I've been sleeping on Flux Klein for too long, so time for my own version. This is based on the official Flux Klein version with a few of my own datasets mixed in to provide some of my aesthetic preferences. It's surprisingly good with NSFW too. In some ways it's like a mixture of Chroma and Z-Image, with a few tradeoffs.
Here are my recommended settings, although I haven't investigated all the options.
Steps: 4-8
CFG: 1.0
Sampler: euler, er_sde, dpmpp_2m
Scheduler: beta, bong_tangent
Description
Initial version. Grab the embedded workflow from one of the showcase images. See the model description for the recommended settings.
None of the showcase images have been refined in any way so they are raw generations.
FAQ
Comments (26)
Any plans for Ernie image engine? It has much loyal LLM text encoder for weaker machines than Klein 9B.
I haven't looked at Ernie yet. I'm sure I'll put out a model for it at some point when I have time.
Looks good, but FLUX 9B is still heavy as fk..
Maybe Anima realistic for those that cant use that heavy models?
I haven't had time to look into Anima yet but I hope I can soon.
i find most of my loras dont work on this idk why im pretty new to this whole thing if anyone could help that would be great
LoRAs — especially multiple ones — are such a thing with Klein. Try to use as few as possible and start at very low values.
I haven't noticed any issues with loras so far. This model is a straight derivative of the official model with some of my own training. It really shouldn't be much different with loras than regular klein.
@GBRX Is it with the same dataset as usual? Because -That’s not meant badly! - I‘m not a big fan of your chroma models, but this one is peak!
@CupRunethOverWithLSD No, a few new datasets. I'll be updating the chroma model again when I have time.
@GBRX Wait for the new Zeta or Radiance, it’s trained in 20M. So it comes with a much larger variety oob.
@CupRunethOverWithLSD thanks!
You f*ckin legend. Bear gonna try it.
Very good model. Try with comfy and Forge Neo. Both - good results. Maybe cigarettes may to be smaller, but anything else is very good. Good job.
do you have a image to image or inpaint workflow?
My sdxl workflow supports a reference image with IPAdapter, https://civitai.red/models/1513601/gonzalomo-sdxl-workflows
This is good best i've seen with how irritating flux can be sometimes. Someone needs to do a dedicated model lika Chroma did for flux 1 schnell. Because the base of flux is just neutered so much that the anatomy is the biggest issue for me even when I use anatomy loras.
Thanks so much G. This model is truly amazing. Trained my own 9B lora to go with and the results are better than any other checkpoint, especially Flux. Thanks again for sharing your incredible work. ❤️
Awesome, thanks for the great feedback!
Thank you so much for this amazing model, I can see two BF16 one of them 16 gigabyte and the other is 25 gigabyte Is it possible please to explain the difference between them
Hey, thanks! The 25GB is an all-in-one (AIO) checkpoint that includes the unet, text encoder and vae. The 16GB version is just the unet.
Hello, can this model be edited?
Yes, since it's Klein you can also use it as a regular image edit model.
Thanks a lot!
how do you generate these amazing prompts for these images mate
I either copy someone else's or create my own.
Fair enough because these prompts create amazing images