Check the version description below for more info and add a ❤️ to receive future updates.
First of all, I would like to thanks @artificialguybr (techguy), CivitAI (dataset) and Redmond.AI (GPU Power) for the training of this checkpoint model.
Finetuned over SDXL1.0.
The idea was to train a robust and consistent model using images generated by CivitAI users. A total of 2,197 images were manually selected and tagged with BLIP-2.
I recommend generating images at least in 1024x1024 resolution for the best results.
However, 832x1216 works fine too!
A good workflow:
Prompt: Whatever you want
Steps: 30
Sampler: DPM++ SDE
CFG scale: 7
Size: 832x1216
Upscaler: 4x-UniScaleV2_Moderate
Upscale by: 1.5
Denoising strength: 0,25~0,35
The model tends to exhibit texture features reminiscent of artistic models from version 1.5 and also tends to displays a kind of oriental eyes.
As soon as possible, I will attempt to generate a new version with wildcards and a more refined training process for photorealism.
Description
Add more styles
Photorealistic resources
Wildcards
FAQ
Comments (4)
so this was trained using images already generated by older AI tech? that kinda feels like inbreeding. wont using real actual photographs be much better? or is this not supposed to be photo real? like making a copy of a copy instead of using the original source material. i know training takes a lot of resources and time so why do this way around?
A lot of xl models were trained the same way
I mean... it is tagged as 'fantasy' and 'artstyle' ;)
so this was trained using images already generated by older AI tech? You just described what is now happening with LLM like Chatgpt and gemini :)