Finetuned SDXL VAE / Decoder only / No retrain SDXL needed
Compare: https://imgsli.com/NDE1MjY1/1/2
Full description:
Description
-
FAQ
Comments (12)
Sorry, cannot see any difference to the original SDXL VAE. Maybe if you would provide more example images. Why do you train on 512 pixels and not 1024? SDXL has 1024.
Very subtle difference, just better. Its not about color, saturation and contrast, more about highlights, shadows.
Because you need a beefy PC with enough memory to train it with a reasonable res. A 512x512 image has 262k pixels, it is not like latent space. I can't even train on 256x256 with my consumer grade GPU. I am positive he has Colab Pro to train the VAE(s)
---
Why did he upload the same vae 6 times on HuggingFace?
Only final_vae/diffusion_pytorch_model.safetensors isn't the same
2P2 i use it like git (git pull) and this is research repo - pls use https://huggingface.co/AiArtLab/sdxl_vae not Simple Vae (it not finished latent upscaler - in dev)
https://imgsli.com/NDA4MTE0
VAE is different from unet - most train it in 256, i try 768 - but for now looks like in 512 train is faster.
TLDR; its trained on random crop from 1k+ image
>Sorry, cannot see any difference to the original SDXL VAE.
It is normal. The main improves of custom VAEs are for training from latents generated by VAEs - less noises -> faster and better convergence with a low loss.
Good job every 1% improvement is nice for me every pixel matters
Nice VAE to train. Better than the default SDXL VAE.
Nice project. Is this VAE still being trained on? I hope the project is progressing fine and dandy. Looking forward to seeing the next version. Congrats!
Just for the record, i bundled in this vae in my last checkpoint (Event Horizon 1.6) as an experiment. This one:
https://civitai.com/models/1645577?modelVersionId=2209234
So you can decide by yourself if it works or not because there's lot of samples. I hope i can use newer vaes of aiartlab in future projects. Thanks. Have a nice day!