# Checkpoint in ongoing research

This is not a release. I am still trying to improve this.
but it is currently better than sd or sdxl vae.
As it SHOULD be for f8c32 instead of f8c4

But it should be better

See https://github.com/ppbrown/sd15_vae-f8c32
for tools I used to train it.




Results from utility "calculate_loss.py"
(Smaller is better)

        image                               l1     rawvgg     edge      lap
    P1 step_010000/vae_sample.webp      0.2104     7.4103   0.2934   0.0729
    P1 step_070000/vae_sample.webp      0.0153     1.1987   0.0686   0.0385
      (LR 1e-5, lpips weight 0.1 lap 0.02 [NO RAWVGG!!] edge_l1_weight 0.1)


    P2 step_960000/vae_sample.webp      0.0121     0.7109   0.0535   0.0355
      (LR 8e-6, lpips weight 0.04 lap 0.02 rawvgg  hires_tiling)

    P3 step_950000/vae_sample.webp      0.0116     0.6232   0.0492   0.0342
      (LR 4e-6, lpips weight 0.04 lap 0.02 rawvgg  hires_tiling)

    P4 step_230000/vae_sample.webp      0.0111     0.6042   0.0488   0.0339
      (LR 2e-6, lpips weight 0.04 lap 0.02 rawvgg  hires_tiling)

As a comparison:


    sampleimg.img_sdxl.webp             0.0174     0.9795   0.0710   0.0439
    sampleimg.img_flux2.webp            0.0075     0.2425   0.0283   0.0281