Stable Diffusion 3

Stable Diffusion 3 (SD3) is the latest generation text-to-image model from Stability AI, featuring a Multimodal Diffusion Transformer (MMDiT) architecture. It excels at handling complex prompts involving multiple subjects, spatial relationships, and diverse text styles. SD3 utilizes three distinct text encoders (CLIP-L, OpenCLIP-G, T5-XXL) for enhanced prompt interpretation. RBLN NPUs can accelerate Stable Diffusion 3 pipelines using Optimum RBLN.

Supported Pipelines

Optimum RBLN supports several Stable Diffusion 3 pipelines:

• Text-to-Image: Generate high-quality images from text prompts.
• Image-to-Image: Modify existing images based on text prompts.
• Inpainting: Fill masked regions of an image guided by text prompts.

Important: Batch Size Configuration for Guidance Scale

Batch Size and Guidance Scale

When using Stable Diffusion 3 with a guidance scale > 1.0 (the default is typically around 5.0-7.0), the MMDiT Transformer's effective batch size is doubled during runtime because of the classifier-free guidance technique.

Since RBLN NPU uses static graph compilation, the Transformer's batch size at compilation time must match its runtime batch size, or you'll encounter errors during inference.

Default Behavior

By default, if you don't explicitly specify the Transformer's batch size, Optimum RBLN will:

• Assume you'll use a guidance scale > 1.0.
• Automatically set the Transformer's batch size to 2× your pipeline's batch size.

If you plan to use the default guidance scale (which is > 1.0), this automatic configuration will work correctly. However, if you plan to use a different guidance scale or want more control, you should explicitly configure the Transformer's batch size.

Example: Explicitly Setting the Transformer Batch Size (guidance_scale > 1.0)

from optimum.rbln import RBLNStableDiffusion3PipelineConfig, RBLNStableDiffusion3Pipeline

# For inference with guidance_scale > 1.0 (e.g., default 7.0)
# Set Transformer's batch size to 2 * your desired inference batch size
config = RBLNStableDiffusion3PipelineConfig(
    batch_size=1,  # Inference batch size
    height=1024,
    width=1024,
    # Configure Transformer with doubled batch size
    transformer=dict(batch_size=2)  # Double the batch size for Transformer
)

pipe = RBLNStableDiffusion3Pipeline.from_pretrained(
    "stabilityai/stable-diffusion-3-medium-diffusers",
    export=True,
    rbln_config=config,
)

# Normal inference with default guidance_scale (e.g., 7.0)
prompt = "A cat riding a bike through a field of sunflowers, realistic photo."
image = pipe(prompt).images[0] # Batch size of 1

# Save the generated image
image.save(f"sd3_generated_image.png")
print(f"Image saved as sd3_generated_image.png")

Example: Disabling Guidance (guidance_scale = 0.0)

If you plan to use a guidance scale of exactly 0.0 (disabling classifier-free guidance), you should explicitly set the Transformer batch size to match your inference batch size:

import torch
from optimum.rbln import RBLNStableDiffusion3PipelineConfig, RBLNStableDiffusion3Pipeline

config = RBLNStableDiffusion3PipelineConfig(
    height=1024,
    width=1024,
    # Transformer batch size matches the inference batch size
    transformer=dict(batch_size=1)
)

pipe = RBLNStableDiffusion3Pipeline.from_pretrained(
    "stabilityai/stable-diffusion-3-medium-diffusers",
    export=True,
    rbln_config=config,
)

# Must use guidance_scale=0.0 during inference
prompt = "A cat riding a bike through a field of sunflowers, realistic photo."
image = pipe(prompt, guidance_scale=0.0).images[0]

# Save the generated image
image.save("sd3_no_guidance_image.png")
print("Image saved as sd3_no_guidance_image.png")

Usage Example (Text-to-Image)

import torch
from optimum.rbln import RBLNStableDiffusion3Pipeline, RBLNStableDiffusion3PipelineConfig

# Create a configuration object (optional, can use defaults)
config = RBLNStableDiffusion3PipelineConfig(
    height=1024,
    width=1024
)

# Load and compile the SD3 model for RBLN NPU
pipe = RBLNStableDiffusion3Pipeline.from_pretrained(
    "stabilityai/stable-diffusion-3-medium-diffusers",
    export=True,
    rbln_config=config,
)

# Generate an image
prompt = "A cat riding a bike through a field of sunflowers, realistic photo."
image = pipe(prompt).images[0]

# Save the generated image
image.save("sd3_cat_bike.png")
print("Image saved as sd3_cat_bike.png")

API Reference

Classes

RBLNStableDiffusion3Pipeline

Bases: RBLNDiffusionMixin, StableDiffusion3Pipeline

RBLN-accelerated implementation of Stable Diffusion 3 pipeline for advanced text-to-image generation.

This pipeline compiles Stable Diffusion 3 models to run efficiently on RBLN NPUs, enabling high-performance inference with improved text understanding, enhanced image quality, and superior prompt adherence.

Functions

from_pretrained(model_id, *, export=None, model_save_dir=None, rbln_config={}, lora_ids=None, lora_weights_names=None, lora_scales=None, **kwargs) classmethod

Load a pretrained diffusion pipeline from a model checkpoint, with optional compilation for RBLN NPUs.

This method has two distinct operating modes

• When export=True: Takes a PyTorch-based diffusion model, compiles it for RBLN NPUs, and loads the compiled model
• When export=False: Loads an already compiled RBLN model from model_id without recompilation

It supports various diffusion pipelines including Stable Diffusion, Kandinsky, ControlNet, and other diffusers-based models.

Parameters:

Name	Type	Description	Default
model_id	`str`	The model ID or path to the pretrained model to load. Can be either: A model ID from the HuggingFace Hub A local path to a saved model directory	required
export	bool	If True, takes a PyTorch model from model_id and compiles it for RBLN NPU execution. If False, loads an already compiled RBLN model from model_id without recompilation.	None
model_save_dir	Optional[PathLike]	Directory to save the compiled model artifacts. Only used when export=True. If not provided and export=True, a temporary directory is used.	None
rbln_config	Dict[str, Any]	Configuration options for RBLN compilation. Can include settings for specific submodules such as text_encoder, unet, and vae. Configuration can be tailored to the specific pipeline being compiled.	{}
lora_ids	Optional[Union[str, List[str]]]	LoRA adapter ID(s) to load and apply before compilation. LoRA weights are fused into the model weights during compilation. Only used when export=True.	None
lora_weights_names	Optional[Union[str, List[str]]]	Names of specific LoRA weight files to load, corresponding to lora_ids. Only used when export=True.	None
lora_scales	Optional[Union[float, List[float]]]	Scaling factor(s) to apply to the LoRA adapter(s). Only used when export=True.	None
kwargs	Any	Additional arguments to pass to the underlying diffusion pipeline constructor or the RBLN compilation process. These may include parameters specific to individual submodules or the particular diffusion pipeline being used.	{}

Returns:

Type	Description
RBLNDiffusionMixin	A compiled or loaded diffusion pipeline that can be used for inference on RBLN NPU. The returned object is an instance of the class that called this method, inheriting from RBLNDiffusionMixin.

Classes

RBLNStableDiffusion3Img2ImgPipeline

Bases: RBLNDiffusionMixin, StableDiffusion3Img2ImgPipeline

RBLN-accelerated implementation of Stable Diffusion 3 pipeline for advanced image-to-image generation.

This pipeline compiles Stable Diffusion 3 models to run efficiently on RBLN NPUs, enabling high-performance inference for transforming input images with superior text understanding and enhanced visual quality.

Functions

from_pretrained(model_id, *, export=None, model_save_dir=None, rbln_config={}, lora_ids=None, lora_weights_names=None, lora_scales=None, **kwargs) classmethod

Load a pretrained diffusion pipeline from a model checkpoint, with optional compilation for RBLN NPUs.

This method has two distinct operating modes

• When export=True: Takes a PyTorch-based diffusion model, compiles it for RBLN NPUs, and loads the compiled model
• When export=False: Loads an already compiled RBLN model from model_id without recompilation

It supports various diffusion pipelines including Stable Diffusion, Kandinsky, ControlNet, and other diffusers-based models.

Parameters:

Name	Type	Description	Default
model_id	`str`	The model ID or path to the pretrained model to load. Can be either: A model ID from the HuggingFace Hub A local path to a saved model directory	required
export	bool	If True, takes a PyTorch model from model_id and compiles it for RBLN NPU execution. If False, loads an already compiled RBLN model from model_id without recompilation.	None
model_save_dir	Optional[PathLike]	Directory to save the compiled model artifacts. Only used when export=True. If not provided and export=True, a temporary directory is used.	None
rbln_config	Dict[str, Any]	Configuration options for RBLN compilation. Can include settings for specific submodules such as text_encoder, unet, and vae. Configuration can be tailored to the specific pipeline being compiled.	{}
lora_ids	Optional[Union[str, List[str]]]	LoRA adapter ID(s) to load and apply before compilation. LoRA weights are fused into the model weights during compilation. Only used when export=True.	None
lora_weights_names	Optional[Union[str, List[str]]]	Names of specific LoRA weight files to load, corresponding to lora_ids. Only used when export=True.	None
lora_scales	Optional[Union[float, List[float]]]	Scaling factor(s) to apply to the LoRA adapter(s). Only used when export=True.	None
kwargs	Any	Additional arguments to pass to the underlying diffusion pipeline constructor or the RBLN compilation process. These may include parameters specific to individual submodules or the particular diffusion pipeline being used.	{}

Returns:

Type	Description
RBLNDiffusionMixin	A compiled or loaded diffusion pipeline that can be used for inference on RBLN NPU. The returned object is an instance of the class that called this method, inheriting from RBLNDiffusionMixin.

Classes

RBLNStableDiffusion3InpaintPipeline

Bases: RBLNDiffusionMixin, StableDiffusion3InpaintPipeline

RBLN-accelerated implementation of Stable Diffusion 3 pipeline for advanced image inpainting.

This pipeline compiles Stable Diffusion 3 models to run efficiently on RBLN NPUs, enabling high-performance inference for filling masked regions with superior text understanding and seamless content generation.

Functions

from_pretrained(model_id, *, export=None, model_save_dir=None, rbln_config={}, lora_ids=None, lora_weights_names=None, lora_scales=None, **kwargs) classmethod

Load a pretrained diffusion pipeline from a model checkpoint, with optional compilation for RBLN NPUs.

This method has two distinct operating modes

• When export=True: Takes a PyTorch-based diffusion model, compiles it for RBLN NPUs, and loads the compiled model
• When export=False: Loads an already compiled RBLN model from model_id without recompilation

It supports various diffusion pipelines including Stable Diffusion, Kandinsky, ControlNet, and other diffusers-based models.

Parameters:

Name	Type	Description	Default
model_id	`str`	The model ID or path to the pretrained model to load. Can be either: A model ID from the HuggingFace Hub A local path to a saved model directory	required
export	bool	If True, takes a PyTorch model from model_id and compiles it for RBLN NPU execution. If False, loads an already compiled RBLN model from model_id without recompilation.	None
model_save_dir	Optional[PathLike]	Directory to save the compiled model artifacts. Only used when export=True. If not provided and export=True, a temporary directory is used.	None
rbln_config	Dict[str, Any]	Configuration options for RBLN compilation. Can include settings for specific submodules such as text_encoder, unet, and vae. Configuration can be tailored to the specific pipeline being compiled.	{}
lora_ids	Optional[Union[str, List[str]]]	LoRA adapter ID(s) to load and apply before compilation. LoRA weights are fused into the model weights during compilation. Only used when export=True.	None
lora_weights_names	Optional[Union[str, List[str]]]	Names of specific LoRA weight files to load, corresponding to lora_ids. Only used when export=True.	None
lora_scales	Optional[Union[float, List[float]]]	Scaling factor(s) to apply to the LoRA adapter(s). Only used when export=True.	None
kwargs	Any	Additional arguments to pass to the underlying diffusion pipeline constructor or the RBLN compilation process. These may include parameters specific to individual submodules or the particular diffusion pipeline being used.	{}

Returns:

Type	Description
RBLNDiffusionMixin	A compiled or loaded diffusion pipeline that can be used for inference on RBLN NPU. The returned object is an instance of the class that called this method, inheriting from RBLNDiffusionMixin.

Classes

RBLNStableDiffusion3PipelineBaseConfig

Bases: RBLNModelConfig

Functions

__init__(transformer=None, text_encoder=None, text_encoder_2=None, text_encoder_3=None, vae=None, *, max_seq_len=None, sample_size=None, image_size=None, batch_size=None, img_height=None, img_width=None, height=None, width=None, guidance_scale=None, **kwargs)

Parameters:

Name	Type	Description	Default
transformer	Optional[RBLNSD3Transformer2DModelConfig]	Configuration for the transformer model component. Initialized as RBLNSD3Transformer2DModelConfig if not provided.	None
text_encoder	Optional[RBLNCLIPTextModelWithProjectionConfig]	Configuration for the primary text encoder. Initialized as RBLNCLIPTextModelWithProjectionConfig if not provided.	None
text_encoder_2	Optional[RBLNCLIPTextModelWithProjectionConfig]	Configuration for the secondary text encoder. Initialized as RBLNCLIPTextModelWithProjectionConfig if not provided.	None
text_encoder_3	Optional[RBLNT5EncoderModelConfig]	Configuration for the tertiary text encoder. Initialized as RBLNT5EncoderModelConfig if not provided.	None
vae	Optional[RBLNAutoencoderKLConfig]	Configuration for the VAE model component. Initialized as RBLNAutoencoderKLConfig if not provided.	None
max_seq_len	Optional[int]	Maximum sequence length for text inputs. Defaults to 256.	None
sample_size	Optional[Tuple[int, int]]	Spatial dimensions for the transformer model.	None
image_size	Optional[Tuple[int, int]]	Dimensions for the generated images. Cannot be used together with img_height/img_width.	None
batch_size	Optional[int]	Batch size for inference, applied to all submodules.	None
img_height	Optional[int]	Height of the generated images.	None
img_width	Optional[int]	Width of the generated images.	None
height	Optional[int]	Height of the generated images.	None
width	Optional[int]	Width of the generated images.	None
guidance_scale	Optional[float]	Scale for classifier-free guidance.	None
kwargs	Any	Additional arguments passed to the parent RBLNModelConfig.	{}

Raises:

Type	Description
ValueError	If both image_size and img_height/img_width are provided.

Note

When guidance_scale > 1.0, the transformer batch size is automatically doubled to accommodate classifier-free guidance.

load(path, **kwargs) classmethod

Load a RBLNModelConfig from a path.

Parameters:

Name	Type	Description	Default
path	str	Path to the RBLNModelConfig file or directory containing the config file.	required
kwargs	Any	Additional keyword arguments to override configuration values. Keys starting with 'rbln_' will have the prefix removed and be used to update the configuration.	{}

Returns:

Name	Type	Description
RBLNModelConfig	RBLNModelConfig	The loaded configuration instance.

Note

This method loads the configuration from the specified path and applies any provided overrides. If the loaded configuration class doesn't match the expected class, a warning will be logged.

RBLNStableDiffusion3PipelineConfig

Bases: RBLNStableDiffusion3PipelineBaseConfig

Config for SD3 Text2Img Pipeline

Functions

__init__(transformer=None, text_encoder=None, text_encoder_2=None, text_encoder_3=None, vae=None, *, max_seq_len=None, sample_size=None, image_size=None, batch_size=None, img_height=None, img_width=None, height=None, width=None, guidance_scale=None, **kwargs)

Parameters:

Name	Type	Description	Default
transformer	Optional[RBLNSD3Transformer2DModelConfig]	Configuration for the transformer model component. Initialized as RBLNSD3Transformer2DModelConfig if not provided.	None
text_encoder	Optional[RBLNCLIPTextModelWithProjectionConfig]	Configuration for the primary text encoder. Initialized as RBLNCLIPTextModelWithProjectionConfig if not provided.	None
text_encoder_2	Optional[RBLNCLIPTextModelWithProjectionConfig]	Configuration for the secondary text encoder. Initialized as RBLNCLIPTextModelWithProjectionConfig if not provided.	None
text_encoder_3	Optional[RBLNT5EncoderModelConfig]	Configuration for the tertiary text encoder. Initialized as RBLNT5EncoderModelConfig if not provided.	None
vae	Optional[RBLNAutoencoderKLConfig]	Configuration for the VAE model component. Initialized as RBLNAutoencoderKLConfig if not provided.	None
max_seq_len	Optional[int]	Maximum sequence length for text inputs. Defaults to 256.	None
sample_size	Optional[Tuple[int, int]]	Spatial dimensions for the transformer model.	None
image_size	Optional[Tuple[int, int]]	Dimensions for the generated images. Cannot be used together with img_height/img_width.	None
batch_size	Optional[int]	Batch size for inference, applied to all submodules.	None
img_height	Optional[int]	Height of the generated images.	None
img_width	Optional[int]	Width of the generated images.	None
height	Optional[int]	Height of the generated images.	None
width	Optional[int]	Width of the generated images.	None
guidance_scale	Optional[float]	Scale for classifier-free guidance.	None
kwargs	Any	Additional arguments passed to the parent RBLNModelConfig.	{}

Raises:

Type	Description
ValueError	If both image_size and img_height/img_width are provided.

Note

When guidance_scale > 1.0, the transformer batch size is automatically doubled to accommodate classifier-free guidance.

load(path, **kwargs) classmethod

Load a RBLNModelConfig from a path.

Parameters:

Name	Type	Description	Default
path	str	Path to the RBLNModelConfig file or directory containing the config file.	required
kwargs	Any	Additional keyword arguments to override configuration values. Keys starting with 'rbln_' will have the prefix removed and be used to update the configuration.	{}

Returns:

Name	Type	Description
RBLNModelConfig	RBLNModelConfig	The loaded configuration instance.

Note

This method loads the configuration from the specified path and applies any provided overrides. If the loaded configuration class doesn't match the expected class, a warning will be logged.

RBLNStableDiffusion3Img2ImgPipelineConfig

Bases: RBLNStableDiffusion3PipelineBaseConfig

Config for SD3 Img2Img Pipeline

Functions

__init__(transformer=None, text_encoder=None, text_encoder_2=None, text_encoder_3=None, vae=None, *, max_seq_len=None, sample_size=None, image_size=None, batch_size=None, img_height=None, img_width=None, height=None, width=None, guidance_scale=None, **kwargs)

Parameters:

Name	Type	Description	Default
transformer	Optional[RBLNSD3Transformer2DModelConfig]	Configuration for the transformer model component. Initialized as RBLNSD3Transformer2DModelConfig if not provided.	None
text_encoder	Optional[RBLNCLIPTextModelWithProjectionConfig]	Configuration for the primary text encoder. Initialized as RBLNCLIPTextModelWithProjectionConfig if not provided.	None
text_encoder_2	Optional[RBLNCLIPTextModelWithProjectionConfig]	Configuration for the secondary text encoder. Initialized as RBLNCLIPTextModelWithProjectionConfig if not provided.	None
text_encoder_3	Optional[RBLNT5EncoderModelConfig]	Configuration for the tertiary text encoder. Initialized as RBLNT5EncoderModelConfig if not provided.	None
vae	Optional[RBLNAutoencoderKLConfig]	Configuration for the VAE model component. Initialized as RBLNAutoencoderKLConfig if not provided.	None
max_seq_len	Optional[int]	Maximum sequence length for text inputs. Defaults to 256.	None
sample_size	Optional[Tuple[int, int]]	Spatial dimensions for the transformer model.	None
image_size	Optional[Tuple[int, int]]	Dimensions for the generated images. Cannot be used together with img_height/img_width.	None
batch_size	Optional[int]	Batch size for inference, applied to all submodules.	None
img_height	Optional[int]	Height of the generated images.	None
img_width	Optional[int]	Width of the generated images.	None
height	Optional[int]	Height of the generated images.	None
width	Optional[int]	Width of the generated images.	None
guidance_scale	Optional[float]	Scale for classifier-free guidance.	None
kwargs	Any	Additional arguments passed to the parent RBLNModelConfig.	{}

Raises:

Type	Description
ValueError	If both image_size and img_height/img_width are provided.

Note

When guidance_scale > 1.0, the transformer batch size is automatically doubled to accommodate classifier-free guidance.

load(path, **kwargs) classmethod

Load a RBLNModelConfig from a path.

Parameters:

Name	Type	Description	Default
path	str	Path to the RBLNModelConfig file or directory containing the config file.	required
kwargs	Any	Additional keyword arguments to override configuration values. Keys starting with 'rbln_' will have the prefix removed and be used to update the configuration.	{}

Returns:

Name	Type	Description
RBLNModelConfig	RBLNModelConfig	The loaded configuration instance.

Note

This method loads the configuration from the specified path and applies any provided overrides. If the loaded configuration class doesn't match the expected class, a warning will be logged.

RBLNStableDiffusion3InpaintPipelineConfig

Bases: RBLNStableDiffusion3PipelineBaseConfig

Config for SD3 Inpainting Pipeline

Functions

__init__(transformer=None, text_encoder=None, text_encoder_2=None, text_encoder_3=None, vae=None, *, max_seq_len=None, sample_size=None, image_size=None, batch_size=None, img_height=None, img_width=None, height=None, width=None, guidance_scale=None, **kwargs)

Parameters:

Name	Type	Description	Default
transformer	Optional[RBLNSD3Transformer2DModelConfig]	Configuration for the transformer model component. Initialized as RBLNSD3Transformer2DModelConfig if not provided.	None
text_encoder	Optional[RBLNCLIPTextModelWithProjectionConfig]	Configuration for the primary text encoder. Initialized as RBLNCLIPTextModelWithProjectionConfig if not provided.	None
text_encoder_2	Optional[RBLNCLIPTextModelWithProjectionConfig]	Configuration for the secondary text encoder. Initialized as RBLNCLIPTextModelWithProjectionConfig if not provided.	None
text_encoder_3	Optional[RBLNT5EncoderModelConfig]	Configuration for the tertiary text encoder. Initialized as RBLNT5EncoderModelConfig if not provided.	None
vae	Optional[RBLNAutoencoderKLConfig]	Configuration for the VAE model component. Initialized as RBLNAutoencoderKLConfig if not provided.	None
max_seq_len	Optional[int]	Maximum sequence length for text inputs. Defaults to 256.	None
sample_size	Optional[Tuple[int, int]]	Spatial dimensions for the transformer model.	None
image_size	Optional[Tuple[int, int]]	Dimensions for the generated images. Cannot be used together with img_height/img_width.	None
batch_size	Optional[int]	Batch size for inference, applied to all submodules.	None
img_height	Optional[int]	Height of the generated images.	None
img_width	Optional[int]	Width of the generated images.	None
height	Optional[int]	Height of the generated images.	None
width	Optional[int]	Width of the generated images.	None
guidance_scale	Optional[float]	Scale for classifier-free guidance.	None
kwargs	Any	Additional arguments passed to the parent RBLNModelConfig.	{}

Raises:

Type	Description
ValueError	If both image_size and img_height/img_width are provided.

Note

When guidance_scale > 1.0, the transformer batch size is automatically doubled to accommodate classifier-free guidance.

load(path, **kwargs) classmethod

Load a RBLNModelConfig from a path.

Parameters:

Name	Type	Description	Default
path	str	Path to the RBLNModelConfig file or directory containing the config file.	required
kwargs	Any	Additional keyword arguments to override configuration values. Keys starting with 'rbln_' will have the prefix removed and be used to update the configuration.	{}

Returns:

Name	Type	Description
RBLNModelConfig	RBLNModelConfig	The loaded configuration instance.

Note

This method loads the configuration from the specified path and applies any provided overrides. If the loaded configuration class doesn't match the expected class, a warning will be logged.