Generated image in blur while training Autoencodertiny Decoder

[phyllispeng123]

Hi, this is Phyllis. Very appreciate your marvelous work on SD-VAE compression and acceleration !!!! I am currently working on Autoencodertiny decoder training from scratch using LDM training structure, but I find that the generated images are not very clear. The training is completely the same as LDM training apart from your extra loss distance(disc(real).mean(), disc(fake).mean()) in my decoder generator (the extra loss indeed helped with stability and FID ,many thanks!!!! ). I train the decoder using SD1.5 encoder output as my input for 190k steps with batch_size=4, lr = 1e-4, but generated images are still not clear. Would you mind give some hints (loss used, steps for training, any finetune stage? ) on how to align with your TAESD results?

this is TAESD result

this is MY result

[madebyollin]

Since the output is a mix of blurry and sharp portions, it looks like the scale of your adversarial loss / gradients is probably too low compared to the reconstruction (MSE/MAE) loss / gradients.

You could try just scaling up the discriminator loss (10 * distance(disc(real).mean(), disc(fake).mean()) or something), or using the automatic scaling from https://github.com/CompVis/latent-diffusion/blob/main/ldm/modules/losses/contperceptual.py#L32.

I also posted a simplified example training notebook using only adversarial loss here https://github.com/madebyollin/seraena/blob/main/TAESDXL_Training_Example.ipynb (which also does some automatic gradient scaling); you could try taking the adversarial loss from that and combining it with your reconstruction losses.

[FangSen9000]

Hi, @phyllispeng123 I would like to know if you could share your complete TAESD training script (Autoencodertiny (TinyVAE)). I'm a bit worried that I might go astray at the beginning of my attempt. If you have any experience, it would be much better for me.

[madebyollin]

@FangSen9000 FWIW I posted a basic TAESDXL example notebook here; there's also a more advanced distillation script by the Hybrid-SD team here (though I haven't personally tried it).