just cast to fp8 on inference. Or is it the space?

@Ashen3 is that true for base too? can we get fp8 for the base model? I have been finding that a base model + turbo is sometimes better than a distilled model.

Q8 quant works great btw

probably the more training that gets done to the base, the more the local turbo lora most people use will alter the outputs more. If you skip the turbo lora and increase steps base would probably still be better

@makiaeveli I've tried using the same prompt with the distilled version as well as the base version, following the same formulas as the examples (50 step/cfg 5 for base, and 8step/cfg1 for distilled) I definitely prefer the results with the same prompt from the distilled. I guess the images just look more realistic, although they aren't as refined, limbs can be a promblem, but the noise/issues often lend the image to have a bit more reality to them over the base version.

@fiyota share few images if you can please , then is this distill model will perform better than the base with distill turbo lora on top of it ?

@fiyota Perhaps cfg 5 is too high.

The case you describe sounds like a prompting issue, I experienced myself. Small changes in prompt can have large impact in Flux Klein, both ways. Sampler and scheduler matters too. SA-Solver + beta57 is a good choice. You may also experiment with higher step values (especially with Base + Turbo LoRa workflow).
In general, Flux Klein is a compact transformer model that prioritizes image quality, speed, and prompt adherence, but its reduced capacity makes it more prone to losing global body structure, especially in full-body scenes and at higher resolutions. For me, a basic workaround is to generate at low resolutions, and then add a second pass upscale, latent upscale or just a simple classic ESRGAN (4x_foolhardy_Remacri is a great choice). Anatomical weaknesses may still occur, but not that often as in higher resolutions.
Older UNet-based models had same issues, but they have been finetuned heavily over time. Compact transformer-based diffusion models are hard to finetune, but we are heading in the right direction. Like SNOFS.

are you per chance using the BASE model without the proper cfg and steps?

Guys, thank you very much for the answers and tips, they helped me a lot! I realized that the biggest mistake I was making was my own mistake. In Load Clip, I was using qwen_3_8b_fp8mixed, when I started using qwen_3_8b, it improved by 90%, it looked great and by far this model is one of the best I've tested.

I use 2 color correction nodes after Klein: Colormatchv2 - for automatic or Hue and Saturation - for manual. It is usually necessary to reduce the yellow tone.

usually adding the term "neutral white balance" to the prompt can remedy that.

Would love this too, cant believe no one made a feets lora yet. Where my foot boys at?

@dungaipara532 if you add just a single image of a foot from the bottom as a reference it helps a lot

Never use a distilled model for training, always use a base model.

@hs244 thanks!

Excuse yourself, as this does not qualify as "dumb question". Thank you for your attention to this matter.

@lantoniak617 what is "edition" ? Oh was that you trying to type "editing" with one hand ?