Inference Scaling¶

Generate K candidates, pick the best by CLIP score (or NVILA when ready).

from strands_sana import sana_inference_scale

result = sana_inference_scale(
    prompt="a cherry blossom branch in soft moonlight",
    model="sana-1.6b-1024",
    n_samples=4,
    score_fn="clip",    # or "first" to skip scoring
    seed_start=200,
)

print(result["best"])   # best candidate path + score
print(result["scores"]) # all 4 scores

How it works¶

graph LR
    P[Prompt] --> G[Generate K with seeds]
    G --> S[Score each with CLIP]
    S --> B[Return best + all]

    style B fill:#5e35b1,stroke:#333,color:#fff

Demo — 4 candidates¶

When to use¶

✅ Production where quality matters more than latency
✅ Difficult prompts (compositional, off-distribution)
✅ Building a "best-of-N" agent loop
❌ Real-time UX (latency is K× single inference)

Score functions¶

`score_fn`	Method	Speed
`"clip"`	OpenAI CLIP-vit-base-patch32	~0.5s/image
`"first"`	No scoring, return seed_start	0s

For more sophisticated scoring (NVILA-2B, ImageReward, MPS), add a custom score_fn — see API reference.