Qwen_3_4B Trained Text Encoder
For use in Z-Image or FLUX Klein4B
Without lora's LLM had similar text errors to BF16 base QWEN, but better adherence to nude prompt.
See the comparison article here
Description
FAQ
Comments (15)
Great work! Will there also be 8b? 🥰
@Felldude Ah, great, thank you. Will def try it out.
@CupRunethOverWithLSD I am working on the FP8-FP32 mix that is a bit faster on ampere or newer cards
Finally, a plumper based Checkpoint. Sick of these skinny tiny teens models all the time. I want real full ladies....
Can I convert to gguf q4 or q5.? coz I want to be able to use on rtx3060
The full diffusers FP32 file would be ideal for that at https://huggingface.co/Felldude/FLUX.2-klein-base-4B/tree/main/text_encoder
you should be able to run the bf16 with no problem. i also use a 3060, and i run the bf16 text encoder with an fp16 diffusion model + seedvr2 upscaling with no issues.
Is this comparable to Fedric, Huihui, Hauhau and Heretic versions of the same idea?
I don't see information for some of those models - however the term obliteration suggest a layer 0 or 1 targeted training that might effect the LLM text generation.
I have not tested or tried any of these models, hopefully someone else can.
@Felldude I see how would you go about testing them? Do you know like what kinds of prompts would an uncensored one do better that the base Qwen TE wouldn't, of course I'm not interested in anything not legal, only improving the NSFW quality maybe.
@ferrrett33 Fixed seed in comfy could be used for just prompt adherence - but regarding the LLM it would be more difficult...Asking for example how do I make a nuclear bomb, list all steps could be illegal in some jurisdictions
Will this work for Klein 9B?
So I've been testing this for the past month or so on Z-Image-Turbo, and no offense but I couldn't find a single instance in which it allowed me to do a concept or made the images higher quality. On my last test if we summarize the results it would be this "pulls the image back so more people are added and it becomes more censored". Yes, correct, more censored not less censored compared to base Qwen3_4b. I appreciate you guys experimenting with stuff if you can't demonstrate what it does and the effects are not reproducible on a large scale without side effects too, then the right choice is not to publish your personal experiments so other people spend weeks or months testing them really not getting whatever you thought it was. Thanks.