add the option of upsample function for tiny vae #7604
Conversation
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
| @@ -926,6 +926,7 @@ def __init__( | |||
| block_out_channels: Tuple[int, ...], | |||
| upsampling_scaling_factor: int, | |||
| act_fn: str, | |||
| upsample_fn: str, | |||
There was a problem hiding this comment.
We need to have the default of Upsample here.
There was a problem hiding this comment.
Maybe this is not necessary. I checked the code, DecoderTiny will only be called when AutoencoderTiny is initialized, and upsample_fn is bound to pass parameters, consistent with act_fn. The default value of nn.Upsample is "nearest", which upsample_fn will pass. So the change will not affect the original TAESD.
We found that using the default nearest upsampler tends to lead to lattice-like artifacts. Although the lattice artifacts can be mitigated by increasing the weights of the GAN loss, it still leads to other texture artifacts in the generated images. The problem is significantly mitigated when the bilinear upsampler is enabled. The problem was discovered when we performed face restoration using Tiny VAE. |
|
would be nice to see an comparison so that we can add a note to our doc |
|
@IDKiro once my comments are addressed, happy to merge the PR. |
Demonstrate a special sample (most samples require a closer look) where images are generated using the model. |
|
Just have a single comment to be addressed and then we're good to go: #7604 (comment). |
|
It would be nice to be able to choose this possibility for |
What does this PR do?
Fixes # (issue)
Before submitting
documentation guidelines, and
here are tips on formatting docstrings.
Who can review?
Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.