Finetune repo: lavinal712/AutoencoderKL
Dataset: CortexLM/midjourney-v6
Eval Dataset: playgroundai/MJHQ-30K
Fine-tuning modules: decoder and quant_conv
Input:
Reconstruction:
| metrics on ImageNet | rFID | PSNR | SSIM | LPIPS | gFID (MJHQ-30K) |
|---|---|---|---|---|---|
| sdxl-vae | 0.665 | 27.376 | 0.794 | 0.122 | 6.558 |
| finetuned (ours) | 1.358 | 27.590 | 0.802 | 0.121 | 6.643 |