I was trying to create Qwen-Image Workflow ASAP ;)
get GGUF: https://huggingface.co/city96/Qwen-Image-gguf/tree/main
get VAE and CLIP: https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/tree/main/split_files
Update: there are at least two Abliterated (Uncensored to a degree) GGUF versions of CLIP text-encoder for this model:
1) https://huggingface.co/mradermacher/Qwen2.5-VL-7B-Instruct-abliterated-GGUF/tree/main
2) https://huggingface.co/mradermacher/Qwen2.5-VL-7B-Abliterated-Caption-it-GGUF/tree/main <--- this one is my personal favorite!
3) for 4-8 steps lora go to this link: https://huggingface.co/lightx2v/Qwen-Image-Lightning/tree/main
P.S. The model is very sensitive to photography settings. Try to be careful with the depth of field and shallow focus in your prompts.
Description
update your Comfyui for fresh GGUF node version!
FAQ
Comments (18)
how much vram is needed?
12 Gb is enough
Depends on which GGUF quant you pick.
everydream All GGUF quants eventually fill the same VRAM: from Q_4 to Q_8 - but the problem is with larger ones - computation takes more time and loading of it from HDD to RAM also sux
OliviaRossi If they're the same size in VRAM something is terribly wrong with GGUF loading in comfy.
everydream comfyui is constantly evolving - what today is ok - the next day (after update) can be as shit ... and vice versa
Not sure why, but my 16GB GPU crashed loading this workflow. Pretty sure it's OOM since it happens when the GPU memory utilization hits peak. I was using Q4_K_M which should be as small as it can get :(
after 10th step turns image to pitch black
dunno why it does in your case - the only advice - update your ComfyUI!!!
OliviaRossi I did!!!) Otherwise, I wouldn't have been able to start generating at all..
Randmeist there is an update from reddit about the black image: https://www.reddit.com/r/comfyui/comments/1mi4ur9/comment/n72ym1h/?utm_source=share&utm_medium=web3x&utm_name=web3xcss&utm_term=1&utm_content=share_button
OliviaRossi it was --use-sage-attention, thanks!
Randmeist just wanted to write: remove the --use-sage-attention argument but you did it urself ;)
Randmeist I have only black images too...without sage-attention
I'm a little perplexed, but the base Qwen models are running more quickly for me than the GGUF, which I understand there might be circumstances where this might be the case, but I can't figure out from where it might be coming.
I'm using 12GB VRAM, I've updated Comfy and the GGUF nodes. The base gives me 12/it while the GGUFs are giving me 16/it.
If anyone has advice, I'd appreciate it. Thanks!
that can be somehow connected to some technical details of everything about your comfy, torch version, whether you use some cache for speed-up ... etc
If you have a 40X0 or 50X0 it makes sense since they support fp8 without conversion.
Thanks for the work, I got the same thing : 3080 Ti and 64 GB RAM it's about 3x faster with the standard 20 GB model (about 3.5 sec / step after the first generation for 1024x1024, lightning 8 steps 1.1 bf16 lora)
Should probably switch to 24 GB of VRAM anyway