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
Flux 2 9B Klein
FAQ
Comments (8)
Nice to see a Klein version! How do you feel about the Klein version compared to Qwen?
I wrote briefly in the model description, but overall I'd say I really like the realism it achieves compared to Qwen, but the rest not so much. It struggles with anything that is not basic anatomy a lot (toe-fingers/detached hands, etc.), and also struggles with complex prompts/scenes from what I have seen. Using the distilled model, it made the issues even worse, but maybe something is wrong with my settings. Also, generating at higher resolutions leads to more weirdness. Of course, it is possible that I have trained the lora somehow badly.
I will maybe test it as a refiner for my qwen images to increase realism, as I think that this is where the model does very well.
What about you? I have not really tested the model apart from this lora.
@absampleph You hit the nail on the head about Klein. Yeah, I feel like if Klein didn't have constant anatomy issues, it would be crazy high tier. With the anatomy issues, it's "just" really good. I have a feeling it can't be fixed with training. I think it's really powerful though. Some say adding steps helps with the anatomy issues, but I'm not so sure. In the end, Klein is really useful even with those errors due to being able to spit out multiple images in the time it takes for one from another Edit model. You can iterate a lot. More slot machine, but you can get lucky early too. The speed also lets you explore ideas faster, while you're okay with things not being perfect.
@Jellai I have not really tried the edit, just T2I. With the settings I used, it took almost as long as Qwen to generate images on the base model. If I generate with the same resolution as Qwen (2MP), it would take even longer.. But yeah, the distilled model is very fast! If you get good results using the distilled model with this lora, let me know, as I'm curious whether that's even possible.
@absampleph
Try this and see how it feels:
https://civitai.com/models/2324315/klein-4b9b-base-to-turbo-lora
For image editing, you can try around lora strength 0.5 (might need to explore this), CFG 1 and 8-10 steps. I get edits in 25 seconds. Better quality than the Turbo model, if you've used that. Don't use those settings for T2I though. For that, you need higher CFG and steps (CFG 3.5 and 10-12 steps), but that means you get some extra speed, AND negative prompt if you use that.
@Jellai Had not seen this, thanks! Will check it out.
Why you don't you seedream or qwen image 2.0?
Is either of those open-source / allow training loras?