Merge qwenImage2512 with 2 LoRAs; add the LoRA https://huggingface.co/Wuli-art/Qwen-Image-2512 -Turbo-LoRA/blob/main/Wuli-Qwen-Image-2512-Turbo-LoRA-4steps-V3.0-bf16.safetensors, use with simple Euler in 8 steps, on an RTX 3060, generated in 30 seconds, 1024×1536
https://huggingface.co/lightx2v/Qwen-Image-Lightning/tree/main
Image generated with SD Forge Neo Classic – don't ask me about ComfyUI workflows, I don't use it
Description
FAQ
Comments (3)
[COMFYUI] If someone tests this with a working workflow, I appreciate. For now:
- Trying to use 'checkpoint loader' (it wont have VAE, you have to load external VAE), or UNET loader, it gives error just after Negative Prompt and Positive prompt processing.
Using 'CLIPLoaderDisTorch2MultiGPU' and variants: ValueError: too many values to unpack (expected 4)
Using 'load clip' node with type 'qwen-image': "mat1 and mat2 shapes cannot be multiplied (1x3072 and 18432x3072)"
Have you tried the UNET Loader? Put the model in /models/diffusion_models. I can load it but i just get a black image as a result, with deis / kl_optimal, steps 25, cfg 2.5: those are the settings i use successfully with other Qwen models, so idk.
Just be careful if using --force-fp16 and --fp16-vae in BAT FILE. In certain COMFYUI environments, that gives black screen.