Fix AutoencoderTiny encoder scaling convention

  * Add [-1, 1] -> [0, 1] rescaling to EncoderTiny (this fixes huggingface#4676)

  * Move [0, 1] -> [-1, 1] rescaling from AutoencoderTiny.decode to DecoderTiny
    (i.e. immediately after the final conv, as earlier as possible)

  * Fix missing [0, 255] -> [0, 1] rescaling in AutoencoderTiny.forward

  * Update AutoencoderTinyIntegrationTests to protect against scaling issues.
    The new test constructs a simple image, round-trips it through AutoencoderTiny,
    and confirms the decoded result is approximately equal to the source image.
    This test will fail if new AutoencoderTiny scaling issues are introduced.

  * Context: Raw TAESD weights expect images in [0, 1], but diffusers'
    convention represents images with zero-centered values in [-1, 1],
    so AutoencoderTiny needs to scale / unscale images at the start of
    encoding and at the end of decoding in order to work with diffusers.

Author: madebyollin

src/diffusers/models/autoencoder_tiny.py

@@ -312,9 +312,6 @@ def decode(self, x: torch.FloatTensor, return_dict: bool = True) -> Union[Decode
             output = torch.cat(output)
         else:
             output = self._tiled_decode(x) if self.use_tiling else self.decoder(x)
-        # Refer to the following discussion to know why this is needed.
-        # https://github.com/huggingface/diffusers/pull/4384#discussion_r1279401854
-        output = output.mul_(2).sub_(1)
 
         if not return_dict:
             return (output,)
@@ -334,7 +331,7 @@ def forward(
         """
         enc = self.encode(sample).latents
         scaled_enc = self.scale_latents(enc).mul_(255).round_().byte()
-        unscaled_enc = self.unscale_latents(scaled_enc)
+        unscaled_enc = self.unscale_latents(scaled_enc / 255.0)
         dec = self.decode(unscaled_enc)
 
         if not return_dict:

src/diffusers/models/vae.py

@@ -732,7 +732,8 @@ def custom_forward(*inputs):
                 x = torch.utils.checkpoint.checkpoint(create_custom_forward(self.layers), x)
 
         else:
-            x = self.layers(x)
+            # scale image from [-1, 1] to [0, 1] to match TAESD convention
+            x = self.layers(x.mul(0.5).add(0.5))
 
         return x
 
@@ -790,4 +791,5 @@ def custom_forward(*inputs):
         else:
             x = self.layers(x)
 
-        return x
+        # scale image from [0, 1] to [-1, 1] to match diffusers convention
+        return x.mul(2).sub(1)

tests/models/test_models_vae.py

@@ -302,19 +302,23 @@ def test_tae_tiling(self, in_shape, out_shape):
             dec = model.decode(zeros).sample
             assert dec.shape == out_shape
 
-    def test_stable_diffusion(self):
+    def test_roundtrip(self):
+        # load the autoencoder
         model = self.get_sd_vae_model()
-        image = self.get_sd_image(seed=33)
 
+        # make a black image with white square in the middle
+        image = -torch.ones(1, 3, 512, 512, device=torch_device)
+        image[..., 128:384, 128:384] = 1.0
+
+        # round-trip the image through the autoencoder
         with torch.no_grad():
             sample = model(image).sample
 
-        assert sample.shape == image.shape
+        # the autoencoder reconstruction should match original image, sorta
+        def downscale(x):
+            return torch.nn.functional.avg_pool2d(x, 8)
+        assert torch_all_close(downscale(sample), downscale(image), atol=0.125)
 
-        output_slice = sample[-1, -2:, -2:, :2].flatten().float().cpu()
-        expected_output_slice = torch.tensor([0.9858, 0.9262, 0.8629, 1.0974, -0.091, -0.2485, 0.0936, 0.0604])
-
-        assert torch_all_close(output_slice, expected_output_slice, atol=3e-3)
 
 
 @slow