Saltar al contenido principal

PixArtSigma

GenerativeModel
DashAI.back.models.hugging_face.PixArtSigma

Diffusion Transformer model for high efficiency text-to-image generation.

Wraps the PixArt-Sigma pipeline, which replaces the U-Net backbone used in Stable Diffusion with a scalable Diffusion Transformer (DiT) architecture. Text conditioning is provided by a T5-XXL encoder, enabling richer semantic understanding than CLIP-based models.

PixArt-Sigma achieves state of the art image quality with 14-25 denoising steps (compared to 20-50 for comparable U-Net models) and supports flexible multiscale resolutions up to 2048 px. Two checkpoint sizes are available: 512 px (lighter) and 1024 px (best quality).

References​

Parameters​

checkpoint : string, default=1024
Which PixArt-Sigma checkpoint to use: '1024' for best quality at 1024x1024 px, or '512' for a faster, lighter model at 512x512 px. Both checkpoints are downloaded together.
negative_prompt
num_inference_steps : integer, default=20
Number of denoising steps. PixArt-Sigma achieves good quality with 14-25 steps due to its efficient transformer architecture. More steps refine details but increase generation time.
guidance_scale : number, default=4.5
Classifier-Free Guidance (CFG) scale. PixArt-Sigma works best with lower values (3.5-5.5) compared to U-Net models. Higher values enforce the prompt more strictly but may saturate colors. The default of 4.5 is recommended.
device : string, default=CPU
Hardware device for inference. GPU is strongly recommended. PixArt-Sigma uses a DiT (Diffusion Transformer) architecture with T5 text encoding, which is faster than U-Net on GPU.
seed : integer, default=-1
Random seed for reproducible generation. A fixed positive integer always produces the same image. Use -1 for a random seed.
width : integer, default=1024
Width of the output image in pixels. Must be a multiple of 8. PixArt-Sigma supports flexible resolutions up to 2048px.
height : integer, default=1024
Height of the output image in pixels. Must be a multiple of 8. PixArt-Sigma supports flexible resolutions up to 2048px.
num_images_per_prompt : integer, default=1
How many images to generate from a single prompt in one batch. Requires proportionally more GPU memory per additional image.

Methods​

generate(self, input: str) -> List[Any]

Defined on PixArtSigma

Generate images from a text prompt.

Parameters

input : str
Text prompt to generate an image from.

Returns

List[Any]
Generated output images in a list.

hf_repos(cls)

Defined on PixArtSigma

Download both the 1024 (full pipeline) and 512 (transformer) repos.

Returns

list of tuple of (str, str)
The 1024 checkpoint (T5, VAE, scheduler, tokenizer, transformer) and the 512 checkpoint (transformer only), so either variant can be used after a single download.

component_dir(cls) -> pathlib.Path

Defined on DownloadableMixin

Return this component's own storage directory.

Returns

pathlib.Path
<COMPONENT_PATH>/<ClassName>.

delete(cls) -> None

Defined on DownloadableMixin

Remove the component's downloaded artifacts.

download(cls, report: Optional[Callable[[Optional[float], Optional[str]], NoneType]] = None) -> None

Defined on HFDownloadableMixin

Download all repos listed in hf_repos() into component_dir().

Parameters

report : ProgressReporter, optional
Callback invoked before each repo download with report(None, "Downloading <repo_id>"). None means no progress reporting.

get_metadata(cls) -> Dict[str, Any]

Defined on BaseGenerativeModel

Get metadata values for the current generative model.

Returns

Dict[str, Any]
Dictionary indicating whether the model requires a download before use and the expected download size in bytes.

get_schema(cls) -> dict

Defined on ConfigObject

Generates the component related Json Schema.

Returns

dict
Dictionary representing the Json Schema of the component.

is_downloaded(cls) -> bool

Defined on HFDownloadableMixin

Return whether all repo directories exist and are non-empty.

Returns

bool
True when every repo listed in hf_repos() has a non-empty local directory; False otherwise (including when the list is empty).

validate_and_transform(self, raw_data: dict) -> dict

Defined on ConfigObject

It takes the data given by the user to initialize the model and returns it with all the objects that the model needs to work.

Parameters

raw_data : dict
A dictionary with the data provided by the user to initialize the model.

Returns

dict
A validated dictionary with the necessary objects.

Compatible with​