Flux 2 Klein 9B Release
Suggested strength: 0.7-0.8 (base model)
The LORA was trained on the base model. From my observations, the model really struggles with anatomy and often times struggles to interpret prompt, but is great at realism.
I was unable to get it to follow my prompts well using the distilled model. If anyone has suggestions or samples you could provide using the distilled model, please let me know.
Qwen image generation model. Suggested strength: 0.6-0.85
Notes
LORA strength -> it seems to work reasonably well at strengths 0.6-0.85, but feel free to experiment. I found that at strengths over 0.85, it significantly affects image composition.
At lower strengths, it can struggle with side-trained aspects of the lora (such as mismatched footwear)
Workflows
Workflows attached to sample images.
The workflows use VAE-utils, that help get rid of artifacts that are common on qwen generated images. Please follow installation tutorial here:
https://github.com/spacepxl/ComfyUI-VAE-Utils/tree/main
Parameters
If you are looking for a quicker generations (but lower fidelity), you can use the following parameters:
Enable Lightning lora
Sampler: res_2s
Steps: 10
CFG: 1
Description
FAQ
Comments (19)
I was waiting for the Qwen version and it works like a charm ! Awesome results ! thanks
Thanks for the feedback and great to hear you are liking it!
Works very well! Thanks!
Thank you!
Hey awesome lora, one question what lora do you use for the wan upscale step, the one that makes the image more amateur
Thank you! With regards to the wan upscale - I used a poorly trained T2V Footjob model I am currently looking to retrain and publish. Also, sorry if it is misleading, but the WAN upscale in all the images I tried to upscale actually had the opposite effect - it made everything look less amateur and less realistic, and also significantly changed the facial features. Probably something wrong with the model config though.
This is incredible! Thank you so much for this! Is there any chance you could please make a Non-POV footjob Qwen Lora? Like from a side angle showing both characters in frame? Thanks again for your amazing work! :)
Thank you and also thanks for the suggestion!
Yes, this is something I have thought of myself as well. The main issue with this is getting the dataset. I prefer to work with real videos (for realism), and finding side views of footjobs in a significant enough quantity and with variation (e.g. bare feet/socks/shoes) could be difficult. If you have a dataset/videos/sources that could be good for this, feel free to message me.
Hi, can i ask how to keep the character face consistent after generation? I actually need to input some animated image (like button style eyes), but the results often appear more like real human face. Sometimes the character remains the same, but if i adjust the pov, the face changes significantly. Do you have any suggestions like face-consistent lora or parameter tuning? Thanks a lot!
For face consistency - as far as I know you effectively have 2 options. 1st -> train your own lora for the character you are looking to generate images for. 2nd -> generate an image describing the face/other characteristics as closely as you can, then use faceswap (search for faceswap workflows on civit) to get the face consistency.
For the results giving real human face outputs - this might be related to the lora being trained on real images, not animated. You could try adding keywords to the prompt indicating that you are looking for an animated output and/or using loras that focus on animated outputs. Other creators have added samples that appear cartoon-like to the model gallery, so you could start there.
Thanks for the info! I have another yet slight problem. Im using the lora with strength=0.6, but it always generates the image with the feet bottom facing the viewer, the feet do not embrace the penis. Can i ask if there is any solution, i tried all your prompts and it seems they dont work well.
@cjh122037981 Can you post your images and/or prompts so I could test them please (or send me your workflow via dm)? If you cannot, the only thing I can advise really is to try increasing the lora strength to higher. If you are using other loras, they might be overriding the footjob lora. Additionally, if you download the images I have posted, they should have the workflow embedded in them. If you drag and drop them in comfyui and then run the workflow, do they not generate the same image as the one you downloaded?
This is uniquely good. Have you considered using the data to train a Z-Image version? It should train pretty fast in comparison.
Thanks! I am waiting for the base model for Z-image, and will train then. I tried training on the distilled model and it did not do well (either overfit significantly or failed to properly generate penis)
@absampleph Makes sense. It sure feels like the Z-Image base model is this looming atomic bomb that's going to make this gigantic daily avalanche of Z-Image loras look cute in comparison.
I cant really reliably generate ankle socks or low-cut socks at all, 99/100 gen are long socks no matter how hard i describe them in the prompt and how i set cfg... i also tryed different loras strenghts from 0.65 to 0.85 but nothing really works.
That's right, the lora/model does seem to have trouble with this. I have a lora trained for the 2512 version, which seems much better at specifically this. I will try to get that released soon.
In the meantime, you could try prompting something like "small sporty ankle sock", and then hope for the best (see sample images that have proper/almost proper ankle socks). I have noticed that prompting "low-cut" sometimes makes the output image have a literal cut in the sock.
@absampleph will try that, thank you. Looking forward to the updated version!
Maybe this is just a gender problem ; men don't know they're called SOCKETTES. Or try nude ankles...
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.