Hello Civitai Community,
I'm thrilled to share an update on a recent project of mine. After some dedicated work, I've developed a highly effective text-to-image model. This innovation results from integrating the DPO model from Hugging Face with several advanced counterparts, including Juggernaut7XL, ALBEDOXL, MEARGEHEAVEN, and a model of my own design. The outcome is a unique fusion that showcases exceptional prompt adherence and semantic understanding, it seems to be a step above base SDXL and a step closer to DALLE-3 in terms of prompt comprehension. Notably, this model excels in interpreting and adhering to the given prompts, focusing more on semantic accuracy than on ultra-high-fidelity image generation.
In my quest to refine this model, I encountered a small hiccup – I inadvertently deleted the model from Civitai. However, I've quickly rectified this and reuploaded the model for community access. I'm also excited to announce that version 1.1 is already underway, aiming to bring further improvements and expanded capabilities.
I'm eager to hear your thoughts and feedback on this version. Your insights are invaluable in guiding the ongoing development of this project. So, please stay tuned for the upcoming iteration and thank you for your continued support
Settings for OpenDalle v1.0
Use these settings for the best results with OpenDalle v1.0:
CFG Scale: Use a CFG scale of 8 to 7
Steps: 60 to 70 steps for more detail, 35 steps for faster results.
Sampler: DPM2
Scheduler: Normal or Karras
Description
FAQ
Comments (7)
You should take more dalle-3 outputs and train them into this model using dreambooth, it should be pretty straight forward and lead to a unique model
doing that right now but its gonna take a week or more. training 100k unique samples
This does give refreshingly different outputs to my prompts. It's a winner and I look forward to future versions. Thanks!
Probably the best prompting model I've used. Hope to see more improvements.
Complex promts work much better in direct comparison to base sdxl! But it loses variety. People look more same-ish, because the beautiful ones are picked more from the set, loosing the personality of ordinay faces. But overall the better model! You just have be more specific if you want different faces...
Yep, just be specific. Control is better than randomness. (Also, use CADS to get randomness out of any model, and IPAdapter to reuse a good face you find.)
Any plans to finetune this model https://huggingface.co/mhdang/dpo-sdxl-text2image-v1 on your dataset?
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.