Add NitroFusion2 as a one-step Text to Image preset

CLAUDE.md

@@ -54,7 +54,7 @@ Load `dist/` in Photoshop with UXP Developer Tool (Add Plugin → `dist/manifest
   Don't grow it: new tools go in `src/ui/tools/<name>.ts` (see `livePainting.ts`, `layerTools.ts`).
 - `src/ui/generationController.ts`: the only way to run a generation (submit → WS progress →
   poll `/history` → fetch `/view` → commit). Exactly one active run; stale runs cannot commit.
-- `src/comfy/presetRegistry.ts`: source of truth for every preset (39): API workflow file, node
+- `src/comfy/presetRegistry.ts`: source of truth for every preset (40): API workflow file, node
   ids + `class_type` + required inputs, injection targets, model loader, required models with
   download URL/size/licence gate. Setup screen, health checks, setup pack and hardware advisor
   derive from it (Live Painting graphs are the exception).

README.md

@@ -351,13 +351,14 @@ Almost every preset is one of these. Install a stack once and every tool that us
 costs 12.3 GB, not 20.7 GB. `ae.safetensors` is shared by Z_image_Turbo and Flux Fill.
 
 <details>
-<summary><b>Text to Image</b> — Klein, Z-Image, Krea-2, or any checkpoint you own</summary>
+<summary><b>Text to Image</b> — Klein, Z-Image, NitroFusion2, Krea-2, or any checkpoint you own</summary>
 
 | Preset | Needs | Extra |
 | :--- | :--- | :--- |
 | Standard checkpoint | any SD 1.x / SDXL checkpoint you already have | — |
 | FLUX.2 Klein | **Klein stack** | — |
 | Z_image_Turbo | **Z_image_Turbo stack** | — |
+| NitroFusion2 | `nf2-bf16.safetensors` → `models/diffusion_models/` + the **Z_image_Turbo stack**'s `qwen_3_4b` and `ae` | +12.3 GB · Apache-2.0 · one step, a Z_image_Turbo fine-tune |
 | Krea-2 Turbo | **Krea-2 Turbo stack** | — |
 | Qwen-Image 2.1 | **Qwen-Image 2.1 stack** | research licence · the one preset with **Transparent Background** · native 2K |
 | Flux1-dev fp8 | `flux1-dev-fp8.safetensors` → `models/checkpoints/` (17.3 GB) | non-commercial licence |

docs/index.html

@@ -119,7 +119,7 @@ <h2 id="features-title">What works now</h2>
             <article class="tool-card">
               <span class="tool-icon"><img src="assets/tools/text-to-image.png" alt=""></span>
               <div><h3>Text to Image</h3><span class="chip chip-ready">Stable</span></div>
-              <p>Prompt, negative prompt, size, steps, CFG, and seed into your checkpoint of choice — SD 1.x, SDXL, FLUX.2 Klein, Z-Image-Turbo, Krea-2, and Qwen-Image 2.1, which can hand back a transparent cut-out layer and letters spelled exactly.</p>
+              <p>Prompt, negative prompt, size, steps, CFG, and seed into your checkpoint of choice — SD 1.x, SDXL, FLUX.2 Klein, Z-Image-Turbo, one-step NitroFusion2, Krea-2, and Qwen-Image 2.1, which can hand back a transparent cut-out layer and letters spelled exactly.</p>
             </article>
             <article class="tool-card">
               <span class="tool-icon"><img src="assets/tools/image-to-image.png" alt=""></span>

docs/llms.txt

@@ -33,7 +33,7 @@ First run: start with FLUX.2 Klein 4B (Apache-2.0, three files, 12.5 GB, no sign
 
 ## Supported models
 
-FLUX.2 Klein 4B (Apache-2.0), SeedVR2 3B for generative upscale (Apache-2.0), Qwen-Image 2.1, Qwen-Image-Layered, Krea-2 Turbo, Z-Image Turbo, Flux Fill, and Stable Diffusion 1.5 checkpoints with your own LoRAs. Qwen-Image 2.1 weights are licensed for research and evaluation only, and Flux Fill is non-commercial, so those presets are opt-in and labelled. Check each model's licence before commercial use.
+FLUX.2 Klein 4B (Apache-2.0), SeedVR2 3B for generative upscale (Apache-2.0), Qwen-Image 2.1, Qwen-Image-Layered, Krea-2 Turbo, Z-Image Turbo, NitroFusion2 (Apache-2.0, a one-step Z-Image Turbo fine-tune), Flux Fill, and Stable Diffusion 1.5 checkpoints with your own LoRAs. Qwen-Image 2.1 weights are licensed for research and evaluation only, and Flux Fill is non-commercial, so those presets are opt-in and labelled. Check each model's licence before commercial use.
 
 ## Links
 

docs/model-guide.md

@@ -37,6 +37,7 @@ These are practical starter guidelines, not strict rules.
 | Inpaint / Flux Fill | Flux Fill Dev | Experimental `inpaint-flux-fill-basic`; not production-ready. `inpaint-flux-fill-cropstitch` uses the same files and adds the `comfyui-inpaint-cropandstitch` node pack. |
 | Outpaint / Canvas Expansion | Flux Fill Dev | Experimental `outpaint-flux-fill-basic` using `ImagePadForOutpaint`. |
 | Z_image_Turbo Text/Image | Z_image_Turbo | Experimental dedicated presets in v0.4.3-alpha. |
+| NitroFusion2 Text to Image | NitroFusion2 (a Z_image_Turbo fine-tune) | `txt2img-nitrofusion2`: 1 step, CFG 1, 1024 × 1024. Needs `nf2-bf16.safetensors` (12.3 GB, Apache-2.0, ungated) plus Z_image_Turbo's `qwen_3_4b` and `ae`. Keep steps at 1 or 2; at 4 it speckles. |
 | Prompt from Layer | Florence-2 PromptGen | Available for alpha testing through `prompt-from-layer-florence2` for describing captured Photoshop sources. |
 | Future Realtime Preview | To be tested | Likely needs a very fast dedicated workflow. |
 

src/comfy/modelCompatibility.ts

@@ -30,6 +30,15 @@ export function detectCheckpointFamily(checkpointName: string): ModelFamily {
     return "zImage";
   }
 
+  // NitroFusion2 is a Z_image_Turbo fine-tune whose files are named nf2-*.
+  // Matched as a prefix rather than a substring: "nf2" is short enough to turn
+  // up inside unrelated names. ComfyUI lists files in subfolders as
+  // "folder/name", so the prefix is checked on the file name alone.
+  const fileName = normalized.slice(Math.max(normalized.lastIndexOf("/"), normalized.lastIndexOf("\\")) + 1);
+  if (fileName.startsWith("nf2-") || fileName.startsWith("nf2_") || normalized.includes("nitrofusion2")) {
+    return "zImage";
+  }
+
   // Must precede the plain "flux" test, which would otherwise swallow every
   // Flux.2 filename by substring. The two are not interchangeable: Flux.1 runs
   // through a plain KSampler, while Flux.2 needs the advanced sampler chain and

src/comfy/presetRegistry.ts

@@ -59,6 +59,10 @@ const COMFY_ORG_FLUX1_DEV_REPO = "https://huggingface.co/Comfy-Org/flux1-dev";
 const COMFY_ORG_FLUX2_DEV_REPO = "https://huggingface.co/Comfy-Org/flux2-dev";
 const CITY96_FLUX2_DEV_GGUF_REPO = "https://huggingface.co/city96/FLUX.2-dev-gguf";
 const COMFY_ORG_Z_IMAGE_TURBO_REPO = "https://huggingface.co/Comfy-Org/z_image_turbo";
+// NitroFusion2 is an Apache-2.0, ungated fine-tune of Z_image_Turbo, published by
+// its authors rather than repackaged by Comfy-Org, so the model URL points at
+// their repo. No licenseGate, for the same reason Klein has none.
+const CHENDY_NITROFUSION2_REPO = "https://huggingface.co/ChenDY/NitroFusion2";
 // Klein ships Apache-2.0 and ungated, unlike FLUX.1-dev and FLUX.2-dev, so
 // there is deliberately no licenseGate on the stack below -- the setup pack can
 // point straight at the file with no click-through.
@@ -386,6 +390,32 @@ const Z_IMAGE_TURBO_STACK = [
   }
 ] as const;
 
+/**
+ * NitroFusion2 swaps only the diffusion model: it is Z_image_Turbo's own
+ * architecture (the bf16 file is byte-for-byte the same size) with
+ * adversarially distilled weights, so it reads the same qwen_3_4b encoder
+ * through CLIPLoader type lumina2 and decodes through the same ae VAE. The two
+ * shared entries are taken from Z_IMAGE_TURBO_STACK itself rather than copied,
+ * so the setup pack de-duplicates them and a user who already runs Z_image_Turbo
+ * downloads only the 12.3 GB model.
+ */
+const NITROFUSION2_STACK = [
+  {
+    kind: "diffusion-model-stack",
+    objectInfoNode: "UNETLoader",
+    inputName: "unet_name",
+    label: "NitroFusion2 diffusion model",
+    modelName: "nf2-bf16.safetensors",
+    setupHint: "Install nf2-bf16.safetensors where ComfyUI's UNETLoader can find it.",
+    downloadUrl: `${CHENDY_NITROFUSION2_REPO}/resolve/main/nf2-bf16.safetensors`,
+    sourcePageUrl: CHENDY_NITROFUSION2_REPO,
+    downloadSizeBytes: 12309866400,
+    sha256: "a1e11422d84c1986ba106f892eda86b955b752627ec902aefff1fb939c7ac19c"
+  },
+  Z_IMAGE_TURBO_STACK[1],
+  Z_IMAGE_TURBO_STACK[2]
+] as const;
+
 /**
  * ControlNet-as-model-patch weights for the Z_image_Turbo stack, applied
  * through the core `ModelPatchLoader` / `ZImageFunControlnet` pair rather than
@@ -2589,6 +2619,30 @@ const Z_IMAGE_TURBO_TXT2IMG_CAPABILITY: WorkflowCapability = {
   }
 };
 
+const NITROFUSION2_TXT2IMG_CAPABILITY: WorkflowCapability = {
+  toolType: "txt2img",
+  loaderType: "diffusion-model-stack",
+  artistLabel: "Text to Image",
+  technicalLabel: "txt2img-nitrofusion2",
+  requiredPhotoshopInputs: [],
+  controls: ["prompt", "negativePrompt", "width", "height", "steps", "guidance", "seed"],
+  output: {
+    kind: "full-image",
+    size: "preset",
+    importBehavior: "new-layer"
+  },
+  uiHints: {
+    showModelSelector: true,
+    modelSelectorLabel: "NitroFusion2 model",
+    primaryActionLabel: "Generate",
+    // CFG 1 means the sampler never reads the negative prompt, so a field for
+    // it would only look like it does something.
+    hiddenControls: ["negativePrompt"],
+    experimentalNote:
+      "NitroFusion2 is a one-step version of Z_image_Turbo: 1 step at CFG 1, at 1024 x 1024. Keep steps at 1 or 2. More steps do not improve it; at 4 it starts to speckle. It is fastest when you re-roll the same prompt with a new seed, because ComfyUI keeps the prompt encoded."
+  }
+};
+
 const Z_IMAGE_TURBO_IMG2IMG_CAPABILITY: WorkflowCapability = {
   toolType: "img2img",
   loaderType: "diffusion-model-stack",
@@ -5197,6 +5251,79 @@ export const WORKFLOW_PRESETS: WorkflowPresetDefinition[] = [
     compatibilityNote:
       "Z_image_Turbo is a diffusion model stack, not a checkpoint. OpenLayer loads it through UNETLoader, CLIPLoader, and VAELoader."
   },
+  {
+    id: "txt2img-nitrofusion2",
+    label: "txt2img-nitrofusion2",
+    displayName: "NitroFusion2",
+    mode: "txt2img",
+    description: "One-step text-to-image preset for NitroFusion2, a distilled fine-tune of Z_image_Turbo.",
+    workflowFile: "workflows/api/txt2img-nitrofusion2.json",
+    status: "stable",
+    recommendedSettings: { steps: 1, cfg: 1, width: 1024, height: 1024 },
+    supportedModelFamilies: ["zImage"],
+    experimentalModelFamilies: ["unknown"],
+    modelSource: DIFFUSION_MODEL_SOURCE,
+    capability: NITROFUSION2_TXT2IMG_CAPABILITY,
+    modelStack: [...NITROFUSION2_STACK],
+    requiredModels: [...NITROFUSION2_STACK],
+    // The graph keeps the Z_image_Turbo graph's node ids, so its injection map
+    // is the same one; only the model file and sampler settings differ.
+    injections: Z_IMAGE_TURBO_TXT2IMG_INJECTIONS,
+    requiredNodes: [
+      {
+        id: Z_IMAGE_TURBO_TXT2IMG_NODES.diffusionModelLoader,
+        classType: "UNETLoader",
+        requiredInputs: ["unet_name", "weight_dtype"]
+      },
+      {
+        id: Z_IMAGE_TURBO_TXT2IMG_NODES.clipLoader,
+        classType: "CLIPLoader",
+        requiredInputs: ["clip_name", "type"]
+      },
+      {
+        id: Z_IMAGE_TURBO_TXT2IMG_NODES.vaeLoader,
+        classType: "VAELoader",
+        requiredInputs: ["vae_name"]
+      },
+      {
+        id: Z_IMAGE_TURBO_TXT2IMG_NODES.modelSampling,
+        classType: "ModelSamplingAuraFlow",
+        requiredInputs: ["model", "shift"]
+      },
+      {
+        id: Z_IMAGE_TURBO_TXT2IMG_NODES.positivePrompt,
+        classType: "CLIPTextEncode",
+        requiredInputs: ["text", "clip"]
+      },
+      {
+        id: Z_IMAGE_TURBO_TXT2IMG_NODES.negativePrompt,
+        classType: "CLIPTextEncode",
+        requiredInputs: ["text", "clip"]
+      },
+      {
+        id: Z_IMAGE_TURBO_TXT2IMG_NODES.latentImage,
+        classType: "EmptySD3LatentImage",
+        requiredInputs: ["width", "height", "batch_size"]
+      },
+      {
+        id: Z_IMAGE_TURBO_TXT2IMG_NODES.sampler,
+        classType: "KSampler",
+        requiredInputs: ["model", "seed", "steps", "cfg", "sampler_name", "scheduler", "positive", "negative", "latent_image", "denoise"]
+      },
+      {
+        id: Z_IMAGE_TURBO_TXT2IMG_NODES.decode,
+        classType: "VAEDecode",
+        requiredInputs: ["samples", "vae"]
+      },
+      {
+        id: Z_IMAGE_TURBO_TXT2IMG_NODES.saveImage,
+        classType: "SaveImage",
+        requiredInputs: ["images"]
+      }
+    ],
+    compatibilityNote:
+      "txt2img-nitrofusion2 is the txt2img-z-image-turbo graph with three values changed, matching the authors' own one-step ComfyUI workflow: nf2-bf16.safetensors in UNETLoader, 1 step instead of 4, and euler instead of res_multistep, with the same ModelSamplingAuraFlow shift 3, CFG 1 and simple scheduler. Those values are the point of the preset rather than a tuning choice. On 2026-11-09, running NitroFusion2 through the Z_image_Turbo preset's own settings (4 steps, res_multistep) on six prompts added white splatter and speckle that the one-step run did not have, which is why the model is its own preset and not just a file picked in Z_image_Turbo. Measured on the same day on a 12 GB RTX 4070 Ti at 1024 x 1024: about 2.5 s against Z_image_Turbo's 5 s when only the seed changes, but 14 s against 17.5 s for a new prompt, because the 8 GB encoder and the 12.3 GB model do not fit on the card together and ComfyUI swaps them for every new prompt. The authors' two-step graph is not used: it needs their de-distillation LoRA switched on for the first step only, across two KSamplerAdvanced nodes, which OpenLayer's LoRA insertion cannot express. It declares no loraInsertion either: no Z_image_Turbo LoRA has been tried on the one-step weights, and the Z_image_Turbo preset stays the place for LoRAs until one has. GGUF quantisations exist but need the ComfyUI-GGUF loader and are not wired up here."
+  },
   {
     id: "img2img-z-image-turbo",
     label: "img2img-z-image-turbo",

src/comfy/types.ts

@@ -7,6 +7,7 @@ export type WorkflowPreset =
   | "img2img-flux2-klein"
   | "edit-flux2-klein"
   | "txt2img-z-image-turbo"
+  | "txt2img-nitrofusion2"
   | "img2img-z-image-turbo"
   | "txt2img-krea2-turbo"
   | "img2img-krea2-turbo"

src/comfy/workflowBuilder.ts

@@ -3,6 +3,7 @@ import img2imgBasicWorkflow from "../workflows/api/img2img-basic.json";
 import txt2imgFlux1DevFp8Workflow from "../workflows/api/txt2img-flux1-dev-fp8.json";
 import txt2imgFlux2DevGgufWorkflow from "../workflows/api/txt2img-flux2-dev-gguf.json";
 import txt2imgZImageTurboWorkflow from "../workflows/api/txt2img-z-image-turbo.json";
+import txt2imgNitroFusion2Workflow from "../workflows/api/txt2img-nitrofusion2.json";
 import img2imgZImageTurboWorkflow from "../workflows/api/img2img-z-image-turbo.json";
 import txt2imgKrea2TurboWorkflow from "../workflows/api/txt2img-krea2-turbo.json";
 import img2imgKrea2TurboWorkflow from "../workflows/api/img2img-krea2-turbo.json";
@@ -73,6 +74,7 @@ const WORKFLOW_TEMPLATES: Partial<Record<WorkflowPreset, ComfyWorkflow>> = {
   "txt2img-flux1-dev-fp8": txt2imgFlux1DevFp8Workflow as ComfyWorkflow,
   "txt2img-flux2-dev-gguf": txt2imgFlux2DevGgufWorkflow as ComfyWorkflow,
   "txt2img-z-image-turbo": txt2imgZImageTurboWorkflow as ComfyWorkflow,
+  "txt2img-nitrofusion2": txt2imgNitroFusion2Workflow as ComfyWorkflow,
   "img2img-z-image-turbo": img2imgZImageTurboWorkflow as ComfyWorkflow,
   "txt2img-krea2-turbo": txt2imgKrea2TurboWorkflow as ComfyWorkflow,
   "img2img-krea2-turbo": img2imgKrea2TurboWorkflow as ComfyWorkflow,

src/workflows/api/txt2img-nitrofusion2.json

@@ -0,0 +1,141 @@
+{
+  "20": {
+    "class_type": "UNETLoader",
+    "inputs": {
+      "unet_name": "nf2-bf16.safetensors",
+      "weight_dtype": "default"
+    },
+    "_meta": {
+      "title": "Load NitroFusion2 Diffusion Model"
+    }
+  },
+  "21": {
+    "class_type": "CLIPLoader",
+    "inputs": {
+      "clip_name": "qwen_3_4b.safetensors",
+      "type": "lumina2",
+      "device": "default"
+    },
+    "_meta": {
+      "title": "Load Qwen CLIP"
+    }
+  },
+  "22": {
+    "class_type": "VAELoader",
+    "inputs": {
+      "vae_name": "ae.safetensors"
+    },
+    "_meta": {
+      "title": "Load AE VAE"
+    }
+  },
+  "23": {
+    "class_type": "ModelSamplingAuraFlow",
+    "inputs": {
+      "model": [
+        "20",
+        0
+      ],
+      "shift": 3
+    },
+    "_meta": {
+      "title": "NitroFusion2 Sampling"
+    }
+  },
+  "5": {
+    "class_type": "EmptySD3LatentImage",
+    "inputs": {
+      "width": 1024,
+      "height": 1024,
+      "batch_size": 1
+    },
+    "_meta": {
+      "title": "Image Size"
+    }
+  },
+  "6": {
+    "class_type": "CLIPTextEncode",
+    "inputs": {
+      "text": "",
+      "clip": [
+        "21",
+        0
+      ]
+    },
+    "_meta": {
+      "title": "Positive Prompt"
+    }
+  },
+  "7": {
+    "class_type": "CLIPTextEncode",
+    "inputs": {
+      "text": "",
+      "clip": [
+        "21",
+        0
+      ]
+    },
+    "_meta": {
+      "title": "Negative Prompt"
+    }
+  },
+  "3": {
+    "class_type": "KSampler",
+    "inputs": {
+      "seed": 1,
+      "steps": 1,
+      "cfg": 1,
+      "sampler_name": "euler",
+      "scheduler": "simple",
+      "denoise": 1,
+      "model": [
+        "23",
+        0
+      ],
+      "positive": [
+        "6",
+        0
+      ],
+      "negative": [
+        "7",
+        0
+      ],
+      "latent_image": [
+        "5",
+        0
+      ]
+    },
+    "_meta": {
+      "title": "Sampler"
+    }
+  },
+  "8": {
+    "class_type": "VAEDecode",
+    "inputs": {
+      "samples": [
+        "3",
+        0
+      ],
+      "vae": [
+        "22",
+        0
+      ]
+    },
+    "_meta": {
+      "title": "Decode"
+    }
+  },
+  "9": {
+    "class_type": "SaveImage",
+    "inputs": {
+      "filename_prefix": "OpenLayer_NitroFusion2",
+      "images": [
+        "8",
+        0
+      ]
+    },
+    "_meta": {
+      "title": "Save Image"
+    }
+  }
+}