ComfyUI_TwinFlow

TwinFlow: Realizing One-step Generation on Large Models with Self-adversarial Flows - use it in ComfyUI

Update

• add z-image dype support，infer 3840*2160 use 2 step，need more Vram
• 新增 z的dype支持，1步质量一般，超大尺寸需要多步数，或许也会崩，当然显存占用也会增大；

Previous

• Fix if diffusers>0.36.0 got flash attn can't use attn mask error and MAC can't use attn2 (z-image only) / 高版本或开发版的diffuser的z image 其attn mask 会报错（已知的官方错误），改成常规的，MAC的attn问题可能也是在这里，做了简单的修复；
• Fix lora can't use error 修复lora无法使用的问题，加入lora adapter 清理代码，避免混用，可以参考示例图或新的lora节点
• upload new z-image gguf model，type is lumia2 now 上传新的z image gguf ，减少大小，解决context 都设置为f16的错误，类型改成lunmia2 ，跟city96大佬的一致；
• Now supports 1-step or any number of steps for image generation. Thanks to QAQdev for the code support - please give them a star!

Requirements

• diffusers >= 0.36.0 (required for Z-Image support)

Tips

• LoRA support added via @oliveagle PR - use LoRA when inferring with 4 steps (untested, no guarantees)
• Offload modes:
  • clip: Only unload ComfyUI CLIP (recommended for most cases)
  • none: Don't unload anything (use if running many prompts repeatedly)
  • all: Unload all models
• Z-Image and Qwen-Image GGUF support - Qwen-Image GGUF has been re-quantized, please update to avoid dtype errors
• Z-Image Q8 performance:
  • 12GB VRAM: 1024x768 in 2-3s/image (without offloading)
  • 24GB VRAM: ~1.3s/image
• Qwen-Image performance:
  • 12GB VRAM (50 blocks): 1024x768 in ~15s/image with GPU offloading
• If VRAM > 16GB, set block number to 0 for maximum inference speed

1. Installation

In the ./ComfyUI/custom_nodes directory, run:

git clone https://github.com/smthemex/ComfyUI_TwinFlow

2. Requirements

pip install -r requirements.txt

3. Checkpoints

• 3.1 Qwen-Image GGUF - GGUF format only (safetensors available from other developers)
• 3.2 Z-Image GGUF and safetensors
• 3.3 Qwen-Image VAE & CLIP
• 3.4 Z-Image VAE & CLIP

├── 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

4. Example

• z-image dype

• qwen-image lora

• z-image lora

5. 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}
}