Image Generation¶

The core tool is sana_generate. It auto-dispatches to PAG when pag_scale > 0, handles aspect-ratio binning, and supports up to 4K with VAE tiling.

sana_generate(prompt: str, ...) -> dict

Basic call¶

from strands_sana import sana_generate

result = sana_generate(
    prompt="a serene koi pond, ink wash painting",
    model="sana-1.6b-1024",
    steps=20,
    seed=42,
)

All parameters¶

arg	type	default	notes
`prompt`	`str`	required	text description
`negative_prompt`	`str`	`""`	what to avoid
`model`	`str?`	`None`	alias from registry
`width` / `height`	`int?`	model resolution	snapped to bucket
`steps`	`int?`	per-model (20 default)	denoising steps
`guidance_scale`	`float?`	per-model (4.5)	CFG strength
`seed`	`int?`	`None`	reproducibility
`pag_scale`	`float`	`0.0`	>0 → SanaPAGPipeline
`use_complex_instruction`	`bool`	`False`	prepend Gemma-2 prompt expander
`use_resolution_binning`	`bool`	`True`	snap to trained bucket
`num_images`	`int`	`1`	per prompt
`output_dir`	`str?`	`./generated`	save directory

Multi-image batch¶

sana_generate(
    prompt="a clay figurine of a frog",
    model="sana-0.6b-512",
    num_images=4,
    seed=20,
)
# Returns: { "paths": [path1, path2, path3, path4], "count": 4, ... }

Multiple prompts¶

For different prompts, use sana_batch:

from strands_sana import sana_batch

sana_batch(
    prompts=["a duck", "a frog", "a cat"],
    model="sana-0.6b-512",
    seed=42,
)

Different aspect ratios¶

sana_generate(prompt="a cinematic landscape",
              width=1280, height=768, model="sana-1.6b-1024")

use_resolution_binning=True (default) snaps to the closest trained bucket — for Sana-1024 buckets are e.g. (1024, 1024), (1280, 768), (768, 1280), etc.

With Gemma-2 prompt expansion¶

Sana uses a Gemma-2 LLM as text encoder. Pass use_complex_instruction=True to prepend the official prompt-expansion template — short prompts get fleshed out automatically.

sana_generate(
    prompt="a duck",   # short
    model="sana-1.6b-1024",
    use_complex_instruction=True,  # expanded server-side via Gemma-2
)