README.md

metadata

license: apache-2.0
base_model: ChenDY/NitroFusion2
base_model_relation: quantized
pipeline_tag: text-to-image
tags:
  - comfyui
  - nitrofusion2
  - z-image-turbo
  - one-step
  - few-step
  - int8
  - w6a8
  - w4a8
  - gptq
  - quantized

NitroFusion2 for ComfyUI: int8, W6A8, W4A8

NitroFusion2 (Dar-Yen Chen and Yi-Zhe Song), the 1–2 step distillation of Z-Image-Turbo, in three smaller formats that run in stock ComfyUI: no custom nodes, the normal Load Diffusion Model node, and ComfyUI's own int8 / W6A8 / W4A8 kernels. They are about 1.7× faster per step than bf16 on an RTX 4090 and half to a quarter of its size. All files are made from the publisher's nf2-bf16.safetensors (sha256 a1e11422…).

Use the GPTQ files. They have the same size, layout and speed as the plain conversions and are closer to bf16 in every format. For 2 steps, each GPTQ file has a step-1 file (the publisher's de-distillation LoRA merged in before quantizing), which cuts the 2-step error by another 21–33%.

Format	Size	1 step, 4090	Error, 1 step	Error, 2 steps with the LoRA	Error, 2 steps with the step-1 file
bf16 (original)	12.31 GB	0.50 s	0	0	–
int8 GPTQ	6.18 GB	0.29 s	11.3%	8.7%	6.9%
W6A8 GPTQ	5.02 GB	0.30 s	13.8%	12.9%	8.6%
W4A8 GPTQ	3.49 GB	0.29 s	25.8%	24.0%	18.5%
int8, plain	6.18 GB	0.29 s	13.4%	9.3%	–
W6A8, plain	5.02 GB	0.30 s	17.5%	14.0%	–
W4A8, plain	3.49 GB	0.29 s	33.6%	26.0%	–

Files, in diffusion_models/:

• nitrofusion2_int8_convrot_gptq.safetensors, nitrofusion2_w6a8_gptq.safetensors, nitrofusion2_w4a8_gptq.safetensors: the GPTQ files, for 1 step and for step 2 of 2
• nitrofusion2_int8_convrot_gptq_step1.safetensors, nitrofusion2_w6a8_gptq_step1.safetensors, nitrofusion2_w4a8_gptq_step1.safetensors: their step-1 files (same size each), for step 1 of 2
• nitrofusion2_int8_convrot.safetensors, nitrofusion2_w6a8.safetensors, nitrofusion2_w4a8.safetensors: the plain conversions published first

Time is the sampler at 1024×1024 (ComfyUI master fb2b26fda, comfy-kitchen 0.2.41), measured on the plain files; the GPTQ files use the same layout and kernels. Error is the relative error of the model's prediction against bf16 for the same input: 12 prompts × 4 noise samples at 1024×1024, the 2-step figures averaged over both steps. A one-step prediction from pure noise reacts strongly to small weight changes, and nearly all of that error is fine texture: the pictures keep bf16's composition, faces and text (below). Use the numbers to rank the files against each other, not to compare with other models.

Use

1. Put the files in ComfyUI/models/diffusion_models/: one GPTQ file for 1 step; the same file and its _step1 file for 2 steps.
2. The text encoder and VAE are Z-Image-Turbo's: qwen_3_4b.safetensors in text_encoders/ and ae.safetensors in vae/.
3. Open a workflow from example_workflows/ (int8 selected; pick another file in the subgraph's loaders):
   • 1 step: the publisher's 1-step workflow.
   • 2 steps with a step-1 file: the publisher's 2-step workflow with a second Load Diffusion Model for step 1 in place of the LoRA. No LoRA file is needed.

Settings as published: euler / simple, CFG 1; 1 step at shift 3; 2 steps at shift 6.

Why a step-1 file. The official 2-step workflow applies the de-distillation LoRA at 0.5 on step 1. On a quantized file ComfyUI re-quantizes the LoRA-patched weights with fresh scales and random rounding, so step 1 loses the calibration (the plain and GPTQ files end up nearly equal there). The step-1 file has the LoRA merged in before quantizing and is calibrated on step 1 itself. Both files are loaded during a 2-step run (W4A8: 7.9 GB VRAM peak). The LoRA route still works with every file: the publisher's 2-step workflow with nf2-dedistill-lora.safetensors in loras/. The plain 1-step workflow is there too.

Pictures

One step with the GPTQ files, seed 42, 1024×1024, the same prompt and noise for every file (top: bf16, int8; bottom: W6A8, W4A8). The plain files' version of each sheet is in assets/ (the same name without _gptq).

How they were made

ComfyUI's own quantization formats (comfy-kitchen):

• What is quantized: all 202 linear layers of the 30 main blocks and the 4 refiner blocks (attention qkv and out, the three feed-forward layers, adaLN): 6.142B of the model's 6.155B weights. Embedders, norms and the final layer stay bf16.
• int8: ConvRot (block Hadamard rotation of 256), one scale per row, 8-bit activations.
• W6A8 / W4A8: 6-bit / 4-bit weights in the rotated basis with an FP8 scale per group of 16, 8-bit activations.
• Plain files: comfy-kitchen's own conversion, no calibration.
• GPTQ files: the rotation and rounding in fp32, a search over each group's FP8 scale, then GPTQ in the rotated basis (act order, damping 0.01), every weight snapped to the exact value the kernel decodes. Calibration: 48 prompts that are not among the evaluation prompts, at about one megapixel, with every model call of both official modes.
• Step-1 files: the de-distillation LoRA merged into bf16 at 0.5 with ComfyUI's own patch code (the merged model matches bf16 + LoRA exactly), then the same GPTQ on statistics of step 1 alone.

License and credit

Apache-2.0, as the original: LICENSE, changes listed in NOTICE. The model is NitroFusion2 by Dar-Yen Chen and Yi-Zhe Song, built on Z-Image-Turbo by Tongyi-MAI; the workflows are the publisher's; the step-1 files contain the publisher's de-distillation LoRA, merged. Paper: arXiv 2611.05648.

@article{chen2026nitrofusion2,
    title={NitroFusion2: Bringing Adversarial Loss to the Text Stream of Multimodal Diffusion Transformers},
    author={Chen, Dar-Yen and Song, Yi-Zhe},
    journal={arXiv preprint arXiv:2611.05648},
    year={2026}
}