PedroMarinhoDev/NitroFusion2-ComfyUI

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 files 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 three are converted from the publisher's nf2-bf16.safetensors (sha256 a1e11422…).

Format	Size	1 step, 4090	Error, 1 step	Error, 2 steps
bf16 (original)	12.31 GB	0.50 s	0	0
int8	6.18 GB	0.29 s	13.4%	9.3%
W6A8	5.02 GB	0.30 s	17.5%	14.0%
W4A8	3.49 GB	0.29 s	33.6%	26.0%

Files, in diffusion_models/:

• nitrofusion2_int8_convrot.safetensors: int8 weights and activations
• nitrofusion2_w6a8.safetensors: 6-bit weights, 8-bit activations
• nitrofusion2_w4a8.safetensors: 4-bit weights, 8-bit activations

Time is the sampler at 1024×1024 (ComfyUI master fb2b26fda, comfy-kitchen 0.2.41). 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 figure 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 three files against each other, not to compare with other models.

Use

1. Put one file in ComfyUI/models/diffusion_models/.
2. The text encoder and VAE are Z-Image-Turbo's: qwen_3_4b.safetensors in text_encoders/ and ae.safetensors in vae/.
3. For two steps, also the publisher's de-distillation LoRA nf2-dedistill-lora.safetensors in loras/.
4. Open a workflow from example_workflows/: the publisher's official 1-step and 2-step workflows with the int8 file selected. Pick another file in the subgraph's unet_name.

Settings as published: euler / simple, CFG 1; 1 step at shift 3; 2 steps at shift 6 with the LoRA at 0.5 on the first step only. The LoRA works on all three files.

Pictures

One step, seed 42, 1024×1024, the same prompt and noise for every file (top: bf16, int8; bottom: W6A8, W4A8).

How they were made

Plain conversion with ComfyUI's own quantization formats (comfy-kitchen), no calibration:

• 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.

Calibrated (GPTQ) versions of the same three formats will follow here once they are tuned.

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. 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}
}