Stable Diffusion 3

Stable Diffusion 3 (SD3)는 Stability AI의 최신 세대 텍스트-이미지 모델로, Multimodal Diffusion Transformer (MMDiT) 아키텍처를 특징으로 합니다. 여러 피사체, 공간 관계, 다양한 텍스트 스타일을 포함하는 복잡한 프롬프트를 처리하는 데 뛰어납니다. SD3는 향상된 프롬프트 해석을 위해 세 가지 고유한 텍스트 인코더(CLIP-L, OpenCLIP-G, T5-XXL)를 활용합니다. RBLN NPU는 Optimum RBLN을 사용하여 Stable Diffusion 3 파이프라인을 가속화할 수 있습니다.

지원하는 파이프라인

Optimum RBLN은 여러 Stable Diffusion 3 파이프라인을 지원합니다:

• 텍스트-이미지 변환(Text-to-Image): 텍스트 프롬프트에서 고품질 이미지 생성.
• 이미지-이미지 변환(Image-to-Image): 텍스트 프롬프트를 기반으로 기존 이미지 수정.
• 인페인팅(Inpainting): 텍스트 프롬프트에 따라 이미지의 마스킹된 영역 채우기.

중요: Guidance Scale에 따른 배치 크기 설정

배치 크기와 Guidance Scale

Stable Diffusion 3을 guidance scale > 1.0으로 사용할 때(기본값은 일반적으로 5.0-7.0), classifier-free guidance 기법으로 인해 MMDiT Transformer의 실제 배치 크기가 실행 시 2배가 됩니다.

RBLN NPU는 정적 그래프 컴파일을 사용하므로, 컴파일 시 Transformer의 배치 크기가 실행 시 배치 크기와 일치해야 합니다. 그렇지 않으면 추론 중에 오류가 발생합니다.

기본 동작

Transformer의 배치 크기를 명시적으로 지정하지 않는 경우, Optimum RBLN은 다음과 같이 동작합니다:

• guidance scale > 1.0을 사용한다고 가정합니다.
• 자동으로 Transformer의 배치 크기를 파이프라인 배치 크기의 2배로 설정합니다.

기본 guidance scale(1.0보다 큰 값)을 사용할 계획이라면, 이 자동 구성이 올바르게 작동합니다. 그러나 다른 guidance scale을 사용하거나 더 많은 제어가 필요한 경우에는 Transformer의 배치 크기를 명시적으로 구성해야 합니다.

예시: Transformer 배치 크기 명시적 설정 (guidance_scale > 1.0)

from optimum.rbln import RBLNStableDiffusion3PipelineConfig, RBLNStableDiffusion3Pipeline

# guidance_scale > 1.0인 경우 (예: 기본값 7.0)
# Transformer의 배치 크기를 원하는 추론 배치 크기의 2배로 설정
config = RBLNStableDiffusion3PipelineConfig(
    batch_size=1,  # 추론 배치 크기
    height=1024,
    width=1024,
    # Transformer의 배치 크기를 2배로 구성
    transformer=dict(batch_size=2)  # Transformer 배치 크기를 2배로 설정
)

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

# 기본 guidance_scale (예: 7.0)로 일반 추론
prompt = "A cat riding a bike through a field of sunflowers, realistic photo."
image = pipe(prompt).images[0] # 배치 크기 1

# 생성된 이미지 저장
image.save(f"sd3_생성된_이미지.png")
print(f"이미지가 sd3_생성된_이미지.png로 저장되었습니다")

예시: Guidance 비활성화 (guidance_scale = 0.0)

guidance_scale을 정확히 0.0으로 사용할 계획이라면 (classifier-free guidance 비활성화), Transformer 배치 크기를 추론 배치 크기와 일치하도록 명시적으로 설정해야 합니다:

import torch
from optimum.rbln import RBLNStableDiffusion3PipelineConfig, RBLNStableDiffusion3Pipeline

config = RBLNStableDiffusion3PipelineConfig(
    height=1024,
    width=1024,
    # Transformer 배치 크기가 추론 배치 크기와 일치
    transformer=dict(batch_size=1)
)

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

# 추론 시 반드시 guidance_scale=0.0 사용
prompt = "A cat riding a bike through a field of sunflowers, realistic photo."
image = pipe(prompt, guidance_scale=0.0).images[0]

# 생성된 이미지 저장
image.save("sd3_가이던스_없는_이미지.png")
print("이미지가 sd3_가이던스_없는_이미지.png로 저장되었습니다")

사용 예제 (텍스트-이미지)

import torch
from optimum.rbln import RBLNStableDiffusion3Pipeline, RBLNStableDiffusion3PipelineConfig

# 구성 객체 생성 (선택 사항, 기본값 사용 가능)
config = RBLNStableDiffusion3PipelineConfig(
    height=1024,
    width=1024
)

# RBLN NPU용 SD3 모델 로드 및 컴파일
pipe = RBLNStableDiffusion3Pipeline.from_pretrained(
    "stabilityai/stable-diffusion-3-medium-diffusers",
    export=True,
    rbln_config=config, # 생성한 경우 config 전달
)

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

# 생성된 이미지 저장
image.save("sd3_고양이_자전거.png")
print("이미지가 sd3_고양이_자전거.png로 저장되었습니다")

API 참조

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.