For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).

Python function

generate_images

generate_images()

max.pipelines.lib.generate_images.generate_images(pipeline, tokenizer, prompt, *, images=None, negative_prompt=None, height=None, width=None, num_inference_steps=None, guidance_scale=None, true_cfg_scale=None, strength=None, num_images_per_prompt=1, seed=0, caller='generate_images()')

source

Generates images for each prompt, one prompt per pipeline run.

An argument left as None takes the model’s default, as a max serve request that omits it does.

Parameters:

• pipeline (ImagePipeline) – The image pipeline to run.
• tokenizer (ImageTokenizer) – The pipeline’s tokenizer, which builds each request’s context.
• prompt (str | Sequence[str]) – One prompt, or a list of prompts.
• images (Sequence[Sequence[bytes | Image]] | None) – One list of input images per prompt, each image encoded bytes or a PIL image, as for generate(). A model reads one input image, so a list holds one image, or none for a text-only prompt. Without height and width, the output size is the input image’s size rounded down to a size the model takes.
• negative_prompt (str | Sequence[str] | None) – What to steer away from, as one string for every prompt or one per prompt.
• height (int | None) – The image height in pixels.
• width (int | None) – The image width in pixels.
• num_inference_steps (int | None) – The number of denoising steps.
• guidance_scale (float | None) – How closely the image follows the prompt.
• true_cfg_scale (float | None) – The true classifier-free guidance scale. Models that use it need it above 1 and a negative_prompt.
• strength (float | None) – How far an input image is noised, in (0, 1]. Z-Image reads it, and FLUX.2 ignores it.
• num_images_per_prompt (int) – The number of images for each prompt.
• seed (int) – The seed for prompt 0. Prompt i uses (seed + i) % 2**32, since numpy takes 32-bit seeds.
• caller (str) – What the user called, which errors name.

Returns:

One list of PIL images per prompt, in prompt order.

Raises:

• TypeError – If the pipeline’s model does not generate images, or images is not a list of image lists.
• ValueError – If there are no prompts, a prompt is empty or too long for the model, negative_prompt or images has the wrong length, a prompt has more than one input image, num_inference_steps or num_images_per_prompt is below 1, seed is outside [0, 2**32), a value is outside the range max serve accepts, the model generates video, or an input image goes to a model that takes text only. A tokenizer’s error names its prompt. Every prompt is checked before the first one runs.
• RuntimeError – If the pipeline returns output that is not an image.
• OSError – If an input image’s bytes do not decode, before any prompt runs.

Return type:

GenerateImagesResult