Skip to content

minimax-h3-image 0.1.0rc1: MiniMax-H3-Image, its prompt expander, lanes (se-146) - #958

Merged
PaulFidika merged 8 commits into
rc/0.26.0from
claude/se-146-h3-image
Nov 17, 2026
Merged

PaulFidika merged 8 commits into
rc/0.26.0from
claude/se-146-h3-image

Conversation

@PaulFidika

@PaulFidika PaulFidika commented Nov 17, 2026 •

Copy link
Copy Markdown
Contributor

se-146: MiniMax-H3-Image, MiniMax's image model, with its prompt expander. Paul, 11-17: "please look into minimax h3
image as well. WE should add this model to our roster when possible", "keep the prompt expander as well".

  • New package minimax-h3-image 0.1.0rc1. One tool, generate ("MiniMax-H3-Image"): text to image, or an edit of up
    to 16 pictures that the instructions mention (<Picture n>). It runs 6 steps with no CFG at 4-20 MP (16 by
    default), 1:1 to 21:9.
  • Slots, with cl-1103 rules:
    • model: the DiT and the VAE decoder.
    • vae_encoder: fetched only when the request has pictures.
    • text_encoder: minimaxai/minimax-h3@1.0.0-rc.8's. It is byte-identical upstream, so no new bytes are
      stored.
    • expander: fetched only with prompt_upsampling. It is MiniMax-H3-Image-PE (a Qwen3.8-27B fine-tune). An edit
      runs MiniMax's edit PE prompt on the same checkpoint, which keeps its vision tower.
  • Lane functions:
    • prepare: bf16 with AdaLN tables at the 6-step grid, and the VAE split into encoder and decoder.
    • expander_bf16: decodes the NVFP4 weights.
    • expander_config: drops ModelOpt's quantization_config, which the Runtime refuses.
    • abliterate: projects the refusal direction out of the residual writers.
  • Recipes: int8 (the default), w4a8 and nvfp4 for the image model; int8 and w4a8 for the expander.
  • Implementation: a port of Diffusers 0.43's minimax_h3_image onto Diffusers 0.41's H3 classes. The velocity-form
    Euler and the draw order are Diffusers'. It reuses P-H3-Image's fused glue.
  • Mirrors published: minimaxai/minimax-h3-image@0.1.0 (bf16-full, bf16, int8, w4a8, nvfp4) and
    minimaxai/minimax-h3-image-pe@0.1.0 (nvfp4 as released, bf16, int8, w4a8).

Proof

All runs used cozy run ./minimax-h3-image/generate in the default home, project ~/cozy-scratch/se-146 (runs SE146-n).
Stack: cozy 0.26.0-dev+gec8ec2ae, Runtime 0.47.0.dev2306+gca841fe7. Lane int8: the DiT and VAE decoder int8.convrot
(GEMM cozy-int8, DiT attention kitchen-int8, VAE attention sdpa), and H3's text encoder int8. The expander is int8
(w4a8 on the 4090). Fused glue minimax-h3-image-glue-fp8 is compiled. A is the upstream example request (4736x3552,
seed 42); C is a 2-picture edit (4096x4096); B and D are the same with the expander. The reference is Diffusers 0.43's
pipeline in bf16 on MiniMax's own files, run on the same pod at the same seed. For B and D the reference gets our
expander's caption.

RTX PRO 6000 Blackwell (runpod US-NC-1):

run lane total executing denoise steady step decode peak VRAM PSNR vs reference
A SE146-12 (cold) int8 29.2 s 14.7 s 8.7 s 1.6 s 0.82 s 94 GiB resident (torch 5.3) 17.0 dB
A SE146-14 (warm, seed 43) int8 12.4 s 11.9 s 9.9 s 1.6 s 0.75 s same –
A SE146-11 bf16, exact attention 58.3 s 30.1 s 17.7 s 2.9 s 0.84 s 94 GiB 21.2 dB
C SE146-18 int8 53.1 s 24.0 s (pictures 2.0 s) 17.9 s 3.0 s 0.70 s – 24.6 dB
C SE146-58 bf16 44.7 s 43.5 s – 5.3 s – – 29.1 dB
B SE146-26 int8 + expander int8 58.2 s 57.3 s (expander 39.8 s) 10.1 s 1.7 s 0.77 s 86 GiB 17.5 dB
B SE146-59 int8 + expander nvfp4 81.3 s 42.4 s (expander 25.1 s) – – – – –
D SE146-27 int8 + expander int8 130 s 92.4 s (expander 68.7 s) 18.2 s 3.0 s 0.70 s – 29.1 dB
  • Reference times, Diffusers bf16 with automatic CPU offload: A 55.7 s at peak 64.7 GiB; C 73.4 s at 67.6 GiB.
  • ComfyUI scout on the same pod: ComfyUI ee34406f with Comfy-Org's minimax_h3_image_pruned_int8_convrot and its
    int8 text encoder, request A, no expander.
    • Warm-up run 29.1 s; timed run 15.5 s (6 x 1.72 s plus a ~4 s decode); PSNR 16.5 dB against the reference.
    • Cozy's warm A takes 12.4 s, so Cozy is 1.25x faster: 1.6 s against 1.72 s per step, and 0.75 s against ~4 s to
      decode.

RTX 4090 (vast FR): the int8 text encoder (25 GiB) and DiT (20 GiB) page.

run lane total executing denoise steady step decode peak VRAM
A SE146-37 (cold) int8 574 s (534 s spent downloading 44 GiB at ~80 MiB/s) 39.8 s 16.7 s 2.8 s 1.6 s 23 GiB
A SE146-49 (warm, seed 43) int8 49.5 s 39.7 s 16.4 s 2.7 s 1.6 s 23 GiB
C SE146-51 int8 87.4 s 53.3 s (pictures 3.4 s) 30.2 s 5.0 s 2.0 s 23 GiB
B SE146-54 int8 + expander w4a8 385 s 182.7 s (expander 110 s) 22.9 s 3.8 s – 23 GiB
D SE146-60 int8 + expander w4a8 274 s 271 s (expander 214 s) 30.9 s 5.2 s 1.1 s 23 GiB

Looked at:

  • A matches the reference's composition (cheese board, bread, bowls); the details differ, as ComfyUI's int8 does.
  • C keeps both hosts, the MONEY TALKS sign and the feast's items from the second picture.
  • The expander picked 4:3 for B, as MiniMax's own example run did.
  • D read the text on the reference's mugs and posters, and the render spells it right ("INVEST IN YOURSELF",
    "COMPOUND CONSISTENCY CONFIDENCE").
  • On the 4090, A, B and C are correct too.

CPU release-gate cell minimax-h3-image (cozy-creator#3695), four cells:

Low-bit study (report only, no default changed), RTX 5060 Ti 16 GB, request A:

lane denoise steady step peak VRAM PSNR vs reference
int8 47.1 s 7.8 s 15 GiB 16.8 dB
w4a8 48.5 s 8.1 s 15 GiB 14.5 dB
nvfp4 38.3 s 6.4 s 15 GiB 15.2 dB
  • Runs: int8 SE146-39, w4a8 SE146-57, nvfp4 SE146-61.
  • 2-bit, ComfyUI with unsloth's Q2_K on an RTX 5080: 66.9 s and 12.6 dB, visibly degraded. Details in se-146.

Removed

n/a: a new package.

@PaulFidika
PaulFidika marked this pull request as ready for review November 17, 2026 11:12
@PaulFidika
PaulFidika added this pull request to the merge queue Nov 17, 2026
Merged via the queue into rc/0.26.0 with commit 1625a97 Nov 17, 2026
18 checks passed
@PaulFidika
PaulFidika deleted the claude/se-146-h3-image branch November 17, 2026 11:18
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant