Here is an FP16 version of Qwen Image that I distilled from the BF16 model.
The BF16 and FP8 versions are available from ComfyUI on Hugging Face.
https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/tree/main
Big thanks to nunchaku.tech for the original SVDQ model! 🙏
Make sure to check out their work:
👉 https://huggingface.co/nunchaku-tech
Description
FAQ
Comments (32)
censored?
I find it less censored than Flux.
denrakeiw so, its not another good boob generator xd?
does breasts decent but anything below the belt is more or less on par as all other models.
fyi; pretty much any base model released these days will probably never do below the belt nudity anymore. so get use to that being the norm.
TheP3NGU1N chroma?
Somewhat. It still understands nudity and anatomical elements even though it creates body horror versions of them. But that means it will be easier for trainers w/ loras and finetunes to correct it. Meanwhile flux doesn't even know what you're talking about
cutetodeath78409597 And it's a really gory model — horror works insanely well with it.
compgamer1337267 it is kind of censored, its not generating clown nose like flux but still the generated boobs are not that great in my opinion
makiaevelio543 :world's smallest violin plays in the background:
It's called reality. Sorry kiddo, no large company is going to put time and money into making a model for people to goon with and get them into legal trouble over.
But that's why finetuning and loras exist.
compgamer1337267 technically chroma is a finetune but yes, it does nsfw somewhat well.
makiaevelio543 You must have just walked into the room.
makiaevelio543 "dedicated to the proliferation of corporate-breaking media sharing" ahahah.
Only thing they are dedicated to is making money, because they are a business. Once NSFW content gets in the way of that you will see how dedicated to keeping it around they really are.
And being as they are literally running out of payment processors.. you might see that happening a lot sooner than you think. Crypto and 'gift cards' to pay for things right now is a band-aid on a gushing cut. It's only a matter of time.
TheP3NGU1N Dude got stomped so hard he erased his own history. Talk about censorship 😭
if you dont mind barbie girls then its fine lol
Fun fact the only uncensored is Grok model from Elon Musk but he wont shared it to run locally unfortunately. And I just don't like censored models not like I need to render something wild.
Sandoogs would never use anything from that dumb motherfucker. didnt he inject nazi behavior into grok? no thanks
@TheP3NGU1N add a nsfw lora & you're good
Forget about these CRAP models like Flux, Qwen. Just use Wan 2.2 14B i2v or t2v and umt5 xxl sent to cpu and fp32 sent to cpu.
@GardaX damn straight. wan 2.2 is the best and its not even close. people need to stop wasting time on qwen. i am so thirsty for good style loras to use with wan its all gooner and anime stuff
@cutetodeath78409597 Wan2.2 is such a pain to train. None of my attempts to train good Wan 2.2 LoRAs have worked out well.
@denrakeiw Good to see someone skilled is making the attempt. Appreciate your work, dude.
@denrakeiw thats really unfortunate ... its such a good model. wonder why its so hard to train?
No realism?
For realism, I'd need to increase the resolution significantly, but rendering takes too long right now. First, I’ll test what the model is capable of.
After couple hours of playing with it -- realism is not it's strong point, especially for the model's speed and sizes.
TheP3NGU1N exactly. we wait for loras now
TheP3NGU1N It's because only 4% of the dataset consists of people.
denrakeiw Damn thats depressing hehe
by itself, it does not appear to be it's strong point. However, with a little messing around and you'd be amazed with either a standalone Qwen workflow or a Qwen to Wan 2.2 (low noise) t2i workflow. Makes a helleuva difference.
MysticMindAi Yes, Qwen2Wan is SOTA
Amazing model :)
Does this still have parts in bf16?


















