TwinFlow: Realizing One-step Generation on Large Models with Self-adversarial Flows — running on ComfyUI's native DiT inference flow for both Z-Image and Qwen-Image.

Update

Use comfyUI core to fit twinflow /使用comfyUI核心来适配twinflow,避免diffuser版本问题

Checkpoints

├── ComfyUI/models/diffusion_models/ 
|     ├── TwinFlow-Qwen-Image-v1.0-int8_convrot.safetensors # or Q6-K, Q8_0, int8_convrot, w4a8 推荐
|     ├── TwinFlow-Z-Image-Turbo-comfy_int8_convrot.safetensors # or Q6-K, Q8_0, BF16
├── ComfyUI/models/gguf
|     ├── TwinFlow-Qwen-Image-diffusers-Q6_K.gguf  # or Q8_0、 BF16
|     ├── TwinFlow-Z-Image-Turbo-diffuser-Q8_0.gguf # or Q6-k，BF16
├── ComfyUI/models/vae
|     ├──qwen_image_vae.safetensors
|     ├──ae.safetensors #z-image use flux vae
├── ComfyUI/models/clip 
|     ├──qwen_2.5_vl_7b_fp8_scaled.safetensors # or bf16
|     ├──qwen_3_4b.safetensors # z image

Example

qwen-image
z-image

Notes

The official TwinFlow demo runs without CFG (cfg_scale=0.0, "should be zero") for TwinFlow-Qwen-Image. In ComfyUI terms that means setting the KSampler CFG = 1.0 (ComfyUI's cfg=1 skips the uncond pass; cfg=0 would output the unconditional prediction instead). CFG > 1 still runs but deviates from the official configuration — with 1–2 step generation there are no later steps to absorb the distortion, so quality usually degrades.
No shift node is needed: the TwinFlow schedule replaces ComfyUI's sigma schedule entirely (ModelSamplingFlux / ModelSamplingDiscreteFlow shift only shapes the stock schedule, and the DiT receives the raw sigma either way).

Citation

@article{cheng2025twinflow,
  title={TwinFlow: Realizing One-step Generation on Large Models with Self-adversarial Flows},
  author={Cheng, Zhenglin and Sun, Peng and Li, Jianguo and Lin, Tao},
  journal={arXiv preprint arXiv:2512.05150},
  year={2025}
}