --- -本页涵盖三个基础 MiniMax H3 模板及原生节点高级技巧。能力、调参与加速见 [MiniMax H3 概览](/zh/tutorials/video/minimax/minimax-h3)。 +本页涵盖三个基础 MiniMax H3 模板及原生节点高级技巧。能力与许可见 [MiniMax H3 概览](/zh/tutorials/video/minimax/minimax-h3)。 ## MiniMax H3 文生视频(T2V) @@ -257,3 +257,22 @@ ComfyUI/ [逐 token 遮罩](https://github.com/Comfy-Org/ComfyUI/pull/15375) 覆盖视频与音频 latent,可只重生成一部分。 遮罩接采样器 `denoise_mask`:**0** 保留,**1** 重生成。视频对齐 2×2 latent 块,音频对齐整帧。用于修补、移除或扩展。 + +### 使用 Sol-H3-VAE 加速解码 (#17411) + +官方视频 VAE 以 256px 分块解码,并在时间方向上重叠。[Sol-H3-VAE](https://huggingface.co/Efficient-Large-Model/Sol-H3-VAE) 是从该解码器蒸馏出的 24 层解码器,一次解码整段视频。Comfy-Org 把它与 H3 编码器打包成 [minimax_h3_sol_vae_int8_convrot.safetensors](https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/vae/minimax_h3_sol_vae_int8_convrot.safetensors),用来替换负责视频解码的那个 `Load VAE` 节点上的视频 VAE,音频 VAE 的加载节点保持不动。需要 ComfyUI **0.44.0** 或更高版本。 + +1344x768、124 帧的片段在 RTX 5090 上的解码耗时: + +| 解码器 | 解码耗时 | 峰值显存 | +| --- | --- | --- | +| H3 VAE int8 | 4.61 s | 3.2 GiB | +| Sol-H3-VAE int8 | 1.70 s | 4.3 GiB | +| H3 VAE fp16 | 10.7 s | 5.5 GiB | +| Sol-H3-VAE fp16 | 3.28 s | 5.9 GiB | + +两种解码器都有 fp16 版本:[minimax_h3_sol_vae_fp16.safetensors](https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/vae/minimax_h3_sol_vae_fp16.safetensors) 与 [minimax_h3_video_vae_fp16.safetensors](https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/vae/minimax_h3_video_vae_fp16.safetensors),都放进 `ComfyUI/models/vae/`。 + +画质上的代价在纹理。蒸馏解码器相对官方解码器在生成的一段 768p 片段上约 39 dB PSNR:整体观感一致,差异集中在细节,切换前先用带纹理的片段对比。 + +显存则相反。在上述片段上,蒸馏解码器的峰值比官方 int8 解码器高 1.1 GiB,而且它不能分块,`VAEDecodeTiled` 在它上面会解码整段。激活值按每个 latent token 约 4.3 KB 增长,同一画布下 362 帧约 3.5 GiB,2048x1152 时约 5.4 GiB。解码也只是一次完整渲染中的小部分,端到端提升远低于表中数字。 From 415a81ae71c237d58cedc0ec03914c6332decf7d Mon Sep 17 00:00:00 2001 From: lin-bot23 Date: Fri, 6 Nov 2026 18:03:57 +0800 Subject: [PATCH 2/7] docs: document chaining H3 clips into a longer shot --- es/tutorials/video/minimax/minimax-h3-native.mdx | 12 ++++++++++-- ja/tutorials/video/minimax/minimax-h3-native.mdx | 12 ++++++++++-- ko/tutorials/video/minimax/minimax-h3-native.mdx | 12 ++++++++++-- tutorials/video/minimax/minimax-h3-native.mdx | 8 ++++++++ zh/tutorials/video/minimax/minimax-h3-native.mdx | 12 ++++++++++-- 5 files changed, 48 insertions(+), 8 deletions(-) diff --git a/es/tutorials/video/minimax/minimax-h3-native.mdx b/es/tutorials/video/minimax/minimax-h3-native.mdx index 2dbf3f969b..190077fb71 100644 --- a/es/tutorials/video/minimax/minimax-h3-native.mdx +++ b/es/tutorials/video/minimax/minimax-h3-native.mdx @@ -2,14 +2,14 @@ title: "Flujos de trabajo de MiniMax H3 en ComfyUI: texto a video, imagen a video y referencia a video" description: "Ejemplos paso a paso de flujos nativos de MiniMax H3 en ComfyUI, con anclaje de guías y máscaras de ruido latente para un control más fino del video." sidebarTitle: "Flujos nativos" -translationSourceHash: b66253f0 +translationSourceHash: 05645ecd translationFrom: tutorials/video/minimax/minimax-h3-native.mdx translationBlockHashes: "_intro": dde651e5 "MiniMax H3 Text to Video (T2V)": 8807c872 "MiniMax H3 Image to Video (I2V)": c59554b0 "MiniMax H3 Reference to Video (R2V)": 961829df - "Advanced workflows with the native nodes": 8d994444 + "Advanced workflows with the native nodes": 6e7219db --- Esta página cubre las tres plantillas base de MiniMax H3 y las técnicas avanzadas con nodos nativos. Para conocer las capacidades y la licencia, consulta la [descripción general de MiniMax H3](/es/tutorials/video/minimax/minimax-h3). @@ -252,6 +252,14 @@ Esto es acondicionamiento, no video a video: el muestreo sigue partiendo del lat Consulta la plantilla [Multiframe Reference](/es/tutorials/video/minimax/minimax-h3-multiframe). +### Continuar una toma entre generaciones de video + +`Add Guide for MiniMax H3` también acepta un lote de varios frames en `image`, así que los últimos frames de un clip pueden iniciar el siguiente: conéctalos en `frame_idx` 0 y describe la continuación en el prompt. Los lotes bajan a 5, 22 o 39 frames y el nodo usa los primeros frames del lote, así que envía exactamente los 5 frames que quieres continuar. Los frames de guía no llegan al codificador de texto, así que el prompt tiene que llevar la continuación. + +Lo que limita la cadena es la deriva, no la unión: cada clip mueve un poco el color y la escena, y una cadena larga termina fuera del aspecto del primer clip aunque no se vea ningún corte. Pasa cada clip nuevo por [Transfer Color](/es/built-in-nodes/ColorTransfer) contra el último frame del clip anterior como `image_ref`, con `source_stats` en `target_frame` y `target_index` en 0, para que una sola transformación cubra todo el clip. `per_frame` iguala cada frame por separado y puede parpadear. + +Para una composición que debe mantenerse en toda la toma, ancla imágenes fijas en lugar de encadenar clips. + ### Relleno inteligente y extensión con máscaras de ruido latente (#15375) Las [máscaras de ruido por token](https://github.com/Comfy-Org/ComfyUI/pull/15375) en los latentes de video y audio te permiten regenerar parte de un clip. diff --git a/ja/tutorials/video/minimax/minimax-h3-native.mdx b/ja/tutorials/video/minimax/minimax-h3-native.mdx index e2ab0094ce..e76158f7dc 100644 --- a/ja/tutorials/video/minimax/minimax-h3-native.mdx +++ b/ja/tutorials/video/minimax/minimax-h3-native.mdx @@ -2,14 +2,14 @@ title: "ComfyUI MiniMax H3 テキストから動画・画像から動画・参照から動画のワークフロー" description: "ComfyUI における MiniMax H3 ネイティブワークフローの手順例。ガイドのアンカー固定や latent ノイズマスクで生成動画をより細かく制御できます。" sidebarTitle: "ネイティブワークフロー" -translationSourceHash: b66253f0 +translationSourceHash: 05645ecd translationFrom: tutorials/video/minimax/minimax-h3-native.mdx translationBlockHashes: "_intro": dde651e5 "MiniMax H3 Text to Video (T2V)": 8807c872 "MiniMax H3 Image to Video (I2V)": c59554b0 "MiniMax H3 Reference to Video (R2V)": 961829df - "Advanced workflows with the native nodes": 8d994444 + "Advanced workflows with the native nodes": 6e7219db --- このページは 3 つの基本 MiniMax H3 テンプレートとネイティブノードの高度なテクニックを扱います。機能とライセンスは [MiniMax H3 概要](/ja/tutorials/video/minimax/minimax-h3) を参照してください。 @@ -253,6 +253,14 @@ H3 ノードの `positive` と latent を `MiniMaxH3AddGuide` に接続: [マルチフレーム参照](/ja/tutorials/video/minimax/minimax-h3-multiframe) テンプレートを参照。 +### 生成をまたいで同じショットを続ける + +`Add Guide for MiniMax H3` は `image` にマルチフレームのバッチも受け取るため、前のクリップの末尾フレームを次のクリップの起点にできます。`frame_idx` 0 に接続し、続きはプロンプトで記述します。バッチは 5、22、39 フレームに切り下げられ、ノードはバッチの先頭フレームを使うので、続けたいちょうど 5 フレームを渡します。ガイドのフレームはテキストエンコーダーに渡らないため、続きの内容はプロンプトに書く必要があります。 + +長いチェーンを制限するのは継ぎ目ではなくドリフトです。クリップごとに色とシーンが少しずつ動き、カットが見えなくても最後には最初のクリップと見た目が変わります。新しいクリップは前のクリップの最終フレームを `image_ref` にして [Transfer Color](/ja/built-in-nodes/ColorTransfer) を通し、`source_stats` を `target_frame`、`target_index` を 0 にするとクリップ全体に 1 つの変換がかかります。`per_frame` はフレームごとに合わせるためちらつくことがあります。 + +ショット全体で構図を固定する場合は、クリップを繋がず静止画を固定します。 + ### latent ノイズマスクによるインペイントと拡張 (#15375) [トークン単位マスク](https://github.com/Comfy-Org/ComfyUI/pull/15375) で動画・音声 latent の一部を再生成。 diff --git a/ko/tutorials/video/minimax/minimax-h3-native.mdx b/ko/tutorials/video/minimax/minimax-h3-native.mdx index 41964bdd40..d3ead2e1dc 100644 --- a/ko/tutorials/video/minimax/minimax-h3-native.mdx +++ b/ko/tutorials/video/minimax/minimax-h3-native.mdx @@ -2,14 +2,14 @@ title: "ComfyUI MiniMax H3 텍스트·이미지·레퍼런스 기반 비디오 워크플로" description: "ComfyUI의 MiniMax H3 네이티브 워크플로 단계별 예제입니다. 가이드 고정과 잠재 노이즈 마스크로 생성 비디오를 더 세밀하게 제어할 수 있습니다." sidebarTitle: "네이티브 워크플로" -translationSourceHash: b66253f0 +translationSourceHash: 05645ecd translationFrom: tutorials/video/minimax/minimax-h3-native.mdx translationBlockHashes: "_intro": dde651e5 "MiniMax H3 Text to Video (T2V)": 8807c872 "MiniMax H3 Image to Video (I2V)": c59554b0 "MiniMax H3 Reference to Video (R2V)": 961829df - "Advanced workflows with the native nodes": 8d994444 + "Advanced workflows with the native nodes": 6e7219db --- 이 페이지는 세 가지 기본 MiniMax H3 템플릿과 네이티브 노드 고급 기법을 다룹니다. 기능과 라이선스는 [MiniMax H3 개요](/ko/tutorials/video/minimax/minimax-h3)를 참조하세요. @@ -253,6 +253,14 @@ H3 노드 `positive`·latent를 `MiniMaxH3AddGuide`에 연결: [멀티프레임 레퍼런스](/ko/tutorials/video/minimax/minimax-h3-multiframe) 템플릿 참고. +### 생성 사이에 샷 이어가기 + +`Add Guide for MiniMax H3`는 `image`에 다중 프레임 배치도 받으므로 앞 클립의 마지막 프레임을 다음 클립의 시작으로 쓸 수 있습니다. `frame_idx` 0에 연결하고 이어질 내용은 프롬프트에 씁니다. 배치는 5, 22, 39 프레임으로 내려가고 노드는 배치의 앞쪽 프레임만 쓰므로, 이어갈 마지막 5 프레임만 정확히 넣습니다. 가이드 프레임은 텍스트 인코더에 전달되지 않으므로 이어질 내용은 프롬프트가 담당합니다. + +긴 체인을 제한하는 것은 이음새가 아니라 드리프트입니다. 클립마다 색과 장면이 조금씩 밀려, 컷이 보이지 않아도 마지막에는 첫 클립과 달라집니다. 새 클립은 이전 클립의 마지막 프레임을 `image_ref`로 [Transfer Color](/ko/built-in-nodes/ColorTransfer)에 통과시키고 `source_stats`를 `target_frame`, `target_index`를 0으로 두면 클립 전체에 변환 하나가 적용됩니다. `per_frame`은 프레임마다 따로 맞춰 깜빡일 수 있습니다. + +샷 전체에서 구도를 고정해야 한다면 클립을 잇지 말고 정지 이미지를 고정합니다. + ### 잠재 노이즈 마스크 인페인팅 및 확장 (#15375) [토큰 단위 마스크](https://github.com/Comfy-Org/ComfyUI/pull/15375)로 비디오·오디오 latent 일부 재생성. diff --git a/tutorials/video/minimax/minimax-h3-native.mdx b/tutorials/video/minimax/minimax-h3-native.mdx index cdeb1b691d..0c50f160d9 100644 --- a/tutorials/video/minimax/minimax-h3-native.mdx +++ b/tutorials/video/minimax/minimax-h3-native.mdx @@ -245,6 +245,14 @@ This is conditioning, not video-to-video: sampling still starts from the workflo See the [Multiframe Reference](/tutorials/video/minimax/minimax-h3-multiframe) template. +### Continuing a shot across generations + +`Add Guide for MiniMax H3` also takes a multi-frame batch on `image`, so the last frames of one clip can start the next one: connect them at `frame_idx` 0 and describe the continuation in the prompt. Batches snap down to 5, 22, or 39 frames and the node keeps the leading frames of the batch, so send exactly the 5 frames you want to continue from. Guide frames never reach the text encoder, so the prompt has to carry the continuation. + +Drift limits the chain, not the join: each clip moves the color and the scene a little, and a long chain ends up off the look of the first clip even where no cut is visible. Run every new clip through [Transfer Color](/built-in-nodes/ColorTransfer) against the previous clip's last frame as `image_ref`, with `source_stats` on `target_frame` and `target_index` at 0, so one transform covers the whole clip. `per_frame` matches each frame on its own and can flicker. + +For a composition that has to hold for the whole shot, anchor stills instead of chaining clips. + ### Inpainting and extension with latent noise masks (#15375) [Per-token noise masks](https://github.com/Comfy-Org/ComfyUI/pull/15375) on video and audio latents let you regenerate part of a clip. diff --git a/zh/tutorials/video/minimax/minimax-h3-native.mdx b/zh/tutorials/video/minimax/minimax-h3-native.mdx index 74880b0524..cb774b02d9 100644 --- a/zh/tutorials/video/minimax/minimax-h3-native.mdx +++ b/zh/tutorials/video/minimax/minimax-h3-native.mdx @@ -2,14 +2,14 @@ title: "ComfyUI MiniMax H3 文生视频、图生视频与参考生视频工作流" description: "ComfyUI 中 MiniMax H3 原生工作流的逐步示例,含帧锚定与潜在噪声蒙版技巧,帮助你对生成视频做更精细的控制。" sidebarTitle: "原生工作流" -translationSourceHash: b66253f0 +translationSourceHash: 05645ecd translationFrom: tutorials/video/minimax/minimax-h3-native.mdx translationBlockHashes: "_intro": dde651e5 "MiniMax H3 Text to Video (T2V)": 8807c872 "MiniMax H3 Image to Video (I2V)": c59554b0 "MiniMax H3 Reference to Video (R2V)": 961829df - "Advanced workflows with the native nodes": 8d994444 + "Advanced workflows with the native nodes": 6e7219db --- 本页涵盖三个基础 MiniMax H3 模板及原生节点高级技巧。能力与许可见 [MiniMax H3 概览](/zh/tutorials/video/minimax/minimax-h3)。 @@ -252,6 +252,14 @@ ComfyUI/ 见 [多帧参考](/zh/tutorials/video/minimax/minimax-h3-multiframe) 模板。 +### 分多次生成续写镜头 + +`Add Guide for MiniMax H3` 的 `image` 也接受多帧批次,因此上一段的末几帧可以作为下一段的起点:接到 `frame_idx` 0,并在提示词里描述续写内容。批次会向下对齐到 5、22、39 帧,节点只取批次开头的帧,所以正好传入要续接的那 5 帧。引导帧不进文本编码器,续写内容只能写进提示词。 + +限制长链的是漂移,不是接缝:每个片段都会让颜色和场景偏移一点,长链到最后即使看不到剪辑点,观感也已偏离首段。新片段用 [Transfer Color](/zh/built-in-nodes/ColorTransfer) 对齐上一段的末帧(`image_ref`),`source_stats` 选 `target_frame`、`target_index` 设为 0,整段共用一个变换。`per_frame` 逐帧匹配,容易闪烁。 + +若整个镜头必须固定构图,锚定静图,不要串片段。 + ### 使用 Latent 噪波遮罩进行 Inpainting 与扩展 (#15375) [逐 token 遮罩](https://github.com/Comfy-Org/ComfyUI/pull/15375) 覆盖视频与音频 latent,可只重生成一部分。 From 47c3ebf6c40fab3ed1a21fe9fff837933b4b4a89 Mon Sep 17 00:00:00 2001 From: lin-bot23 Date: Sat, 7 Nov 2026 18:07:33 +0800 Subject: [PATCH 3/7] docs: document H3 camera motion phrasing and guide clips for exact camera tracks --- .../video/minimax/minimax-h3-prompt-guide.mdx | 34 ++++++++++++++++-- .../video/minimax/minimax-h3-prompt-guide.mdx | 34 ++++++++++++++++-- .../video/minimax/minimax-h3-prompt-guide.mdx | 34 ++++++++++++++++-- .../video/minimax/minimax-h3-prompt-guide.mdx | 29 ++++++++++++++- .../video/minimax/minimax-h3-prompt-guide.mdx | 36 ++++++++++++++++--- 5 files changed, 153 insertions(+), 14 deletions(-) diff --git a/es/tutorials/video/minimax/minimax-h3-prompt-guide.mdx b/es/tutorials/video/minimax/minimax-h3-prompt-guide.mdx index 7d48c43bc4..1694140562 100644 --- a/es/tutorials/video/minimax/minimax-h3-prompt-guide.mdx +++ b/es/tutorials/video/minimax/minimax-h3-prompt-guide.mdx @@ -2,18 +2,19 @@ title: "Guía de prompts de ComfyUI MiniMax H3: guías oficiales, consejos y embeddings de prompt" description: "Escribe mejores prompts de MiniMax H3 en ComfyUI: guías oficiales, consejos generales, reglas de diálogo y hablantes, y embeddings de estilo." sidebarTitle: "Guía de prompts" -translationSourceHash: 0f1ded35 +translationSourceHash: 6b0fda32 translationFrom: tutorials/video/minimax/minimax-h3-prompt-guide.mdx translationBlockHashes: - "_intro": da22cc9d + "_intro": ed6e8344 "Official guides": 3d762c67 "General tips": fab4ad35 "Shots and cuts": 8184dae6 + "Camera motion": 55c8463a "Dialogue and speakers": 4985438e "Prompt embeddings": 9c899169 --- -H3 espera planos, cámara, diálogo y SFX en un solo prompt multimodal. Esta página reúne guías oficiales, consejos entre flujos de trabajo, reglas de planos y diálogo, y embeddings de estilo. +H3 espera planos, cámara, diálogo y SFX en un solo prompt multimodal. Esta página reúne las guías oficiales, consejos entre flujos de trabajo, reglas de planos, cámara y diálogo, y embeddings de estilo. ## Guías oficiales @@ -62,6 +63,33 @@ Numera las tomas `[Shot 1]`, `[Shot 2]`, … `[Shot 1]` empieza en 0.00s. Las to Inicia una toma nueva cuando cambien el sujeto, el espacio, el estado, el punto de vista o el momento. Los reencuadres pequeños siguen siendo una sola toma con movimiento de cámara. +## Movimiento de cámara + +Un movimiento por toma, escrito dentro del párrafo de la toma y con aquello sobre lo que actúa. La amplitud y la velocidad son opcionales y solo existen cuatro frases: `with small amplitude`, `with large amplitude`, `at slow speed`, `at fast speed`. No hay frase para la amplitud media ni para la velocidad normal, así que omitirlas significa media y normal. + +```text +The camera pushes in with small amplitude at slow speed toward the folded letter in her hands. +``` + +Frases de movimiento que el modelo acepta: + +| Movimiento | Qué hace | +| :--- | :--- | +| `Zoom In` / `Zoom Out` | Solo la distancia focal; la cámara no se mueve | +| `Push In` / `Pull Out` | La cámara avanza o retrocede | +| `Pan Left` / `Pan Right` | La cámara no se mueve; el objetivo gira en horizontal | +| `Truck Left` / `Truck Right` | La cámara se desplaza de lado | +| `Tilt Up` / `Tilt Down` | La cámara no se mueve; el objetivo gira en vertical | +| `Pedestal Up` / `Pedestal Down` | La cámara sube o baja entera | +| `Arc Shot` | La cámara recorre un arco alrededor del sujeto | +| `Tracking Shot` | La cámara sigue a un sujeto en movimiento | +| `Static Shot` | La cámara y el objetivo no se mueven | +| `Shake Slightly` / `Shake Strongly` | Vibración de cámara suave o fuerte | +| `POV` | Vista desde los ojos del sujeto | +| `Roll Clockwise` / `Roll Counterclockwise` | La cámara gira sobre el eje del objetivo | + +Si nombras dos movimientos a la vez, como un arco mientras la cámara se acerca, normalmente solo sale uno, y un movimiento que no cabe en la duración del clip se comprime para entrar. Para una trayectoria que tenga que salir exacta, pásala como guía: renderiza un doble moviéndose por la trayectoria y conecta ese clip en `image` de [`Add Guide for MiniMax H3`](/es/built-in-nodes/MiniMaxH3AddGuide) con `vae` conectado y `frame_idx` 0, con el cableado de [Anclaje de guías en cualquier fotograma](/es/tutorials/video/minimax/minimax-h3-native). El clip debe tener al menos 22 fotogramas y no durar más que el vídeo: las guías se recortan a 5, 22 o 39 fotogramas, y un clip que pasa del último fotograma da error. + ## Diálogo y hablantes - **IDs de hablante**: `(S1)`, `(S2)`, … para cada voz (incluidas las que están fuera de pantalla y las que cantan). Los IDs indexan hablantes, no números de personaje. Dúo: `(S1,S2)` diff --git a/ja/tutorials/video/minimax/minimax-h3-prompt-guide.mdx b/ja/tutorials/video/minimax/minimax-h3-prompt-guide.mdx index 6964451d70..e991e57cac 100644 --- a/ja/tutorials/video/minimax/minimax-h3-prompt-guide.mdx +++ b/ja/tutorials/video/minimax/minimax-h3-prompt-guide.mdx @@ -2,19 +2,20 @@ title: "ComfyUI MiniMax H3 プロンプトガイド: 公式ガイド、ヒント、プロンプト埋め込み" description: "ComfyUI でより良い MiniMax H3 プロンプトを書く方法: MiniMax の公式プロンプト作成ガイド、一般的なプロンプトのコツ、セリフと話者のルール、スタイル埋め込み。" sidebarTitle: "プロンプトガイド" -translationSourceHash: 0f1ded35 +translationSourceHash: 6b0fda32 translationFrom: tutorials/video/minimax/minimax-h3-prompt-guide.mdx translationBlockHashes: - "_intro": da22cc9d + "_intro": ed6e8344 "Official guides": 3d762c67 "General tips": fab4ad35 "Shots and cuts": 8184dae6 + "Camera motion": 55c8463a "Dialogue and speakers": 4985438e "Prompt embeddings": 9c899169 --- -H3 は 1 つのマルチモーダルプロンプトにショット、カメラ、台詞、SFX を載せます。このページは公式ガイド、ワークフロー横断のヒント、ショット/台詞ルール、スタイル埋め込みをまとめます。 +H3 は 1 つのマルチモーダルプロンプトにショット、カメラ、台詞、SFX を載せます。このページは公式ガイド、ワークフロー横断のヒント、ショット・カメラワーク・台詞のルール、スタイル埋め込みをまとめます。 ## 公式ガイド @@ -63,6 +64,33 @@ non_diegetic_music: ... 被写体・空間・状態・視点・時間が変わるときに新ショット。微リフレームは 1 ショットでカメラモーション。 +## カメラワーク + +1 ショットに 1 つの動きだけを、そのショットの段落に、何に対して動くかと一緒に書く。振幅と速度は任意で、使えるフレーズは 4 つだけ: `with small amplitude`、`with large amplitude`、`at slow speed`、`at fast speed`。中程度の振幅と通常速度にはフレーズがなく、省略すると中程度・通常の意味になる。 + +```text +The camera pushes in with small amplitude at slow speed toward the folded letter in her hands. +``` + +モデルが受け付けるカメラワークのフレーズ: + +| 動き | 効果 | +| :--- | :--- | +| `Zoom In` / `Zoom Out` | 焦点距離のみ変更、カメラ位置は固定 | +| `Push In` / `Pull Out` | カメラが前後へ移動 | +| `Pan Left` / `Pan Right` | 位置は固定、レンズが左右に振れる | +| `Truck Left` / `Truck Right` | カメラが左右に平行移動 | +| `Tilt Up` / `Tilt Down` | 位置は固定、レンズが上下に振れる | +| `Pedestal Up` / `Pedestal Down` | カメラ全体が上下に移動 | +| `Arc Shot` | 被写体の周りを弧を描いて移動 | +| `Tracking Shot` | 動く被写体を追従 | +| `Static Shot` | カメラもレンズも動かない | +| `Shake Slightly` / `Shake Strongly` | 弱い/強いカメラの揺れ | +| `POV` | 被写体の視点 | +| `Roll Clockwise` / `Roll Counterclockwise` | レンズ軸まわりの回転 | + +弧を描きながら寄るなど 2 つを同時に書くと、通常は片方だけが反映され、尺に収まらない動きはその尺に圧縮される。正確に再現したいトラックは文字ではなくガイドで渡す。スタンドインが軌道を動くクリップをレンダリングし、[`Add Guide for MiniMax H3`](/ja/built-in-nodes/MiniMaxH3AddGuide) の `image` に `vae` を接続して `frame_idx` 0 でつなぐ。配線は [H3 ネイティブワークフロー](/ja/tutorials/video/minimax/minimax-h3-native) のガイド固定の節を参照。クリップは 22 フレーム以上、動画より長くしない。ガイドは 5・22・39 フレームに切り詰められ、末尾を超えるクリップはエラーになる。 + ## セリフと話者 - **話者 ID**: `(S1)`、`(S2)`…(オフ画面/歌唱含む); デュエット `(S1,S2)` diff --git a/ko/tutorials/video/minimax/minimax-h3-prompt-guide.mdx b/ko/tutorials/video/minimax/minimax-h3-prompt-guide.mdx index 65f9ade5f0..752b53d34d 100644 --- a/ko/tutorials/video/minimax/minimax-h3-prompt-guide.mdx +++ b/ko/tutorials/video/minimax/minimax-h3-prompt-guide.mdx @@ -2,19 +2,20 @@ title: "ComfyUI MiniMax H3 프롬프트 가이드: 공식 가이드, 팁, 프롬프트 임베딩" description: "ComfyUI에서 더 나은 MiniMax H3 프롬프트 작성하기: MiniMax 공식 프롬프트 작성 가이드, 일반 프롬프트 팁, 대사와 화자 규칙, 스타일 임베딩." sidebarTitle: "프롬프트 가이드" -translationSourceHash: 0f1ded35 +translationSourceHash: 6b0fda32 translationFrom: tutorials/video/minimax/minimax-h3-prompt-guide.mdx translationBlockHashes: - "_intro": da22cc9d + "_intro": ed6e8344 "Official guides": 3d762c67 "General tips": fab4ad35 "Shots and cuts": 8184dae6 + "Camera motion": 55c8463a "Dialogue and speakers": 4985438e "Prompt embeddings": 9c899169 --- -H3는 하나의 멀티모달 프롬프트에 샷, 카메라, 대사, SFX를 담습니다. 이 페이지는 공식 가이드, 워크플로 공통 팁, 샷/대사 규칙, 스타일 임베딩을 모읍니다. +H3는 하나의 멀티모달 프롬프트에 샷, 카메라, 대사, SFX를 담습니다. 이 페이지는 공식 가이드, 워크플로 공통 팁, 샷·카메라 무브·대사 규칙, 스타일 임베딩을 모읍니다. ## 공식 가이드 @@ -63,6 +64,33 @@ non_diegetic_music: ... 피사체·공간·상태·시점·시간 변경 시 새 샷. 미세 리프레임은 한 샷·카메라 모션. +## 카메라 무브 + +한 샷에 움직임 하나만, 그 샷 문단 안에, 무엇에 작용하는지와 함께 쓴다. 진폭과 속도는 선택 사항이고 쓸 수 있는 표현은 네 가지뿐이다: `with small amplitude`, `with large amplitude`, `at slow speed`, `at fast speed`. 중간 진폭과 보통 속도에는 표현이 없어서, 생략하면 중간·보통을 뜻한다. + +```text +The camera pushes in with small amplitude at slow speed toward the folded letter in her hands. +``` + +모델이 받아들이는 카메라 무브 표현: + +| 움직임 | 효과 | +| :--- | :--- | +| `Zoom In` / `Zoom Out` | 초점 거리만 바뀌고 카메라 위치는 고정 | +| `Push In` / `Pull Out` | 카메라가 앞뒤로 이동 | +| `Pan Left` / `Pan Right` | 위치 고정, 렌즈가 좌우로 회전 | +| `Truck Left` / `Truck Right` | 카메라가 좌우로 평행 이동 | +| `Tilt Up` / `Tilt Down` | 위치 고정, 렌즈가 상하로 회전 | +| `Pedestal Up` / `Pedestal Down` | 카메라 전체가 상하로 이동 | +| `Arc Shot` | 피사체 주위를 호를 그리며 이동 | +| `Tracking Shot` | 움직이는 피사체를 따라감 | +| `Static Shot` | 카메라와 렌즈 모두 움직이지 않음 | +| `Shake Slightly` / `Shake Strongly` | 약한/강한 카메라 흔들림 | +| `POV` | 피사체 시점 | +| `Roll Clockwise` / `Roll Counterclockwise` | 렌즈 축 기준 회전 | + +호를 그리며 밀고 들어가기처럼 두 개를 함께 쓰면 보통 하나만 반영되고, 길이 안에 끝나지 못하는 움직임은 그 길이에 맞춰 압축된다. 정확히 나와야 하는 궤적은 글로 쓰지 말고 가이드로 넘긴다. 스탠드인이 경로를 따라 움직이는 클립을 렌더링해 [`Add Guide for MiniMax H3`](/ko/built-in-nodes/MiniMaxH3AddGuide)의 `image`에 `vae`를 연결하고 `frame_idx` 0으로 붙인다. 배선은 [H3 네이티브 워크플로](/ko/tutorials/video/minimax/minimax-h3-native)의 가이드 고정 절을 참고한다. 클립은 22프레임 이상, 영상보다 길지 않게 한다. 가이드는 5·22·39프레임으로 잘리고, 마지막 프레임을 넘는 클립은 오류가 난다. + ## 대사와 화자 - **화자 ID**: `(S1)`, `(S2)`…(오프스크린/노래 포함); 듀엣 `(S1,S2)` diff --git a/tutorials/video/minimax/minimax-h3-prompt-guide.mdx b/tutorials/video/minimax/minimax-h3-prompt-guide.mdx index f973cd7f8b..ad91e6982d 100644 --- a/tutorials/video/minimax/minimax-h3-prompt-guide.mdx +++ b/tutorials/video/minimax/minimax-h3-prompt-guide.mdx @@ -4,7 +4,7 @@ description: "Write better MiniMax H3 prompts in ComfyUI: MiniMax's official pro sidebarTitle: "Prompt guide" --- -H3 expects shots, camera, dialogue, and SFX in one multimodal prompt. This page collects official guides, cross-workflow tips, shot/dialogue rules, and style embeddings. +H3 expects shots, camera, dialogue, and SFX in one multimodal prompt. This page collects the official guides, cross-workflow tips, shot, camera-motion, and dialogue rules, and style embeddings. ## Official guides @@ -53,6 +53,33 @@ Number shots `[Shot 1]`, `[Shot 2]`, … `[Shot 1]` starts at 0.00s. Later shots New shot when subject, space, state, viewpoint, or time changes. Small reframes stay one shot with camera motion. +## Camera motion + +One movement per shot, named in the shot paragraph with what it acts on. Amplitude and speed are optional and only four phrases exist for them: `with small amplitude`, `with large amplitude`, `at slow speed`, `at fast speed`. There is no phrase for medium amplitude or normal speed, so leaving them out means medium and normal. + +```text +The camera pushes in with small amplitude at slow speed toward the folded letter in her hands. +``` + +Camera movement phrases the model accepts: + +| Movement | What it does | +| :--- | :--- | +| `Zoom In` / `Zoom Out` | Lens only, the camera stays in place | +| `Push In` / `Pull Out` | Camera travels forward or back | +| `Pan Left` / `Pan Right` | Lens pivots horizontally | +| `Truck Left` / `Truck Right` | Camera slides sideways | +| `Tilt Up` / `Tilt Down` | Lens pivots vertically | +| `Pedestal Up` / `Pedestal Down` | Camera rises or drops | +| `Arc Shot` | Camera travels around the subject | +| `Tracking Shot` | Camera follows a moving subject | +| `Static Shot` | Camera and lens hold still | +| `Shake Slightly` / `Shake Strongly` | Slight or strong camera shake | +| `POV` | View from the subject's eyes | +| `Roll Clockwise` / `Roll Counterclockwise` | Camera rolls on the lens axis | + +Naming two movements at once, such as an arc shot while pushing in, usually resolves to only one of them, and a move that cannot finish inside the clip time is squeezed into it. For a track that has to come out exact, feed it as a guide: render a stand-in moving along the path, then connect that clip at `image` on [`Add Guide for MiniMax H3`](/built-in-nodes/MiniMaxH3AddGuide) with `vae` and `frame_idx` 0, following [Anchoring guides at any frame](/tutorials/video/minimax/minimax-h3-native#anchoring-guides-at-any-frame-15439). Keep the clip at least 22 frames and no longer than the video: guides snap to 5, 22 or 39 frames, and a clip that runs past the last frame raises an error. + ## Dialogue and speakers - **Speaker IDs**: `(S1)`, `(S2)`, … for every voice (including off-screen/singing). IDs index speakers, not character numbers. Duet: `(S1,S2)` diff --git a/zh/tutorials/video/minimax/minimax-h3-prompt-guide.mdx b/zh/tutorials/video/minimax/minimax-h3-prompt-guide.mdx index a1904a068b..89a3ea90ee 100644 --- a/zh/tutorials/video/minimax/minimax-h3-prompt-guide.mdx +++ b/zh/tutorials/video/minimax/minimax-h3-prompt-guide.mdx @@ -2,19 +2,20 @@ title: "ComfyUI MiniMax H3 提示词指南:官方指南、技巧与提示词嵌入" description: "在 ComfyUI 中写出更好的 MiniMax H3 提示词:MiniMax 官方提示词编写指南、通用提示词技巧、对白与说话人规则,以及风格嵌入。" sidebarTitle: "提示词指南" -translationSourceHash: 0f1ded35 +translationSourceHash: 6b0fda32 translationFrom: tutorials/video/minimax/minimax-h3-prompt-guide.mdx translationBlockHashes: - "_intro": da22cc9d + "_intro": ed6e8344 "Official guides": 3d762c67 "General tips": fab4ad35 "Shots and cuts": 8184dae6 + "Camera motion": 55c8463a "Dialogue and speakers": 4985438e "Prompt embeddings": 9c899169 --- -H3 在单一提示词中承载镜头、运镜、对白与音效。本页汇总官方指南、跨工作流技巧、镜头/对白规则与风格嵌入。 +H3 在单一提示词中承载镜头、运镜、对白与音效。本页汇总官方指南、跨工作流技巧、镜头、运镜与对白规则,以及风格嵌入。 ## 官方指南 @@ -61,7 +62,34 @@ non_diegetic_music: ... - 写单点时间(`At 00:05.000`),不写区间 - 用 `the camera cuts to` 等动作描述;需要时可写溶解/淡入淡出 -仅景别/微角度变化时保持同一镜头,用运镜表达。 +主体、空间、状态、视角或时间变化时开新镜头;仅景别/微角度变化时保持同一镜头,用运镜表达。 + +## 运镜 + +一个镜头只写一种运镜,写在该镜头的段落里,并说明它作用在什么上。幅度和速度是可选项,只有四个短语:`with small amplitude`、`with large amplitude`、`at slow speed`、`at fast speed`。中幅度和常速没有对应短语,省略即表示中幅度与常速。 + +```text +The camera pushes in with small amplitude at slow speed toward the folded letter in her hands. +``` + +模型可用的运镜短语: + +| 运镜 | 效果 | +| :--- | :--- | +| `Zoom In` / `Zoom Out` | 只改焦距,机位不动 | +| `Push In` / `Pull Out` | 机位前后移动 | +| `Pan Left` / `Pan Right` | 机位不动,镜头水平转动 | +| `Truck Left` / `Truck Right` | 机位水平横移 | +| `Tilt Up` / `Tilt Down` | 机位不动,镜头上下转动 | +| `Pedestal Up` / `Pedestal Down` | 机位整体升降 | +| `Arc Shot` | 机位绕主体画弧 | +| `Tracking Shot` | 机位跟随移动的主体 | +| `Static Shot` | 机位与镜头都不动 | +| `Shake Slightly` / `Shake Strongly` | 轻微或强烈的镜头抖动 | +| `POV` | 主体视角 | +| `Roll Clockwise` / `Roll Counterclockwise` | 机位沿镜头轴翻滚 | + +同时写两个运镜(例如一边环绕一边推近)通常只实现其中一个;在镜头时长内走不完的运镜会被压缩进这段时长。需要精确复现的轨迹不要用文字描述,改成引导:渲染一段替身沿轨迹移动的片段,接到 [`Add Guide for MiniMax H3`](/zh/built-in-nodes/MiniMaxH3AddGuide) 的 `image`(接好 `vae`,`frame_idx` 设 0),连线方式见 [H3 原生工作流](/zh/tutorials/video/minimax/minimax-h3-native) 的锚定引导一节。片段至少 22 帧且不长于目标视频:引导会裁到 5、22、39 帧,超出视频末尾帧会报错。 ## 对白与说话人 From b8860063ff7096ce4229a5522ae0d6f5efa4eb51 Mon Sep 17 00:00:00 2001 From: lin-bot23 Date: Sat, 7 Nov 2026 18:10:50 +0800 Subject: [PATCH 4/7] docs: drop the stale anchor from the first generation link --- snippets/tutorials/basic/installation-models.mdx | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/snippets/tutorials/basic/installation-models.mdx b/snippets/tutorials/basic/installation-models.mdx index 54640e09af..3d5cdc4457 100644 --- a/snippets/tutorials/basic/installation-models.mdx +++ b/snippets/tutorials/basic/installation-models.mdx @@ -1,4 +1,4 @@ -Please ensure that you have at least one SD1.5 model file in the `ComfyUI/models/checkpoints` folder. If you are not sure how to install a model, please refer to the [Start ComfyUI's AI Drawing Journey](/get_started/first_generation#3-model-installation) section for instructions on model installation. +Please ensure that you have at least one SD1.5 model file in the `ComfyUI/models/checkpoints` folder. If you are not sure how to install a model, please refer to the [Start ComfyUI's AI Drawing Journey](/get_started/first_generation) section for instructions on model installation. You can use the following models: - [v1-5-pruned-emaonly-fp16.safetensors](https://huggingface.co/Comfy-Org/stable-diffusion-v1-5-archive/blob/main/v1-5-pruned-emaonly-fp16.safetensors) From db098340cd9025d41d24bc52d3bfefcc98112aa5 Mon Sep 17 00:00:00 2001 From: lin-bot23 Date: Sat, 7 Nov 2026 18:14:16 +0800 Subject: [PATCH 5/7] docs: keep the model install link in the localized snippets in step with English --- snippets/es/tutorials/basic/installation-models.mdx | 4 ++-- snippets/ja/tutorials/basic/installation-models.mdx | 4 ++-- snippets/ko/tutorials/basic/installation-models.mdx | 4 ++-- snippets/zh/tutorials/basic/installation-models.mdx | 4 ++-- 4 files changed, 8 insertions(+), 8 deletions(-) diff --git a/snippets/es/tutorials/basic/installation-models.mdx b/snippets/es/tutorials/basic/installation-models.mdx index 394d9e39d8..1a98b40d2b 100644 --- a/snippets/es/tutorials/basic/installation-models.mdx +++ b/snippets/es/tutorials/basic/installation-models.mdx @@ -1,5 +1,5 @@ -{/* translationSourceHash: 950ff9f3 */} -Asegúrate de tener al menos un archivo de modelo SD1.5 en la carpeta `ComfyUI/models/checkpoints`. Si no estás seguro de cómo instalar un modelo, consulta la sección [Inicia el viaje de dibujo con IA de ComfyUI](/es/get_started/first_generation#3-instalación-de-modelos) para ver las instrucciones de instalación de modelos. +{/* translationSourceHash: 50168c2a */} +Asegúrate de tener al menos un archivo de modelo SD1.5 en la carpeta `ComfyUI/models/checkpoints`. Si no estás seguro de cómo instalar un modelo, consulta la sección [Inicia el viaje de dibujo con IA de ComfyUI](/es/get_started/first_generation) para ver las instrucciones de instalación de modelos. Puedes usar los siguientes modelos: - [v1-5-pruned-emaonly-fp16.safetensors](https://huggingface.co/Comfy-Org/stable-diffusion-v1-5-archive/blob/main/v1-5-pruned-emaonly-fp16.safetensors) diff --git a/snippets/ja/tutorials/basic/installation-models.mdx b/snippets/ja/tutorials/basic/installation-models.mdx index 43713b09c0..5b21532629 100644 --- a/snippets/ja/tutorials/basic/installation-models.mdx +++ b/snippets/ja/tutorials/basic/installation-models.mdx @@ -1,5 +1,5 @@ -{/* translationSourceHash: 950ff9f3 */} -`ComfyUI/models/checkpoints` フォルダーに、少なくとも1つの SD1.5 モデルファイルが存在することを確認してください。モデルのインストール方法が不明な場合は、モデルのインストール手順について詳しく説明している「[ComfyUI による AI 描画を始める](/ja/get_started/first_generation#3-モデルのインストール)」セクションをご参照ください。 +{/* translationSourceHash: 50168c2a */} +`ComfyUI/models/checkpoints` フォルダーに、少なくとも1つの SD1.5 モデルファイルが存在することを確認してください。モデルのインストール方法が不明な場合は、モデルのインストール手順について詳しく説明している「[ComfyUI による AI 描画を始める](/ja/get_started/first_generation)」セクションをご参照ください。 以下のモデルをご利用いただけます: - [v1-5-pruned-emaonly-fp16.safetensors](https://huggingface.co/Comfy-Org/stable-diffusion-v1-5-archive/blob/main/v1-5-pruned-emaonly-fp16.safetensors) diff --git a/snippets/ko/tutorials/basic/installation-models.mdx b/snippets/ko/tutorials/basic/installation-models.mdx index cc2ea882af..3fb29b8124 100644 --- a/snippets/ko/tutorials/basic/installation-models.mdx +++ b/snippets/ko/tutorials/basic/installation-models.mdx @@ -1,5 +1,5 @@ -{/* translationSourceHash: 950ff9f3 */} -`ComfyUI/models/checkpoints` 폴더에 SD1.5 모델 파일이 최소 하나 이상 포함되어 있는지 확인해 주세요. 모델 설치 방법이 확실하지 않은 경우, [ComfyUI의 AI 드로잉 여정 시작하기](/ko/get_started/first_generation#3-모델-설치) 섹션을 참조하여 모델 설치 지침을 확인해 주세요. +{/* translationSourceHash: 50168c2a */} +`ComfyUI/models/checkpoints` 폴더에 SD1.5 모델 파일이 최소 하나 이상 포함되어 있는지 확인해 주세요. 모델 설치 방법이 확실하지 않은 경우, [ComfyUI의 AI 드로잉 여정 시작하기](/ko/get_started/first_generation) 섹션을 참조하여 모델 설치 지침을 확인해 주세요. 다음과 같은 모델을 사용할 수 있습니다: - [v1-5-pruned-emaonly-fp16.safetensors](https://huggingface.co/Comfy-Org/stable-diffusion-v1-5-archive/blob/main/v1-5-pruned-emaonly-fp16.safetensors) diff --git a/snippets/zh/tutorials/basic/installation-models.mdx b/snippets/zh/tutorials/basic/installation-models.mdx index 8c7adfd4e9..7d107f06cb 100644 --- a/snippets/zh/tutorials/basic/installation-models.mdx +++ b/snippets/zh/tutorials/basic/installation-models.mdx @@ -1,5 +1,5 @@ -{/* translationSourceHash: 950ff9f3 */} -请确保你已经在 `ComfyUI/models/checkpoints` 文件夹至少有一个 SD1.5 的模型文件,如果你还不了解如何安装模型,请参[开始 ComfyUI 的 AI 绘图之旅](/zh/get_started/first_generation#3-安装绘图模型)章节中关于模型安装的部分说明。 +{/* translationSourceHash: 50168c2a */} +请确保你已经在 `ComfyUI/models/checkpoints` 文件夹至少有一个 SD1.5 的模型文件,如果你还不了解如何安装模型,请参[开始 ComfyUI 的 AI 绘图之旅](/zh/get_started/first_generation)章节中关于模型安装的部分说明。 你可以使用下面的这些模型: - [v1-5-pruned-emaonly-fp16.safetensors](https://huggingface.co/Comfy-Org/stable-diffusion-v1-5-archive/blob/main/v1-5-pruned-emaonly-fp16.safetensors) From a5bc79b7d3a68b36eee126837b4808f9e4696a68 Mon Sep 17 00:00:00 2001 From: lin-bot23 Date: Sat, 7 Nov 2026 18:20:04 +0800 Subject: [PATCH 6/7] docs: state the real minimum clip length for H3 guide clips --- es/tutorials/video/minimax/minimax-h3-prompt-guide.mdx | 6 +++--- ja/tutorials/video/minimax/minimax-h3-prompt-guide.mdx | 6 +++--- ko/tutorials/video/minimax/minimax-h3-prompt-guide.mdx | 6 +++--- tutorials/video/minimax/minimax-h3-prompt-guide.mdx | 2 +- zh/tutorials/video/minimax/minimax-h3-prompt-guide.mdx | 6 +++--- 5 files changed, 13 insertions(+), 13 deletions(-) diff --git a/es/tutorials/video/minimax/minimax-h3-prompt-guide.mdx b/es/tutorials/video/minimax/minimax-h3-prompt-guide.mdx index 1694140562..0dc30b3919 100644 --- a/es/tutorials/video/minimax/minimax-h3-prompt-guide.mdx +++ b/es/tutorials/video/minimax/minimax-h3-prompt-guide.mdx @@ -2,14 +2,14 @@ title: "Guía de prompts de ComfyUI MiniMax H3: guías oficiales, consejos y embeddings de prompt" description: "Escribe mejores prompts de MiniMax H3 en ComfyUI: guías oficiales, consejos generales, reglas de diálogo y hablantes, y embeddings de estilo." sidebarTitle: "Guía de prompts" -translationSourceHash: 6b0fda32 +translationSourceHash: 7dd3d6e7 translationFrom: tutorials/video/minimax/minimax-h3-prompt-guide.mdx translationBlockHashes: "_intro": ed6e8344 "Official guides": 3d762c67 "General tips": fab4ad35 "Shots and cuts": 8184dae6 - "Camera motion": 55c8463a + "Camera motion": c3ff48b2 "Dialogue and speakers": 4985438e "Prompt embeddings": 9c899169 --- @@ -88,7 +88,7 @@ Frases de movimiento que el modelo acepta: | `POV` | Vista desde los ojos del sujeto | | `Roll Clockwise` / `Roll Counterclockwise` | La cámara gira sobre el eje del objetivo | -Si nombras dos movimientos a la vez, como un arco mientras la cámara se acerca, normalmente solo sale uno, y un movimiento que no cabe en la duración del clip se comprime para entrar. Para una trayectoria que tenga que salir exacta, pásala como guía: renderiza un doble moviéndose por la trayectoria y conecta ese clip en `image` de [`Add Guide for MiniMax H3`](/es/built-in-nodes/MiniMaxH3AddGuide) con `vae` conectado y `frame_idx` 0, con el cableado de [Anclaje de guías en cualquier fotograma](/es/tutorials/video/minimax/minimax-h3-native). El clip debe tener al menos 22 fotogramas y no durar más que el vídeo: las guías se recortan a 5, 22 o 39 fotogramas, y un clip que pasa del último fotograma da error. +Si nombras dos movimientos a la vez, como un arco mientras la cámara se acerca, normalmente solo sale uno, y un movimiento que no cabe en la duración del clip se comprime para entrar. Para una trayectoria que tenga que salir exacta, pásala como guía: renderiza un doble moviéndose por la trayectoria y conecta ese clip en `image` de [`Add Guide for MiniMax H3`](/es/built-in-nodes/MiniMaxH3AddGuide) con `vae` conectado y `frame_idx` 0, con el cableado de [Anclaje de guías en cualquier fotograma](/es/tutorials/video/minimax/minimax-h3-native). El clip debe tener al menos 5 fotogramas y no durar más que el vídeo: un lote de menos de 5 se usa solo como imagen fija, las guías se recortan a 5, 22 o 39 fotogramas, y un clip que pasa del último fotograma da error. ## Diálogo y hablantes diff --git a/ja/tutorials/video/minimax/minimax-h3-prompt-guide.mdx b/ja/tutorials/video/minimax/minimax-h3-prompt-guide.mdx index e991e57cac..c7a3488178 100644 --- a/ja/tutorials/video/minimax/minimax-h3-prompt-guide.mdx +++ b/ja/tutorials/video/minimax/minimax-h3-prompt-guide.mdx @@ -2,14 +2,14 @@ title: "ComfyUI MiniMax H3 プロンプトガイド: 公式ガイド、ヒント、プロンプト埋め込み" description: "ComfyUI でより良い MiniMax H3 プロンプトを書く方法: MiniMax の公式プロンプト作成ガイド、一般的なプロンプトのコツ、セリフと話者のルール、スタイル埋め込み。" sidebarTitle: "プロンプトガイド" -translationSourceHash: 6b0fda32 +translationSourceHash: 7dd3d6e7 translationFrom: tutorials/video/minimax/minimax-h3-prompt-guide.mdx translationBlockHashes: "_intro": ed6e8344 "Official guides": 3d762c67 "General tips": fab4ad35 "Shots and cuts": 8184dae6 - "Camera motion": 55c8463a + "Camera motion": c3ff48b2 "Dialogue and speakers": 4985438e "Prompt embeddings": 9c899169 --- @@ -89,7 +89,7 @@ The camera pushes in with small amplitude at slow speed toward the folded letter | `POV` | 被写体の視点 | | `Roll Clockwise` / `Roll Counterclockwise` | レンズ軸まわりの回転 | -弧を描きながら寄るなど 2 つを同時に書くと、通常は片方だけが反映され、尺に収まらない動きはその尺に圧縮される。正確に再現したいトラックは文字ではなくガイドで渡す。スタンドインが軌道を動くクリップをレンダリングし、[`Add Guide for MiniMax H3`](/ja/built-in-nodes/MiniMaxH3AddGuide) の `image` に `vae` を接続して `frame_idx` 0 でつなぐ。配線は [H3 ネイティブワークフロー](/ja/tutorials/video/minimax/minimax-h3-native) のガイド固定の節を参照。クリップは 22 フレーム以上、動画より長くしない。ガイドは 5・22・39 フレームに切り詰められ、末尾を超えるクリップはエラーになる。 +弧を描きながら寄るなど 2 つを同時に書くと、通常は片方だけが反映され、尺に収まらない動きはその尺に圧縮される。正確に再現したいトラックは文字ではなくガイドで渡す。スタンドインが軌道を動くクリップをレンダリングし、[`Add Guide for MiniMax H3`](/ja/built-in-nodes/MiniMaxH3AddGuide) の `image` に `vae` を接続して `frame_idx` 0 でつなぐ。配線は [H3 ネイティブワークフロー](/ja/tutorials/video/minimax/minimax-h3-native) のガイド固定の節を参照。クリップは 5 フレーム以上、動画より長くしない。5 フレーム未満のバッチは静止画として扱われる。ガイドは 5・22・39 フレームに切り詰められ、末尾を超えるクリップはエラーになる。 ## セリフと話者 diff --git a/ko/tutorials/video/minimax/minimax-h3-prompt-guide.mdx b/ko/tutorials/video/minimax/minimax-h3-prompt-guide.mdx index 752b53d34d..909b6166dd 100644 --- a/ko/tutorials/video/minimax/minimax-h3-prompt-guide.mdx +++ b/ko/tutorials/video/minimax/minimax-h3-prompt-guide.mdx @@ -2,14 +2,14 @@ title: "ComfyUI MiniMax H3 프롬프트 가이드: 공식 가이드, 팁, 프롬프트 임베딩" description: "ComfyUI에서 더 나은 MiniMax H3 프롬프트 작성하기: MiniMax 공식 프롬프트 작성 가이드, 일반 프롬프트 팁, 대사와 화자 규칙, 스타일 임베딩." sidebarTitle: "프롬프트 가이드" -translationSourceHash: 6b0fda32 +translationSourceHash: 7dd3d6e7 translationFrom: tutorials/video/minimax/minimax-h3-prompt-guide.mdx translationBlockHashes: "_intro": ed6e8344 "Official guides": 3d762c67 "General tips": fab4ad35 "Shots and cuts": 8184dae6 - "Camera motion": 55c8463a + "Camera motion": c3ff48b2 "Dialogue and speakers": 4985438e "Prompt embeddings": 9c899169 --- @@ -89,7 +89,7 @@ The camera pushes in with small amplitude at slow speed toward the folded letter | `POV` | 피사체 시점 | | `Roll Clockwise` / `Roll Counterclockwise` | 렌즈 축 기준 회전 | -호를 그리며 밀고 들어가기처럼 두 개를 함께 쓰면 보통 하나만 반영되고, 길이 안에 끝나지 못하는 움직임은 그 길이에 맞춰 압축된다. 정확히 나와야 하는 궤적은 글로 쓰지 말고 가이드로 넘긴다. 스탠드인이 경로를 따라 움직이는 클립을 렌더링해 [`Add Guide for MiniMax H3`](/ko/built-in-nodes/MiniMaxH3AddGuide)의 `image`에 `vae`를 연결하고 `frame_idx` 0으로 붙인다. 배선은 [H3 네이티브 워크플로](/ko/tutorials/video/minimax/minimax-h3-native)의 가이드 고정 절을 참고한다. 클립은 22프레임 이상, 영상보다 길지 않게 한다. 가이드는 5·22·39프레임으로 잘리고, 마지막 프레임을 넘는 클립은 오류가 난다. +호를 그리며 밀고 들어가기처럼 두 개를 함께 쓰면 보통 하나만 반영되고, 길이 안에 끝나지 못하는 움직임은 그 길이에 맞춰 압축된다. 정확히 나와야 하는 궤적은 글로 쓰지 말고 가이드로 넘긴다. 스탠드인이 경로를 따라 움직이는 클립을 렌더링해 [`Add Guide for MiniMax H3`](/ko/built-in-nodes/MiniMaxH3AddGuide)의 `image`에 `vae`를 연결하고 `frame_idx` 0으로 붙인다. 배선은 [H3 네이티브 워크플로](/ko/tutorials/video/minimax/minimax-h3-native)의 가이드 고정 절을 참고한다. 클립은 5프레임 이상, 영상보다 길지 않게 한다. 5프레임 미만 배치는 정지 이미지로만 쓰인다. 가이드는 5·22·39프레임으로 잘리고, 마지막 프레임을 넘는 클립은 오류가 난다. ## 대사와 화자 diff --git a/tutorials/video/minimax/minimax-h3-prompt-guide.mdx b/tutorials/video/minimax/minimax-h3-prompt-guide.mdx index ad91e6982d..50b40a9551 100644 --- a/tutorials/video/minimax/minimax-h3-prompt-guide.mdx +++ b/tutorials/video/minimax/minimax-h3-prompt-guide.mdx @@ -78,7 +78,7 @@ Camera movement phrases the model accepts: | `POV` | View from the subject's eyes | | `Roll Clockwise` / `Roll Counterclockwise` | Camera rolls on the lens axis | -Naming two movements at once, such as an arc shot while pushing in, usually resolves to only one of them, and a move that cannot finish inside the clip time is squeezed into it. For a track that has to come out exact, feed it as a guide: render a stand-in moving along the path, then connect that clip at `image` on [`Add Guide for MiniMax H3`](/built-in-nodes/MiniMaxH3AddGuide) with `vae` and `frame_idx` 0, following [Anchoring guides at any frame](/tutorials/video/minimax/minimax-h3-native#anchoring-guides-at-any-frame-15439). Keep the clip at least 22 frames and no longer than the video: guides snap to 5, 22 or 39 frames, and a clip that runs past the last frame raises an error. +Naming two movements at once, such as an arc shot while pushing in, usually resolves to only one of them, and a move that cannot finish inside the clip time is squeezed into it. For a track that has to come out exact, feed it as a guide: render a stand-in moving along the path, then connect that clip at `image` on [`Add Guide for MiniMax H3`](/built-in-nodes/MiniMaxH3AddGuide) with `vae` and `frame_idx` 0, following [Anchoring guides at any frame](/tutorials/video/minimax/minimax-h3-native#anchoring-guides-at-any-frame-15439). Keep the clip at least 5 frames and no longer than the video: a batch under 5 frames is treated as a still, guides snap to 5, 22 or 39 frames, and a clip that runs past the last frame raises an error. ## Dialogue and speakers diff --git a/zh/tutorials/video/minimax/minimax-h3-prompt-guide.mdx b/zh/tutorials/video/minimax/minimax-h3-prompt-guide.mdx index 89a3ea90ee..f042883fd6 100644 --- a/zh/tutorials/video/minimax/minimax-h3-prompt-guide.mdx +++ b/zh/tutorials/video/minimax/minimax-h3-prompt-guide.mdx @@ -2,14 +2,14 @@ title: "ComfyUI MiniMax H3 提示词指南:官方指南、技巧与提示词嵌入" description: "在 ComfyUI 中写出更好的 MiniMax H3 提示词:MiniMax 官方提示词编写指南、通用提示词技巧、对白与说话人规则,以及风格嵌入。" sidebarTitle: "提示词指南" -translationSourceHash: 6b0fda32 +translationSourceHash: 7dd3d6e7 translationFrom: tutorials/video/minimax/minimax-h3-prompt-guide.mdx translationBlockHashes: "_intro": ed6e8344 "Official guides": 3d762c67 "General tips": fab4ad35 "Shots and cuts": 8184dae6 - "Camera motion": 55c8463a + "Camera motion": c3ff48b2 "Dialogue and speakers": 4985438e "Prompt embeddings": 9c899169 --- @@ -89,7 +89,7 @@ The camera pushes in with small amplitude at slow speed toward the folded letter | `POV` | 主体视角 | | `Roll Clockwise` / `Roll Counterclockwise` | 机位沿镜头轴翻滚 | -同时写两个运镜(例如一边环绕一边推近)通常只实现其中一个;在镜头时长内走不完的运镜会被压缩进这段时长。需要精确复现的轨迹不要用文字描述,改成引导:渲染一段替身沿轨迹移动的片段,接到 [`Add Guide for MiniMax H3`](/zh/built-in-nodes/MiniMaxH3AddGuide) 的 `image`(接好 `vae`,`frame_idx` 设 0),连线方式见 [H3 原生工作流](/zh/tutorials/video/minimax/minimax-h3-native) 的锚定引导一节。片段至少 22 帧且不长于目标视频:引导会裁到 5、22、39 帧,超出视频末尾帧会报错。 +同时写两个运镜(例如一边环绕一边推近)通常只实现其中一个;在镜头时长内走不完的运镜会被压缩进这段时长。需要精确复现的轨迹不要用文字描述,改成引导:渲染一段替身沿轨迹移动的片段,接到 [`Add Guide for MiniMax H3`](/zh/built-in-nodes/MiniMaxH3AddGuide) 的 `image`(接好 `vae`,`frame_idx` 设 0),连线方式见 [H3 原生工作流](/zh/tutorials/video/minimax/minimax-h3-native) 的锚定引导一节。片段至少 5 帧且不长于目标视频:不足 5 帧的批次只当静帧用,引导会裁到 5、22、39 帧,超出视频末尾帧会报错。 ## 对白与说话人 From cdb989436b03726ed4834f1658aa9eb366379e2a Mon Sep 17 00:00:00 2001 From: lin-bot23 Date: Sun, 8 Nov 2026 18:08:07 +0800 Subject: [PATCH 7/7] docs: document long clips with H3 context windows --- es/tutorials/video/minimax/minimax-h3-native.mdx | 12 ++++++++++-- ja/tutorials/video/minimax/minimax-h3-native.mdx | 12 ++++++++++-- ko/tutorials/video/minimax/minimax-h3-native.mdx | 12 ++++++++++-- tutorials/video/minimax/minimax-h3-native.mdx | 8 ++++++++ zh/tutorials/video/minimax/minimax-h3-native.mdx | 12 ++++++++++-- 5 files changed, 48 insertions(+), 8 deletions(-) diff --git a/es/tutorials/video/minimax/minimax-h3-native.mdx b/es/tutorials/video/minimax/minimax-h3-native.mdx index 190077fb71..568ae2a527 100644 --- a/es/tutorials/video/minimax/minimax-h3-native.mdx +++ b/es/tutorials/video/minimax/minimax-h3-native.mdx @@ -2,14 +2,14 @@ title: "Flujos de trabajo de MiniMax H3 en ComfyUI: texto a video, imagen a video y referencia a video" description: "Ejemplos paso a paso de flujos nativos de MiniMax H3 en ComfyUI, con anclaje de guías y máscaras de ruido latente para un control más fino del video." sidebarTitle: "Flujos nativos" -translationSourceHash: 05645ecd +translationSourceHash: 3d160290 translationFrom: tutorials/video/minimax/minimax-h3-native.mdx translationBlockHashes: "_intro": dde651e5 "MiniMax H3 Text to Video (T2V)": 8807c872 "MiniMax H3 Image to Video (I2V)": c59554b0 "MiniMax H3 Reference to Video (R2V)": 961829df - "Advanced workflows with the native nodes": 6e7219db + "Advanced workflows with the native nodes": ed209512 --- Esta página cubre las tres plantillas base de MiniMax H3 y las técnicas avanzadas con nodos nativos. Para conocer las capacidades y la licencia, consulta la [descripción general de MiniMax H3](/es/tutorials/video/minimax/minimax-h3). @@ -260,6 +260,14 @@ Lo que limita la cadena es la deriva, no la unión: cada clip mueve un poco el c Para una composición que debe mantenerse en toda la toma, ancla imágenes fijas en lugar de encadenar clips. +### Clips más largos con ventanas de contexto (#17425) + +[`MiniMaxH3ContextWindows`](/es/built-in-nodes/MiniMaxH3ContextWindows) muestrea la toma en ventanas superpuestas, lo que extiende un clip más allá del rango de entrenamiento de unos 124 a 362 frames (5 a 15 segundos a 24fps). Insértalo en la ruta del modelo, entre el cargador del modelo y el sampler (ComfyUI **0.44.0** o posterior). H3 divide la pista de audio en los mismos límites de frame, así que el sonido se mantiene sincronizado en las uniones. + +Mantén `context_length` dentro del rango de entrenamiento y deja que el número de ventanas aporte la duración: `context_length` 362 con `context_overlap` 39 y un clip de 1654 frames alcanza unos 69 segundos a 24fps en cinco ventanas. El muestreo mantiene una ventana a la vez, así que la VRAM máxima sigue a la ventana y no al clip completo. El movimiento es donde falla: cada ventana resuelve su propia copia del solape, así que una cámara que recorre un espacio deforma el fondo en las uniones. + +Un frame de guía se recorta a la ventana en la que cae, así que una imagen fija en el frame 0 condiciona solo la primera ventana y un frame final solo la última. Las ventanas intermedias no llevan guía, así que un flujo de trabajo con primer y último frame deriva o alucina en la parte central. Las referencias no dependen de una posición en la línea de tiempo y llegan a todas las ventanas, así que una toma larga que deba mantener un personaje o una voz va en [R2V](/es/tutorials/video/minimax/minimax-h3-native#minimax-h3-referencia-a-video-r2v). + ### Relleno inteligente y extensión con máscaras de ruido latente (#15375) Las [máscaras de ruido por token](https://github.com/Comfy-Org/ComfyUI/pull/15375) en los latentes de video y audio te permiten regenerar parte de un clip. diff --git a/ja/tutorials/video/minimax/minimax-h3-native.mdx b/ja/tutorials/video/minimax/minimax-h3-native.mdx index e76158f7dc..a79d1dc435 100644 --- a/ja/tutorials/video/minimax/minimax-h3-native.mdx +++ b/ja/tutorials/video/minimax/minimax-h3-native.mdx @@ -2,14 +2,14 @@ title: "ComfyUI MiniMax H3 テキストから動画・画像から動画・参照から動画のワークフロー" description: "ComfyUI における MiniMax H3 ネイティブワークフローの手順例。ガイドのアンカー固定や latent ノイズマスクで生成動画をより細かく制御できます。" sidebarTitle: "ネイティブワークフロー" -translationSourceHash: 05645ecd +translationSourceHash: 3d160290 translationFrom: tutorials/video/minimax/minimax-h3-native.mdx translationBlockHashes: "_intro": dde651e5 "MiniMax H3 Text to Video (T2V)": 8807c872 "MiniMax H3 Image to Video (I2V)": c59554b0 "MiniMax H3 Reference to Video (R2V)": 961829df - "Advanced workflows with the native nodes": 6e7219db + "Advanced workflows with the native nodes": ed209512 --- このページは 3 つの基本 MiniMax H3 テンプレートとネイティブノードの高度なテクニックを扱います。機能とライセンスは [MiniMax H3 概要](/ja/tutorials/video/minimax/minimax-h3) を参照してください。 @@ -261,6 +261,14 @@ H3 ノードの `positive` と latent を `MiniMaxH3AddGuide` に接続: ショット全体で構図を固定する場合は、クリップを繋がず静止画を固定します。 +### コンテキストウィンドウで長いクリップを生成する (#17425) + +[`MiniMaxH3ContextWindows`](/ja/built-in-nodes/MiniMaxH3ContextWindows) はオーバーラップするウィンドウに分けてサンプリングし、クリップを訓練範囲の約 124〜362 フレーム(24fps で 5〜15 秒)より長くします。モデルローダーとサンプラーの間のモデルパスに挿入します(ComfyUI **0.44.0** 以降)。H3 は同じフレーム境界で音声トラックを分割するため、継ぎ目でも映像と音声は同期します。 + +`context_length` は訓練範囲内に保ち、長さはウィンドウ数で稼ぎます。`context_length` 362、`context_overlap` 39、1654 フレームのクリップなら 24fps で 5 ウィンドウ、約 69 秒に届きます。サンプリングは同時に 1 ウィンドウだけを保持するため、ピーク VRAM はクリップ全体ではなくウィンドウに従います。動きの多いショットでは、ウィンドウごとにオーバーラップを独自に解くため、カメラが空間を移動すると継ぎ目で背景が変形します。 + +ガイドフレームは属するウィンドウに切り取られるため、フレーム 0 の静止画が条件づけるのは最初のウィンドウだけ、末尾フレームは最後のウィンドウだけです。中間のウィンドウにはガイドがなく、最初と最後のフレームを使うワークフローは中盤でドリフトやハルシネーションが起きます。リファレンスはタイムライン上の位置に縛られず全ウィンドウに届くため、キャラクターや声を長く保つショットは [R2V](/ja/tutorials/video/minimax/minimax-h3-native#minimax-h3-リファレンスからビデオ-r2v) を使います。 + ### latent ノイズマスクによるインペイントと拡張 (#15375) [トークン単位マスク](https://github.com/Comfy-Org/ComfyUI/pull/15375) で動画・音声 latent の一部を再生成。 diff --git a/ko/tutorials/video/minimax/minimax-h3-native.mdx b/ko/tutorials/video/minimax/minimax-h3-native.mdx index d3ead2e1dc..5aa483fb63 100644 --- a/ko/tutorials/video/minimax/minimax-h3-native.mdx +++ b/ko/tutorials/video/minimax/minimax-h3-native.mdx @@ -2,14 +2,14 @@ title: "ComfyUI MiniMax H3 텍스트·이미지·레퍼런스 기반 비디오 워크플로" description: "ComfyUI의 MiniMax H3 네이티브 워크플로 단계별 예제입니다. 가이드 고정과 잠재 노이즈 마스크로 생성 비디오를 더 세밀하게 제어할 수 있습니다." sidebarTitle: "네이티브 워크플로" -translationSourceHash: 05645ecd +translationSourceHash: 3d160290 translationFrom: tutorials/video/minimax/minimax-h3-native.mdx translationBlockHashes: "_intro": dde651e5 "MiniMax H3 Text to Video (T2V)": 8807c872 "MiniMax H3 Image to Video (I2V)": c59554b0 "MiniMax H3 Reference to Video (R2V)": 961829df - "Advanced workflows with the native nodes": 6e7219db + "Advanced workflows with the native nodes": ed209512 --- 이 페이지는 세 가지 기본 MiniMax H3 템플릿과 네이티브 노드 고급 기법을 다룹니다. 기능과 라이선스는 [MiniMax H3 개요](/ko/tutorials/video/minimax/minimax-h3)를 참조하세요. @@ -261,6 +261,14 @@ H3 노드 `positive`·latent를 `MiniMaxH3AddGuide`에 연결: 샷 전체에서 구도를 고정해야 한다면 클립을 잇지 말고 정지 이미지를 고정합니다. +### 컨텍스트 윈도우로 긴 클립 생성하기 (#17425) + +[`MiniMaxH3ContextWindows`](/ko/built-in-nodes/MiniMaxH3ContextWindows)는 겹치는 윈도우로 샘플링해 클립을 학습 범위(약 124~362 프레임, 24fps에서 5~15초)보다 길게 만듭니다. 모델 로더와 샘플러 사이의 모델 경로에 넣습니다(ComfyUI **0.44.0** 이상). H3는 같은 프레임 경계에서 오디오 트랙을 나누므로 이음새에서도 영상과 소리가 동기화됩니다. + +`context_length`는 학습 범위 안에 두고 길이는 윈도우 개수로 확보합니다. `context_length` 362, `context_overlap` 39, 1654 프레임 클립이면 24fps에서 다섯 윈도우로 약 69초에 이릅니다. 샘플링은 한 번에 윈도우 하나만 유지하므로 최대 VRAM은 클립 전체가 아니라 윈도우를 따릅니다. 움직임이 많은 샷에서는 윈도우마다 겹치는 구간을 따로 계산해, 카메라가 공간을 지나가면 이음새에서 배경이 변형됩니다. + +가이드 프레임은 속한 윈도우로 잘리므로 프레임 0의 정지 이미지는 첫 윈도우만, 마지막 프레임은 마지막 윈도우만 조건으로 작용합니다. 가운데 윈도우에는 가이드가 없어 처음/마지막 프레임 워크플로는 중반에 드리프트하거나 환각이 생깁니다. 레퍼런스는 타임라인 위치에 묶이지 않아 모든 윈도우에 전달되므로, 캐릭터나 목소리를 길게 유지해야 하는 샷은 [R2V](/ko/tutorials/video/minimax/minimax-h3-native#minimax-h3-레퍼런스-기반-비디오-생성-r2v)를 씁니다. + ### 잠재 노이즈 마스크 인페인팅 및 확장 (#15375) [토큰 단위 마스크](https://github.com/Comfy-Org/ComfyUI/pull/15375)로 비디오·오디오 latent 일부 재생성. diff --git a/tutorials/video/minimax/minimax-h3-native.mdx b/tutorials/video/minimax/minimax-h3-native.mdx index 0c50f160d9..4d9a01bcf1 100644 --- a/tutorials/video/minimax/minimax-h3-native.mdx +++ b/tutorials/video/minimax/minimax-h3-native.mdx @@ -253,6 +253,14 @@ Drift limits the chain, not the join: each clip moves the color and the scene a For a composition that has to hold for the whole shot, anchor stills instead of chaining clips. +### Longer clips with context windows (#17425) + +[`MiniMaxH3ContextWindows`](/built-in-nodes/MiniMaxH3ContextWindows) samples a take in overlapping windows, which extends a clip past the trained range of about 124 to 362 frames (5 to 15 seconds at 24fps). Insert it on the model path between the model loader and the sampler (ComfyUI **0.44.0** or later). H3 splits the audio track at the same frame boundaries, so sound stays in sync across the seams. + +Keep `context_length` inside the trained range and let the number of windows carry the duration: `context_length` 362 with `context_overlap` 39 and a 1654-frame clip reaches about 69 seconds at 24fps in five windows. Sampling holds one window at a time, so peak VRAM follows the window, not the whole clip. Motion is where it breaks down: each window resolves its own copy of the overlap, so a camera that travels through a space morphs the background at the seams. + +A guide frame is cropped into the window it falls in, so a still at frame 0 conditions the first window only and a closing frame the last. Middle windows carry no guide, so a first/last-frame workflow drifts or hallucinates through the middle. References are not tied to a timeline position and reach every window, so a long take that has to hold a character or a voice belongs in [R2V](/tutorials/video/minimax/minimax-h3-native#minimax-h3-reference-to-video-r2v). + ### Inpainting and extension with latent noise masks (#15375) [Per-token noise masks](https://github.com/Comfy-Org/ComfyUI/pull/15375) on video and audio latents let you regenerate part of a clip. diff --git a/zh/tutorials/video/minimax/minimax-h3-native.mdx b/zh/tutorials/video/minimax/minimax-h3-native.mdx index cb774b02d9..2685d94d62 100644 --- a/zh/tutorials/video/minimax/minimax-h3-native.mdx +++ b/zh/tutorials/video/minimax/minimax-h3-native.mdx @@ -2,14 +2,14 @@ title: "ComfyUI MiniMax H3 文生视频、图生视频与参考生视频工作流" description: "ComfyUI 中 MiniMax H3 原生工作流的逐步示例,含帧锚定与潜在噪声蒙版技巧,帮助你对生成视频做更精细的控制。" sidebarTitle: "原生工作流" -translationSourceHash: 05645ecd +translationSourceHash: 3d160290 translationFrom: tutorials/video/minimax/minimax-h3-native.mdx translationBlockHashes: "_intro": dde651e5 "MiniMax H3 Text to Video (T2V)": 8807c872 "MiniMax H3 Image to Video (I2V)": c59554b0 "MiniMax H3 Reference to Video (R2V)": 961829df - "Advanced workflows with the native nodes": 6e7219db + "Advanced workflows with the native nodes": ed209512 --- 本页涵盖三个基础 MiniMax H3 模板及原生节点高级技巧。能力与许可见 [MiniMax H3 概览](/zh/tutorials/video/minimax/minimax-h3)。 @@ -260,6 +260,14 @@ ComfyUI/ 若整个镜头必须固定构图,锚定静图,不要串片段。 +### 使用上下文窗口生成长片段 (#17425) + +[`MiniMaxH3ContextWindows`](/zh/built-in-nodes/MiniMaxH3ContextWindows) 以重叠窗口采样,让片段超出约 124 至 362 帧(24fps 下 5 至 15 秒)的训练范围。把它插在模型加载器与采样器之间的模型路径上(需 ComfyUI **0.44.0** 或更高版本)。H3 在相同的帧边界切分音轨,因此接缝处音画保持同步。 + +把 `context_length` 留在训练范围内,时长交给窗口数量承担:`context_length` 362、`context_overlap` 39 搭配 1654 帧的片段,在五个窗口内可达到 24fps 下约 69 秒。采样一次只保留一个窗口,峰值显存取决于窗口大小,而非整段片段。问题出在运动:每个窗口各自解析重叠部分,镜头穿过空间时背景会在接缝处变形。 + +引导帧会被裁进它所在的窗口,因此第 0 帧的静图只约束第一个窗口,收尾帧只约束最后一个窗口。中间窗口没有引导帧,首尾帧工作流在中段会漂移或出现幻觉。参考不绑定时间轴位置,每个窗口都能看到,需要长时间保持角色或声音的镜头应改用 [R2V](/zh/tutorials/video/minimax/minimax-h3-native#minimax-h3-参考转视频(r2v))。 + ### 使用 Latent 噪波遮罩进行 Inpainting 与扩展 (#15375) [逐 token 遮罩](https://github.com/Comfy-Org/ComfyUI/pull/15375) 覆盖视频与音频 latent,可只重生成一部分。