[Feature] TwinFlow: Qwen Image and Z-Image-Turbo

[JohnLoveJoy]

Feature Summary

TwinFlow: Realizing One-step Generation on Large Models with Self-adversarial Flows

Detailed Description

It's quite impressive how they made Qwen Image and ZIT exponentially faster.

https://huggingface.co/inclusionAI/TwinFlow-Z-Image-Turbo
https://huggingface.co/inclusionAI/TwinFlow
https://github.com/inclusionAI/TwinFlow

https://zhenglin-cheng.com/twinflow/

"Recent advances in large multi-modal generative models have demonstrated impressive capabilities in multi-modal generation, including image and video generation. These models are typically built upon multi-step frameworks like diffusion and flow matching, which inherently limits their inference efficiency, requiring 40-100 Number of Function Evaluations (NFEs). While various few-step methods aim to accelerate the inference, existing solutions have clear limitations. Prominent distillation-based methods, such as progressive and consistency distillation, either require an iterative distillation procedure or show significant degradation at very few steps (< 4-NFE). Meanwhile, integrating adversarial training into distillation (e.g., DMD/DMD2 and SANA-Sprint) to enhance performance introduces training instability, added complexity, and high GPU memory overhead due to the auxiliary trained models. To this end, we propose TwinFlow, a simple yet effective framework for training 1-step generative models that bypasses the need for fixed pretrained teacher models and avoids standard adversarial networks during training, making it ideal for building large-scale, efficient models. On text-to-image tasks, our method achieves a GenEval score of 0.83 in 1-NFE, outperforming strong baselines like SANA-Sprint (a GAN loss-based framework) and RCGM (a consistency-based framework). Notably, we demonstrate the scalability of TwinFlow by full-parameter training on Qwen-Image-20B and transform it into an efficient few-step generator. With just 1-NFE, our approach matches the performance of the original 100-NFE model on both the GenEval and DPG-Bench benchmarks, reducing computational cost by 100× with minor quality degradation."

Alternatives you considered

No response

Additional context

No response

[limingchina]

I've tried this model: https://huggingface.co/azazeal2/TwinFlow-Z-Image-Turbo-repacked/blob/main/ComfyUI/TwinFlow_Z_Image_Turbo_exp_bf16.safetensors

There are some error logging. However it can still generate an image in 2 steps with minor flaws for one prompt. And it works completely fine with some other prompts. However, the images generated with 2 steps are still a bit blurry.

[DEBUG] model.cpp:1373 - loading tensors from ComfyUI/models/diffusion_models/TwinFlow_Z_Image_Turbo_exp_bf16.safetensors
|====================> | 443/1099 - 17.52it/s[INFO ] model.cpp:1611 - unknown tensor 'model.diffusion_model.t_embedder_2.mlp.0.bias | bf16 | 1 [1024, 1, 1, 1, 1]' in model file
[INFO ] model.cpp:1611 - unknown tensor 'model.diffusion_model.t_embedder_2.mlp.0.weight | bf16 | 2 [256, 1024, 1, 1, 1]' in model file
[INFO ] model.cpp:1611 - unknown tensor 'model.diffusion_model.t_embedder_2.mlp.2.bias | bf16 | 1 [256, 1, 1, 1, 1]' in model file
[INFO ] model.cpp:1611 - unknown tensor 'model.diffusion_model.t_embedder_2.mlp.2.weight | bf16 | 2 [1024, 256, 1, 1, 1]' in model file

[wbruna]

I've tried this model: https://huggingface.co/azazeal2/TwinFlow-Z-Image-Turbo-repacked/blob/main/ComfyUI/TwinFlow_Z_Image_Turbo_exp_bf16.safetensors

Indeed, seems to work fine, even after quantization. An extra step seems to help with the blurriness:

[JohnLoveJoy]

I've tried this model: https://huggingface.co/azazeal2/TwinFlow-Z-Image-Turbo-repacked/blob/main/ComfyUI/TwinFlow_Z_Image_Turbo_exp_bf16.safetensors

Indeed, seems to work fine, even after quantization. An extra step seems to help with the blurriness:

Did you test it on SD.cpp?

[wbruna]

Did you test it on SD.cpp?

Yes, I've generated those images with the current release (master-453-4ff2c8c).

[limingchina]

I'm using very recent master (about 2 days ago) as well.

Compiled myself and run "bin/sd_cli".

[JohnLoveJoy]

That's great, but I'm having trouble converting to Q5 or Q4. The link above is down.
I tried this one, but the strange errors made me lose patience: https://huggingface.co/smthem/TwinFlow-Z-Image-Turbo-diffuser-gguf/tree/main

[wbruna]

That's great, but I'm having trouble converting to Q5 or Q4

Try this one: https://huggingface.co/wbruna/TwinFlow-Z-Image-Turbo-sdcpp-GGUF

[JohnLoveJoy]

Thank you. It works, although the output image has a lot of noise even with 4 steps.

[limingchina]

@JohnLoveJoy , have you compared it with the the original zimage with the same quantization? With the same steps, sampler and scheduler, does it have any improvement? I haven't tried it much yet. Just one image with 2 steps as the model advertises, the result is not that impressive.

[wbruna]

Thank you. It works, although the output image has a lot of noise even with 4 steps.

https://huggingface.co/azazeal2/TwinFlow-Z-Image-Turbo-repacked/discussions/1 also mentions high frequency noise in the outputs.

I've had some success fiddling with the flow-shift parameter, and different schedulers. This is with --sampling-method euler --steps 3 --scheduler smoothstep --flow-shift 0.5:

[JohnLoveJoy]

It seems the original model generates images with less noise and fewer artifacts than this one; I tested it using both 3 and 4 steps.

Twinflow + your parameters, 3steps >

Z Image Turbo Q4