Fix schedulers zero SNR and rescale classifier free guidance (#3664)

* Implement option for rescaling betas to zero terminal SNR

* Implement rescale classifier free guidance in pipeline_stable_diffusion.py

* focus on DDIM

* make style

* make style

* make style

* make style

* Apply suggestions from Peter Lin

* Apply suggestions from Peter Lin

* make style

* Apply suggestions from code review

* Apply suggestions from code review

* make style

* make style

---------

Co-authored-by: MaxWe00 <[email protected]>
Co-authored-by: Patrick von Platen <[email protected]>

Author: 3 people

docs/source/en/api/pipelines/stable_diffusion/stable_diffusion_2.mdx

@@ -71,6 +71,64 @@ image = pipe(prompt, guidance_scale=9, num_inference_steps=25).images[0]
 image.save("astronaut.png")
 ```
 
+#### Experimental: "Common Diffusion Noise Schedules and Sample Steps are Flawed": 
+
+The paper **[Common Diffusion Noise Schedules and Sample Steps are Flawed](https://arxiv.org/abs/2305.08891)** 
+claims that a mismatch between the training and inference settings leads to suboptimal inference generation results for Stable Diffusion.
+
+The abstract reads as follows:
+
+*We discover that common diffusion noise schedules do not enforce the last timestep to have zero signal-to-noise ratio (SNR),
+and some implementations of diffusion samplers do not start from the last timestep.
+Such designs are flawed and do not reflect the fact that the model is given pure Gaussian noise at inference, creating a discrepancy between training and inference.
+We show that the flawed design causes real problems in existing implementations. 
+In Stable Diffusion, it severely limits the model to only generate images with medium brightness and 
+prevents it from generating very bright and dark samples. We propose a few simple fixes: 
+- (1) rescale the noise schedule to enforce zero terminal SNR; 
+- (2) train the model with v prediction; 
+- (3) change the sampler to always start from the last timestep; 
+- (4) rescale classifier-free guidance to prevent over-exposure. 
+These simple changes ensure the diffusion process is congruent between training and inference and 
+allow the model to generate samples more faithful to the original data distribution.*
+
+You can apply all of these changes in `diffusers` when using [`DDIMScheduler`]:
+- (1) rescale the noise schedule to enforce zero terminal SNR; 
+```py
+pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config, rescale_betas_zero_snr=True)
+```
+- (2) train the model with v prediction; 
+Continue fine-tuning a checkpoint with [`train_text_to_image.py`](https://github.com/huggingface/diffusers/blob/main/examples/text_to_image/train_text_to_image.py) or [`train_text_to_image_lora.py`](https://github.com/huggingface/diffusers/blob/main/examples/text_to_image/train_text_to_image_lora.py)
+and `--prediction_type="v_prediction"`.
+- (3) change the sampler to always start from the last timestep; 
+```py
+pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config, timestep_scaling="trailing")
+```
+- (4) rescale classifier-free guidance to prevent over-exposure. 
+```py
+pipe(..., guidance_rescale=0.7)
+```
+
+An example is to use [this checkpoint](https://huggingface.co/ptx0/pseudo-journey-v2) 
+which has been fine-tuned using the `"v_prediction"`.
+
+The checkpoint can then be run in inference as follows:
+
+```py
+from diffusers import DiffusionPipeline, DDIMScheduler
+
+pipe = DiffusionPipeline.from_pretrained("ptx0/pseudo-journey-v2", torch_dtype=torch.float16)
+pipe.scheduler = DDIMScheduler.from_config(
+    pipe.scheduler.config, rescale_betas_zero_snr=True, timestep_scaling="trailing"
+)
+pipe.to("cuda")
+
+prompt = "A lion in galaxies, spirals, nebulae, stars, smoke, iridescent, intricate detail, octane render, 8k"
+image = pipeline(prompt, guidance_rescale=0.7).images[0]
+```
+
+## DDIMScheduler
+[[autodoc]] DDIMScheduler
+
 ### Image Inpainting
 
 - *Image Inpainting (512x512 resolution)*: [stabilityai/stable-diffusion-2-inpainting](https://huggingface.co/stabilityai/stable-diffusion-2-inpainting) with [`StableDiffusionInpaintPipeline`]

docs/source/en/api/schedulers/ddim.mdx

@@ -18,10 +18,71 @@ specific language governing permissions and limitations under the License.
 
 The abstract of the paper is the following:
 
-Denoising diffusion probabilistic models (DDPMs) have achieved high quality image generation without adversarial training, yet they require simulating a Markov chain for many steps to produce a sample. To accelerate sampling, we present denoising diffusion implicit models (DDIMs), a more efficient class of iterative implicit probabilistic models with the same training procedure as DDPMs. In DDPMs, the generative process is defined as the reverse of a Markovian diffusion process. We construct a class of non-Markovian diffusion processes that lead to the same training objective, but whose reverse process can be much faster to sample from. We empirically demonstrate that DDIMs can produce high quality samples 10× to 50× faster in terms of wall-clock time compared to DDPMs, allow us to trade off computation for sample quality, and can perform semantically meaningful image interpolation directly in the latent space.
+*Denoising diffusion probabilistic models (DDPMs) have achieved high quality image generation without adversarial training, 
+yet they require simulating a Markov chain for many steps to produce a sample. 
+To accelerate sampling, we present denoising diffusion implicit models (DDIMs), a more efficient class of iterative implicit probabilistic models
+with the same training procedure as DDPMs. In DDPMs, the generative process is defined as the reverse of a Markovian diffusion process. 
+We construct a class of non-Markovian diffusion processes that lead to the same training objective, but whose reverse process can be much faster to sample from.
+We empirically demonstrate that DDIMs can produce high quality samples 10× to 50× faster in terms of wall-clock time compared to DDPMs, allow us to trade off 
+computation for sample quality, and can perform semantically meaningful image interpolation directly in the latent space.*
 
 The original codebase of this paper can be found here: [ermongroup/ddim](https://github.com/ermongroup/ddim).
 For questions, feel free to contact the author on [tsong.me](https://tsong.me/).
 
+### Experimental: "Common Diffusion Noise Schedules and Sample Steps are Flawed": 
+
+The paper **[Common Diffusion Noise Schedules and Sample Steps are Flawed](https://arxiv.org/abs/2305.08891)** 
+claims that a mismatch between the training and inference settings leads to suboptimal inference generation results for Stable Diffusion.
+
+The abstract reads as follows:
+
+*We discover that common diffusion noise schedules do not enforce the last timestep to have zero signal-to-noise ratio (SNR),
+and some implementations of diffusion samplers do not start from the last timestep.
+Such designs are flawed and do not reflect the fact that the model is given pure Gaussian noise at inference, creating a discrepancy between training and inference.
+We show that the flawed design causes real problems in existing implementations. 
+In Stable Diffusion, it severely limits the model to only generate images with medium brightness and 
+prevents it from generating very bright and dark samples. We propose a few simple fixes: 
+- (1) rescale the noise schedule to enforce zero terminal SNR; 
+- (2) train the model with v prediction; 
+- (3) change the sampler to always start from the last timestep; 
+- (4) rescale classifier-free guidance to prevent over-exposure. 
+These simple changes ensure the diffusion process is congruent between training and inference and 
+allow the model to generate samples more faithful to the original data distribution.*
+
+You can apply all of these changes in `diffusers` when using [`DDIMScheduler`]:
+- (1) rescale the noise schedule to enforce zero terminal SNR; 
+```py
+pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config, rescale_betas_zero_snr=True)
+```
+- (2) train the model with v prediction; 
+Continue fine-tuning a checkpoint with [`train_text_to_image.py`](https://github.com/huggingface/diffusers/blob/main/examples/text_to_image/train_text_to_image.py) or [`train_text_to_image_lora.py`](https://github.com/huggingface/diffusers/blob/main/examples/text_to_image/train_text_to_image_lora.py)
+and `--prediction_type="v_prediction"`.
+- (3) change the sampler to always start from the last timestep; 
+```py
+pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config, timestep_scaling="trailing")
+```
+- (4) rescale classifier-free guidance to prevent over-exposure. 
+```py
+pipe(..., guidance_rescale=0.7)
+```
+
+An example is to use [this checkpoint](https://huggingface.co/ptx0/pseudo-journey-v2) 
+which has been fine-tuned using the `"v_prediction"`.
+
+The checkpoint can then be run in inference as follows:
+
+```py
+from diffusers import DiffusionPipeline, DDIMScheduler
+
+pipe = DiffusionPipeline.from_pretrained("ptx0/pseudo-journey-v2", torch_dtype=torch.float16)
+pipe.scheduler = DDIMScheduler.from_config(
+    pipe.scheduler.config, rescale_betas_zero_snr=True, timestep_scaling="trailing"
+)
+pipe.to("cuda")
+
+prompt = "A lion in galaxies, spirals, nebulae, stars, smoke, iridescent, intricate detail, octane render, 8k"
+image = pipeline(prompt, guidance_rescale=0.7).images[0]
+```
+
 ## DDIMScheduler
 [[autodoc]] DDIMScheduler

examples/text_to_image/train_text_to_image.py

@@ -307,6 +307,12 @@ def parse_args():
     parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
     parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.")
     parser.add_argument("--hub_token", type=str, default=None, help="The token to use to push to the Model Hub.")
+    parser.add_argument(
+        "--prediction_type",
+        type=str,
+        default=None,
+        help="The prediction_type that shall be used for training. Choose between 'epsilon' or 'v_prediction' or leave `None`. If left to `None` the default prediction type of the scheduler: `noise_scheduler.config.prediciton_type` is chosen.",
+    )
     parser.add_argument(
         "--hub_model_id",
         type=str,
@@ -848,6 +854,10 @@ def collate_fn(examples):
                 encoder_hidden_states = text_encoder(batch["input_ids"])[0]
 
                 # Get the target for loss depending on the prediction type
+                if args.prediction_type is not None:
+                    # set prediction_type of scheduler if defined
+                    noise_scheduler.register_to_config(prediction_type=args.prediction_type)
+
                 if noise_scheduler.config.prediction_type == "epsilon":
                     target = noise
                 elif noise_scheduler.config.prediction_type == "v_prediction":

examples/text_to_image/train_text_to_image_lora.py

@@ -272,6 +272,12 @@ def parse_args():
     parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
     parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.")
     parser.add_argument("--hub_token", type=str, default=None, help="The token to use to push to the Model Hub.")
+    parser.add_argument(
+        "--prediction_type",
+        type=str,
+        default=None,
+        help="The prediction_type that shall be used for training. Choose between 'epsilon' or 'v_prediction' or leave `None`. If left to `None` the default prediction type of the scheduler: `noise_scheduler.config.prediciton_type` is chosen.",
+    )
     parser.add_argument(
         "--hub_model_id",
         type=str,
@@ -749,6 +755,10 @@ def collate_fn(examples):
                 encoder_hidden_states = text_encoder(batch["input_ids"])[0]
 
                 # Get the target for loss depending on the prediction type
+                if args.prediction_type is not None:
+                    # set prediction_type of scheduler if defined
+                    noise_scheduler.register_to_config(prediction_type=args.prediction_type)
+
                 if noise_scheduler.config.prediction_type == "epsilon":
                     target = noise
                 elif noise_scheduler.config.prediction_type == "v_prediction":

src/diffusers/pipelines/alt_diffusion/pipeline_alt_diffusion.py

@@ -51,6 +51,21 @@
 """
 
 
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg
+def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
+    """
+    Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and
+    Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4
+    """
+    std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True)
+    std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
+    # rescale the results from guidance (fixes overexposure)
+    noise_pred_rescaled = noise_cfg * (std_text / std_cfg)
+    # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images
+    noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg
+    return noise_cfg
+
+
 # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline with Stable->Alt, CLIPTextModel->RobertaSeriesModelWithTransformation, CLIPTokenizer->XLMRobertaTokenizer, AltDiffusionSafetyChecker->StableDiffusionSafetyChecker
 class AltDiffusionPipeline(DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin):
     r"""
@@ -567,6 +582,7 @@ def __call__(
         callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
         callback_steps: int = 1,
         cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        guidance_rescale: float = 0.0,
     ):
         r"""
         Function invoked when calling the pipeline for generation.
@@ -627,6 +643,11 @@ def __call__(
                 A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
                 `self.processor` in
                 [diffusers.cross_attention](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/cross_attention.py).
+            guidance_rescale (`float`, *optional*, defaults to 0.7):
+                Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are
+                Flawed](https://arxiv.org/pdf/2305.08891.pdf) `guidance_scale` is defined as `φ` in equation 16. of
+                [Common Diffusion Noise Schedules and Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf).
+                Guidance rescale factor should fix overexposure when using zero terminal SNR.
 
         Examples:
 
@@ -717,6 +738,10 @@ def __call__(
                     noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
                     noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
 
+                if do_classifier_free_guidance and guidance_rescale > 0.0:
+                    # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
+                    noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale)
+
                 # compute the previous noisy sample x_t -> x_t-1
                 latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
 

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py

@@ -55,6 +55,20 @@
 """
 
 
+def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
+    """
+    Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and
+    Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4
+    """
+    std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True)
+    std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
+    # rescale the results from guidance (fixes overexposure)
+    noise_pred_rescaled = noise_cfg * (std_text / std_cfg)
+    # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images
+    noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg
+    return noise_cfg
+
+
 class StableDiffusionPipeline(DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin, FromCkptMixin):
     r"""
     Pipeline for text-to-image generation using Stable Diffusion.
@@ -568,6 +582,7 @@ def __call__(
         callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
         callback_steps: int = 1,
         cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        guidance_rescale: float = 0.0,
     ):
         r"""
         Function invoked when calling the pipeline for generation.
@@ -628,6 +643,11 @@ def __call__(
                 A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
                 `self.processor` in
                 [diffusers.cross_attention](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/cross_attention.py).
+            guidance_rescale (`float`, *optional*, defaults to 0.7):
+                Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are
+                Flawed](https://arxiv.org/pdf/2305.08891.pdf) `guidance_scale` is defined as `φ` in equation 16. of
+                [Common Diffusion Noise Schedules and Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf).
+                Guidance rescale factor should fix overexposure when using zero terminal SNR.
 
         Examples:
 
@@ -718,6 +738,10 @@ def __call__(
                     noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
                     noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
 
+                if do_classifier_free_guidance and guidance_rescale > 0.0:
+                    # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
+                    noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale)
+
                 # compute the previous noisy sample x_t -> x_t-1
                 latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]