SRPO-Refine-Quantized-v1.0
本模型为 https://huggingface.co/tencent/SRPO 模型的 精调 和 8bit/4bit (fp8_e4m3fn/Q8_0/Q4_1) 量化版本，主要提升出图的清晰度和模型的兼容性。
This model is the refine and quantized version of the model: https://huggingface.co/tencent/SRPO, it improve the clarity of the generated images and the compatibility of the models.
For bf16 version, pls refer to: https://civarchive.com/models/1961797 or my: https://www.modelscope.cn/models/wikeeyang/SRPO-Refine-Quantized
官方原始模型的 bf16 / fp8_e4m3fn / Q8_0 / Q4_1 版本，请参阅：https://huggingface.co/wikeeyang/SRPO-for-ComfyUI 和 https://www.modelscope.cn/models/wikeeyang/SRPO-for-ComfyUI
The quantized version of the offical model please refer to: https://huggingface.co/wikeeyang/SRPO-for-ComfyUI and https://www.modelscope.cn/models/wikeeyang/SRPO-for-ComfyUI
Compare SRPO offical and R&Q v1.0 in the same quantized accuracy:
Example workflow:
Please refer to posted image: https://civarchive.com/images/100069974
Also on: https://huggingface.co/wikeeyang/SRPO-Refine-Quantized-v1.0 ; https://www.modelscope.cn/models/wikeeyang/SRPO-Refine-Quantized
License Agreement
Please fall under SRPO license refer license.txt file and refer to the FLUX.1 [dev] Non-Commercial License.
Description
Refine and Quantized v1.0
FAQ
Comments (35)
刚在B站看到其他博主介绍这个，你这里就有量化版本了，必须5星好评！
谢谢！
adding this Q_8 as recommended to my WF
Thnaks a lot!
Hello! Thanks for the contribution. Where can I find a free Workflow to test this model?
Pls refer to posted image: https://civitai.com/images/100069974
@wikeeyang great workflow, do you have one for upscale (with load image/ image to image)? Thank you
@mich3lang3lo pls refer to https://civitai.com/images/60396559 for UltimateSDUpscale, and https://civitai.com/posts/22919755 for TTP Upscale.
4step lora please =)
use the one that already exists for Flux
LoRA is compatible with the offical Flux.1-Dev model.
nice. thanks
Thanks for the contribution can you plz add Nunchaku 4-bit version
Haha, I want to do it, but the quantize method of Nunchaku - deepcompressor is not easy for deployment, I haven't enviroment to do it, So sorry for this now.
This is awesome, thank you. Can we get an fp16 that works in Forge? The fp8 seems to work, but the fp16 from huggingface leads to 'model type not recognized' error for some reason
Can you access to ModelScope website? https://www.modelscope.cn/, if you can access it, I will upload fp16/bf16 to my repo under there, due to it is too hard for me to upload huge file to HF or Civitai over VPN network. 🤝🤣
@wikeeyang yeah that site loads for me :)
@cutetodeath78409597 Pls download from https://www.modelscope.cn/models/wikeeyang/SRPO-Refine-Quantized/file/view/master/Flux1-Dev-SRPO-v1-fp16.safetensors
@wikeeyang im downloading it but its going extremely slow...but i can wait lol. i can host a mirror of it on here once it's finished if you like, with full credit to you of course.
@cutetodeath78409597 👌，Tks。
@wikeeyang release on huggingface please modelscope is too slow..
Where to download the t5xxl-PrefGRPO.safetensors from the example workflow?
ooh! This is my attempt TE model, extracted from the PrefGRPO project. It's just an experiment, no worries, you can use t5-xxl / flan-t5-xxl or GNER-T5-xxl, etc. about GNER-T5-xxl, you can find something from my https://civitai.com/models/941929
Hi, please release the fp16 version on hugginface
So sorry, It is too hard for me to upload huge file to HF the same as Civitai over low network speed VPN. just as you pull from Modelscope. 🤝🤣
its in civitai.
@Mech4nimaL Tks!
@Zojix @cutetodeath78409597 help downloaded it from my ModelScope repo and put to https://civitai.com/models/1961797
@wikeeyang thanks
Is it feasible to train LORA using this FP8 model? Would it yield better results than training LORA directly with the original Flux-Dev model? I mean, when generating images using this FP8 model, would it be more suitable compared to using LORA trained with the Flux-Dev model?
I provide bf16 version, I think should be better, due to SRPO like Krea or De-Distilled model of Flux famliy, There are significant differences in style and detail.
I will research this model a little more and then edit the comment, the srpo model is usually great for portraits, however in some thematic-genre images it can go astray and give results that are not related to the prompt ! For creating faces the model is really excellent and I recommend it, thanks for sharing
Has anyone tested FP8 vs Q_8 version and which would you say is the best for quality?