Please make it comfyUI compatible like Z image.

#2
by krigeta - opened

please make a version for comfyUI like Z image.

Alibaba-PAI org

Could you please test this? I've adapted a ComfyUI version in VideoX-Fun:
https://github.com/aigc-apps/VideoX-Fun/tree/main/comfyui/qwenimage

the problem with these node is they are not compatible with comfyUI core nodes and they are not usable in custom workflows.

Hey @bubbliiiing please give us some hope mate.

The adaptation works, but doesn't use ComfyUI's diffusion model loader, or the KSampler.

This prevents full usage of LORAs, such as the lightx2v/Qwen-Image-2512-Lightning 4 and 8 step Loras, style loras, and character LORAs. The LORA capability in the VideoX-Fun nodes seems to work randomly.

Not using the KSampler (as well as RES4LYF's more complex samplers) prevents usage of the full set of sampler and scheduler types.

Additionally, loading models through ComfyUI's nodes is far faster.

The ideal would be an update to the Apply ControlNet node.

Qwen has amazing prompt adherence compared to other models out there, especially when an controlnet is applied, so it would be amazing to get your controlnet model going with Qwen Image 2512.

I'm compensating for now by applying controlnets on Qwen Image (original) with the KSampler, and then unsampling and resampling the resultant image through the ClownsharkSampler and chain sampler on Qwen Image 2512 using res_2s and bong_tangent with a style LORA and the 8-step lightning LORA. This works most of the time, but not all of the time.

created the solution for now and it is working perfectly with core nodes:

https://github.com/Comfy-Org/ComfyUI/pull/12359

krigeta changed discussion status to closed

Sign up or log in to comment