Description #

Image generation processor (doubao-seedream)

Key features: #

• Text-to-image: Generate an image by entering only the prompt
• Image-to-image: Input reference_images (one or more reference images) + prompt to generate an image based on the reference images
• Batch image generation: Use sequential_image_generation="auto" + sequential_image_generation_options={"max_images": N} to generate multiple images at once
• Supported output formats:
• response_format="url": Returns image links (links expire within 24 hours after generation)
• response_format="b64_json": Returns base64 strings
• Prompt optimization (supported in some versions only): optimize_prompt_options={"mode": "standard" | "fast"}

Model support: #

• doubao-seedream-4.0（version: 250828）
• doubao-seedream-4.5（version: 251128）

Input and output specifications: #

• Input columns:
• Prompt (prompts): pa.array[str]
• Reference images (reference_images): pa.array[str | list[str]] | None
• Output columns:
• Generation result: pa.array[list[str] | None] (Each row outputs a list of image URLs or base64 strings, depending on response_format)

Operator parameters #

Input #

Output #

Returns the list of images generated for each row:

• Type: list[str] | None
• When response_format="url", list[str] contains image URLs
• When response_format="b64_json", list[str] contains base64 strings
• Returns None for failed/skipped rows

Parameters #

If a parameter does not have a default value, it is required

Examples #

The following code demonstrates how to use Daft + las_udf to perform batch inference with DoubaoImageGenerate, covering both text-to-image and image-to-image scenarios.

import daft
from daft import col
from daft.las.functions.ark_llm.doubao_image_generate import DoubaoImageGenerate
from daft.las.functions.udf import las_udf
import os

if os.getenv("DAFT_RUNNER", "native") == "ray":
    import logging

    import ray

    def configure_logging():
        logging.basicConfig(
            level=logging.INFO,
            format="%(asctime)s - %(name)s - %(levelname)s - %(message)s",
            datefmt="%Y-%m-%d %H:%M:%S.%s".format(),
        )
        logging.getLogger("tracing.span").setLevel(logging.WARNING)
        logging.getLogger("daft_io.stats").setLevel(logging.WARNING)
        logging.getLogger("DaftStatisticsManager").setLevel(logging.WARNING)
        logging.getLogger("DaftFlotillaScheduler").setLevel(logging.WARNING)
        logging.getLogger("DaftFlotillaDispatcher").setLevel(logging.WARNING)

    ray.init(dashboard_host="0.0.0.0", runtime_env={"worker_process_setup_hook": configure_logging})
    daft.set_runner_ray()

daft.set_execution_config(min_cpu_per_task=0)

if __name__ == "__main__":
    # Environment variables must be configured before running:
    # - LAS_BASE_URL, create and obtain from the LAS service page
    # - LAS_INFERENCE_TYPE=image_generate
    # - LAS_IMAGE_GENERATE_ENDPOINT=/api/v1/online/image/generate (optional, usually default)
    # - LAS_API_KEY or API_KEY, create and obtain from the LAS service page
    tos_dir_url = os.getenv("TOS_DIR_URL", "las-cn-beijing-public-online.tos-cn-beijing.volces.com")
    samples = {
        "prompt": [
            "生成一张长城的照片，真实摄影风格",                 # Text-to-image
            "保持构图不变，把图片风格改成水彩画风格",            # Image-to-image (with reference_images)
            None,                                         # Example of an abnormal row: this row outputs None and does not affect other rows
        ],
        "reference_images": [
            None,
            f"https://{tos_dir_url}/public/shared_image_dataset/seedream_test_image.png",
            None,
        ],
    }

    df = daft.from_pydict(samples)

    df = df.with_column(
        "images",
        las_udf(
            DoubaoImageGenerate,
            construct_args={
                "model": "doubao-seedream-4.5",
                "version": 251128,
                "response_format": "url",
                "watermark": False,
                "sequential_image_generation": "disabled",
                "size": "2048x2048",
            },
            batch_size=3,
            concurrency=3,
            num_gpus=0,
        )(col("prompt"), col("reference_images")),
    )

    df.show()

    # Examples (the results of each large model inference may vary)
    # ╭────────────────────────────────────────┬────────────────────────────────┬────────────────────────────────╮
    # │ prompt                                 ┆ reference_images               ┆ images                         │
    # │ ---                                    ┆ ---                            ┆ ---                            │
    # │ String                                 ┆ String                         ┆ List[String]                   │
    # ╞════════════════════════════════════════╪════════════════════════════════╪════════════════════════════════╡
    # │ Generate a photo of the Great Wall in a realistic photography style ┆ None                           ┆ [https://ark-content-generati… │
    # ├╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌┼╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌┼╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌┤
    # │ Keep the composition unchanged and change the image style to watercolor ┆ https://ark-project.tos-cn-be… ┆ [https://ark-content-generati… │
    # ├╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌┼╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌┼╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌┤
    # │ None                                   ┆ None                           ┆ None                           │
    # ╰────────────────────────────────────────┴────────────────────────────────┴────────────────────────────────╯