make scaling factor a config arg of vae/vqvae (#1860)

* make scaling factor cnfig arg of vae

* fix

* make flake happy

* fix ldm

* fix upscaler

* qualirty

* Apply suggestions from code review
Co-authored-by: Anton Lozhkov <anton@huggingface.co>
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* solve conflicts, addres some comments

* examples

* examples min version

* doc

* fix type

* typo

* Update src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_upscale.py
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>

* remove duplicate line

* Apply suggestions from code review
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
Co-authored-by: Anton Lozhkov <anton@huggingface.co>
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

examples/community/clip_guided_stable_diffusion.py

@@ -150,7 +150,7 @@ class CLIPGuidedStableDiffusion(DiffusionPipeline):
         else:
             raise ValueError(f"scheduler type {type(self.scheduler)} not supported")
 
-        sample = 1 / 0.18215 * sample
+        sample = 1 / self.vae.config.scaling_factor * sample
         image = self.vae.decode(sample).sample
         image = (image / 2 + 0.5).clamp(0, 1)
 
@@ -336,7 +336,7 @@ class CLIPGuidedStableDiffusion(DiffusionPipeline):
             latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
 
         # scale and decode the image latents with vae
-        latents = 1 / 0.18215 * latents
+        latents = 1 / self.vae.config.scaling_factor * latents
         image = self.vae.decode(latents).sample
 
         image = (image / 2 + 0.5).clamp(0, 1)

examples/dreambooth/train_dreambooth.py

@@ -803,7 +803,7 @@ def main(args):
             with accelerator.accumulate(unet):
                 # Convert images to latent space
                 latents = vae.encode(batch["pixel_values"].to(dtype=weight_dtype)).latent_dist.sample()
-                latents = latents * 0.18215
+                latents = latents * vae.config.scaling_factor
 
                 # Sample noise that we'll add to the latents
                 noise = torch.randn_like(latents)

examples/dreambooth/train_dreambooth_flax.py

@@ -533,7 +533,7 @@ def main():
             latents = vae_outputs.latent_dist.sample(sample_rng)
             # (NHWC) -> (NCHW)
             latents = jnp.transpose(latents, (0, 3, 1, 2))
-            latents = latents * 0.18215
+            latents = latents * vae.config.scaling_factor
 
             # Sample noise that we'll add to the latents
             noise_rng, timestep_rng = jax.random.split(sample_rng)

examples/dreambooth/train_dreambooth_lora.py

@@ -853,7 +853,7 @@ def main(args):
             with accelerator.accumulate(unet):
                 # Convert images to latent space
                 latents = vae.encode(batch["pixel_values"].to(dtype=weight_dtype)).latent_dist.sample()
-                latents = latents * 0.18215
+                latents = latents * vae.config.scaling_factor
 
                 # Sample noise that we'll add to the latents
                 noise = torch.randn_like(latents)

examples/research_projects/colossalai/train_dreambooth_colossalai.py

@@ -607,7 +607,7 @@ def main(args):
             optimizer.zero_grad()
 
             latents = vae.encode(batch["pixel_values"].to(dtype=weight_dtype)).latent_dist.sample()
-            latents = latents * 0.18215
+            latents = latents * vae.config.scaling_factor
 
             # Sample noise that we'll add to the latents
             noise = torch.randn_like(latents)

examples/research_projects/dreambooth_inpaint/train_dreambooth_inpaint.py

@@ -33,7 +33,7 @@ from transformers import CLIPTextModel, CLIPTokenizer
 
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.10.0.dev0")
+check_min_version("0.13.0.dev0")
 
 logger = get_logger(__name__)
 
@@ -699,13 +699,13 @@ def main():
                 # Convert images to latent space
 
                 latents = vae.encode(batch["pixel_values"].to(dtype=weight_dtype)).latent_dist.sample()
-                latents = latents * 0.18215
+                latents = latents * vae.config.scaling_factor
 
                 # Convert masked images to latent space
                 masked_latents = vae.encode(
                     batch["masked_images"].reshape(batch["pixel_values"].shape).to(dtype=weight_dtype)
                 ).latent_dist.sample()
-                masked_latents = masked_latents * 0.18215
+                masked_latents = masked_latents * vae.config.scaling_factor
 
                 masks = batch["masks"]
                 # resize the mask to latents shape as we concatenate the mask to the latents

examples/research_projects/intel_opts/textual_inversion/textual_inversion_bf16.py

@@ -51,7 +51,7 @@ else:
 
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.10.0.dev0")
+check_min_version("0.13.0.dev0")
 
 
 logger = get_logger(__name__)
@@ -555,7 +555,7 @@ def main():
                 with accelerator.accumulate(text_encoder):
                     # Convert images to latent space
                     latents = vae.encode(batch["pixel_values"]).latent_dist.sample().detach()
-                    latents = latents * 0.18215
+                    latents = latents * vae.config.scaling_factor
 
                     # Sample noise that we'll add to the latents
                     noise = torch.randn(latents.shape).to(latents.device)

examples/research_projects/multi_subject_dreambooth/train_multi_subject_dreambooth.py

@@ -31,7 +31,7 @@ from transformers import AutoTokenizer, PretrainedConfig
 
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.10.0.dev0")
+check_min_version("0.13.0.dev0")
 
 logger = get_logger(__name__)
 
@@ -788,7 +788,7 @@ def main(args):
             with accelerator.accumulate(unet):
                 # Convert images to latent space
                 latents = vae.encode(batch["pixel_values"].to(dtype=weight_dtype)).latent_dist.sample()
-                latents = latents * 0.18215
+                latents = latents * vae.config.scaling_factor
 
                 # Sample noise that we'll add to the latents
                 noise = torch.randn_like(latents)

examples/text_to_image/train_text_to_image.py

@@ -636,7 +636,7 @@ def main():
             with accelerator.accumulate(unet):
                 # Convert images to latent space
                 latents = vae.encode(batch["pixel_values"].to(weight_dtype)).latent_dist.sample()
-                latents = latents * 0.18215
+                latents = latents * vae.config.scaling_factor
 
                 # Sample noise that we'll add to the latents
                 noise = torch.randn_like(latents)

examples/text_to_image/train_text_to_image_flax.py

@@ -438,7 +438,7 @@ def main():
             latents = vae_outputs.latent_dist.sample(sample_rng)
             # (NHWC) -> (NCHW)
             latents = jnp.transpose(latents, (0, 3, 1, 2))
-            latents = latents * 0.18215
+            latents = latents * vae.config.scaling_factor
 
             # Sample noise that we'll add to the latents
             noise_rng, timestep_rng = jax.random.split(sample_rng)

examples/text_to_image/train_text_to_image_lora.py

@@ -689,7 +689,7 @@ def main():
             with accelerator.accumulate(unet):
                 # Convert images to latent space
                 latents = vae.encode(batch["pixel_values"].to(dtype=weight_dtype)).latent_dist.sample()
-                latents = latents * 0.18215
+                latents = latents * vae.config.scaling_factor
 
                 # Sample noise that we'll add to the latents
                 noise = torch.randn_like(latents)

examples/textual_inversion/textual_inversion.py

@@ -711,7 +711,7 @@ def main():
             with accelerator.accumulate(text_encoder):
                 # Convert images to latent space
                 latents = vae.encode(batch["pixel_values"].to(dtype=weight_dtype)).latent_dist.sample().detach()
-                latents = latents * 0.18215
+                latents = latents * vae.config.scaling_factor
 
                 # Sample noise that we'll add to the latents
                 noise = torch.randn_like(latents)

examples/textual_inversion/textual_inversion_flax.py

@@ -525,7 +525,7 @@ def main():
             latents = vae_outputs.latent_dist.sample(sample_rng)
             # (NHWC) -> (NCHW)
             latents = jnp.transpose(latents, (0, 3, 1, 2))
-            latents = latents * 0.18215
+            latents = latents * vae.config.scaling_factor
 
             noise_rng, timestep_rng = jax.random.split(sample_rng)
             noise = jax.random.normal(noise_rng, latents.shape)

src/diffusers/models/autoencoder_kl.py

@@ -54,8 +54,15 @@ class AutoencoderKL(ModelMixin, ConfigMixin):
         block_out_channels (`Tuple[int]`, *optional*, defaults to :
             obj:`(64,)`): Tuple of block output channels.
         act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
-        latent_channels (`int`, *optional*, defaults to `4`): Number of channels in the latent space.
+        latent_channels (`int`, *optional*, defaults to 4): Number of channels in the latent space.
         sample_size (`int`, *optional*, defaults to `32`): TODO
+        scaling_factor (`float`, *optional*, defaults to 0.18215):
+            The component-wise standard deviation of the trained latent space computed using the first batch of the
+            training set. This is used to scale the latent space to have unit variance when training the diffusion
+            model. The latents are scaled with the formula `z = z * scaling_factor` before being passed to the
+            diffusion model. When decoding, the latents are scaled back to the original scale with the formula: `z = 1
+            / scaling_factor * z`. For more details, refer to sections 4.3.2 and D.1 of the [High-Resolution Image
+            Synthesis with Latent Diffusion Models](https://arxiv.org/abs/2112.10752) paper.
     """
 
     @register_to_config
@@ -71,6 +78,7 @@ class AutoencoderKL(ModelMixin, ConfigMixin):
         latent_channels: int = 4,
         norm_num_groups: int = 32,
         sample_size: int = 32,
+        scaling_factor: float = 0.18215,
     ):
         super().__init__()
 

src/diffusers/models/vae_flax.py

@@ -752,8 +752,15 @@ class FlaxAutoencoderKL(nn.Module, FlaxModelMixin, ConfigMixin):
             Latent space channels
         norm_num_groups (:obj:`int`, *optional*, defaults to `32`):
             Norm num group
-        sample_size (:obj:`int`, *optional*, defaults to `32`):
+        sample_size (:obj:`int`, *optional*, defaults to 32):
             Sample input size
+        scaling_factor (`float`, *optional*, defaults to 0.18215):
+            The component-wise standard deviation of the trained latent space computed using the first batch of the
+            training set. This is used to scale the latent space to have unit variance when training the diffusion
+            model. The latents are scaled with the formula `z = z * scaling_factor` before being passed to the
+            diffusion model. When decoding, the latents are scaled back to the original scale with the formula: `z = 1
+            / scaling_factor * z`. For more details, refer to sections 4.3.2 and D.1 of the [High-Resolution Image
+            Synthesis with Latent Diffusion Models](https://arxiv.org/abs/2112.10752) paper.
         dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
             parameters `dtype`
     """
@@ -767,6 +774,7 @@ class FlaxAutoencoderKL(nn.Module, FlaxModelMixin, ConfigMixin):
     latent_channels: int = 4
     norm_num_groups: int = 32
     sample_size: int = 32
+    scaling_factor: float = 0.18215
     dtype: jnp.dtype = jnp.float32
 
     def setup(self):

src/diffusers/models/vq_model.py

@@ -57,6 +57,13 @@ class VQModel(ModelMixin, ConfigMixin):
         sample_size (`int`, *optional*, defaults to `32`): TODO
         num_vq_embeddings (`int`, *optional*, defaults to `256`): Number of codebook vectors in the VQ-VAE.
         vq_embed_dim (`int`, *optional*): Hidden dim of codebook vectors in the VQ-VAE.
+        scaling_factor (`float`, *optional*, defaults to `0.18215`):
+            The component-wise standard deviation of the trained latent space computed using the first batch of the
+            training set. This is used to scale the latent space to have unit variance when training the diffusion
+            model. The latents are scaled with the formula `z = z * scaling_factor` before being passed to the
+            diffusion model. When decoding, the latents are scaled back to the original scale with the formula: `z = 1
+            / scaling_factor * z`. For more details, refer to sections 4.3.2 and D.1 of the [High-Resolution Image
+            Synthesis with Latent Diffusion Models](https://arxiv.org/abs/2112.10752) paper.
     """
 
     @register_to_config
@@ -74,6 +81,7 @@ class VQModel(ModelMixin, ConfigMixin):
         num_vq_embeddings: int = 256,
         norm_num_groups: int = 32,
         vq_embed_dim: Optional[int] = None,
+        scaling_factor: float = 0.18215,
     ):
         super().__init__()
 

src/diffusers/pipelines/alt_diffusion/pipeline_alt_diffusion.py

@@ -369,7 +369,7 @@ class AltDiffusionPipeline(DiffusionPipeline):
         return image, has_nsfw_concept
 
     def decode_latents(self, latents):
-        latents = 1 / 0.18215 * latents
+        latents = 1 / self.vae.config.scaling_factor * latents
         image = self.vae.decode(latents).sample
         image = (image / 2 + 0.5).clamp(0, 1)
         # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16

src/diffusers/pipelines/alt_diffusion/pipeline_alt_diffusion_img2img.py

@@ -391,7 +391,7 @@ class AltDiffusionImg2ImgPipeline(DiffusionPipeline):
         return image, has_nsfw_concept
 
     def decode_latents(self, latents):
-        latents = 1 / 0.18215 * latents
+        latents = 1 / self.vae.config.scaling_factor * latents
         image = self.vae.decode(latents).sample
         image = (image / 2 + 0.5).clamp(0, 1)
         # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
@@ -490,7 +490,7 @@ class AltDiffusionImg2ImgPipeline(DiffusionPipeline):
         else:
             init_latents = self.vae.encode(image).latent_dist.sample(generator)
 
-        init_latents = 0.18215 * init_latents
+        init_latents = self.vae.config.scaling_factor * init_latents
 
         if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] == 0:
             # expand init_latents for batch_size

src/diffusers/pipelines/audio_diffusion/pipeline_audio_diffusion.py

@@ -153,7 +153,7 @@ class AudioDiffusionPipeline(DiffusionPipeline):
                 input_images = self.vqvae.encode(torch.unsqueeze(input_images, 0)).latent_dist.sample(
                     generator=generator
                 )[0]
-                input_images = 0.18215 * input_images
+                input_images = self.vqvae.config.scaling_factor * input_images
 
             if start_step > 0:
                 images[0, 0] = self.scheduler.add_noise(input_images, noise, self.scheduler.timesteps[start_step - 1])
@@ -195,7 +195,7 @@ class AudioDiffusionPipeline(DiffusionPipeline):
 
         if self.vqvae is not None:
             # 0.18215 was scaling factor used in training to ensure unit variance
-            images = 1 / 0.18215 * images
+            images = 1 / self.vqvae.config.scaling_factor * images
             images = self.vqvae.decode(images)["sample"]
 
         images = (images / 2 + 0.5).clamp(0, 1)

src/diffusers/pipelines/dit/pipeline_dit.py

@@ -182,7 +182,7 @@ class DiTPipeline(DiffusionPipeline):
         else:
             latents = latent_model_input
 
-        latents = 1 / 0.18215 * latents
+        latents = 1 / self.vae.config.scaling_factor * latents
         samples = self.vae.decode(latents).sample
 
         samples = (samples / 2 + 0.5).clamp(0, 1)