I LOVE this model :)
This is a Forge UI compatible fp16 version of SRPO for those with 24gb+ vram.
All credit to the wonderful @wikeeyang for the work done quantizing the model.
Other versions for lower vram here :
https://civarchive.com/models/1953067/srpo-refine-quantized?modelVersionId=2210447
(the fp8 on this link is also compatible w/ Forge)
Refer to that page for more information on the model. Again, all credit to @wikeeyang. This is just a mirror of the model from https://www.modelscope.cn/models/wikeeyang/SRPO-Refine-Quantized/file/view/master/Flux1-Dev-SRPO-v1-fp16.safetensors. Uploading here because the download speed on that site is very slow, took me about a whole day to download the whole thing.
24gb vram + 64gb RAM minimum. Tested and works in Forge. Adjust the GPU Weights slider to slightly lower than default if you are getting an error message. Make sure to load clip_l, ae, and t5 like normal.
Flux loras will work about 70% or so, depending on the lora. Slightly better than Flux Krea in my experience. The fp16 version definitely fixes the 'hand' problem on a lot of seeds that I noticed with the fp8 version. It still has the butt-chin problem, mostly this model fixes the plastic problem and gives better, more realistic textures.
More info about SRPO here: https://huggingface.co/tencent/SRPO
Description
FAQ
Comments (10)
Tks, bro! this is the spirit of open source and sharing, let's go ...
<3
I was hyped, but sadly another useless Flux bastard
yeah it still has a lot of the hallmarks of flux sadly
Maybe it is not the model. Maybe it is you.
Will it work with 16GB VRAM?
I dont think so. use the link https://civitai.com/models/1953067/srpo-refine-quantized?modelVersionId=2210447 for other versions
It works with 12 perfectly.
Very nice model that creates a unique result with some nice variations from base Flux, one of my faves currently.
The best result for realistic image with excellent result in background face. The best to me so far. Watch out : Very slow... hard work to master prompt...
Details
Files
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.






