>>65753593>>65752869I uploaded this version here -
https://huggingface.co/CluelessC/hll-test/blob/main/lyco/hll6.3-fluff-a7-test.safetensorsIt still fails at 3d, but colors and backgrounds look better? -
https://litter.catbox.moe/r5zski.jpg https://litter.catbox.moe/6vpm8i.jpgIt's resumed from a6, old tags are same.
New tags: ai-generated, artshading.
New artists (not all of them actually work): by neocoill, by yuumei, by mitsukai-inki, by stable diffusion, by wlop, by rhads, by puffyart, by artgerm, by dominikmayerart, by hikage eiji, by razaras, by shadyfox, by size80, by phungdinhdung, by trungbui42, by kycaka, by guillemhp, by nixeu, by taekwonkim, by rutkowski, by kagami hirotaka, by hiroaki samura, by guweiz, by nanora, by imoko \(imonatsuki\), by danielkamarudin
>>65752924Best of luck to you.
To avoid manual scoring, you can try AliignProp with HPSv2 score (implemented here, but it's very slow -
https://github.com/Nerogar/OneTrainer), or
https://arxiv.org/abs/2311.17946 (no public implementation as far i know)
For tagging, have you tried the usual LLM tricks like CoT / reflexion? Maybe it reduces hallucinations? After asking it to describe the image, ask it something like this: "Evaluate the accuracy of your previous response. Fix mistakes if necessary. Provide a revised image caption, describing the image in detail"
I also wanted to tag everything with cogVLM, but i calculated that tagging 800k images with 1 4090 would take over 20 days. so i gave up