SD XL finetuned with Direct Preference Optimization. Tested and working with both ComfyUI and Automatic1111 WebUI.
Credit goes to the fine-tune here: https://huggingface.co/mhdang/dpo-sdxl-text2image-v1. See the Huggingface page for more details.
Text Encoder and VAE from SD XL v1.0 VAE Fix [E6BB9EA85B] https://civarchive.com/models/101055/sd-xl
The U-Net of this model (ComfyUI-only model here: https://civarchive.com/models/237681/dpo-sdxl-fp16) has been converted into a format compatible with both A1111 and ComfyUI using this tool: https://github.com/arenasys/stable-diffusion-webui-model-toolkit
Description
FAQ
Comments (9)
thanks for sharing! may i ask what is this preference optimization about?
From https://huggingface.co/mhdang/dpo-sdxl-text2image-v1
"Direct Preference Optimization (DPO) for text-to-image diffusion models is a method to align diffusion models to text human preferences by directly optimizing on human comparison data."
@adempotent In caveman please?
@pandaasylum404354 Most models are capable of generating excellent or very poor images, right? From what I understand, DPO basically means a human steps in and says, "Hey, I like image A better than B. Let's lean toward A". So it's fine-tuned with user intervention.
This is how I interpret it at least so take that with a grain of salt.
seems usual loras dont work
looks allways overtrained also with weight 0.1
Lora trained for SD XL Base are working for me. ComfyUI seems to output better images than A1111 with this checkpoint. Unsure why though...
I've just checked and loras are working perfect. Don't know maybe smth wrong with a1111. I use comfy
im sorry but whats the difference with this?
there seem to be some rendering bugs in this one compared to the other one. I recommend using other one.
See https://postimg.cc/9DrJZTFp
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.