This is a poor copy of the following excellent application:

https://huggingface.co/spaces/tori29umai/Qwen-Image-2509-CharacterSheet

• You can switch between FP8 and GGUF (models and clips).

• The lightning 4-step is attractive, but the 8-step is more stable.

• If the output is mosaic, remove the "fast" option from Comfyui's bat.

• This is usually where I start organizing for a release, but I run out of energy and it ends up being shelved. So I'll upload it as is.

• I tried to make it using as many "normal" nodes as possible, but if you can't find it, please let me know.

• (10/25) I tried to create a setting for 12GB of VRAM, but ended up with a setting of 6GB.

Time taken on 4070tis (second pass)

qwen_image_edit_2509_fp8_e4m3fn.safetensors (19GB)

qwen_2.5_vl_7b_fp8_scaled.safetensors (9GB)

Uses 15.2GB of VRAM.

140sec @ 8step w/Qwen-Image-Lightning-8steps-V2.0-bf16.safetensors

88sec @ 4step w/Qwen-Image-Lightning-4steps-V2.0-bf16.safetensors

=====

Qwen-Image-Edit-2509-Q3_K_M.gguf (9GB)

qwen_2.5_vl_7b_fp8_scaled.safetensors (9GB)

Uses 12.1GB of VRAM (low quality)

105sec @ 4step w/Qwen-Image-Lightning-4steps-V2.0-bf16.safetensors

=====

Qwen-Image-Edit-2509-Q3_K_M.gguf (9GB)

Qwen2.5-VL-7B-Instruct-Q3_K_M.gguf (3.5GB)

Uses 13.8GB of VRAM (low quality. There is no point in using GGUF as the text encoder.)

105sec @ 4step w/Qwen-Image-Lightning-4steps-V2.0-bf16.safetensors

=====

Qwen-Image-Edit-2509-Q5_K_M.gguf (14GB)

qwen_2.5_vl_7b_fp8_scaled.safetensors (9GB)

DisTorch2MultiGPU virtual vram : 10GB

Uses 7.1GB of VRAM.

=====

Qwen-Image-Edit-2509-Q4_K_M.gguf (12GB)

qwen_2.5_vl_7b_fp8_scaled.safetensors (9GB)

DisTorch2MultiGPU virtual vram : 10GB

Uses 5.2GB of VRAM. / 28GB MEM

125sec @ 4step w/Qwen-Image-Lightning-4steps-V2.0-bf16.safetensors

=====

Qwen-Image-Edit-2509-Q3_K_M.gguf (9GB)

qwen_2.5_vl_7b_fp8_scaled.safetensors (9GB)

DisTorch2MultiGPU virtual vram : 10GB

Uses 3.1GB of VRAM. (Uses 5.2GB of VRAM @ VAE)

=====