Skip to content

strands-sana

PAG

cagataycali/strands-sana

PAG — Perturbed Attention Guidance¶

PAG (Perturbed Attention Guidance) is a CFG-style guidance technique that improves quality without text conditioning. It costs ~2× compute.

from strands_sana import sana_generate

sana_generate(
    prompt="a stained glass dragon",
    model="sana-1.5-1.6b-1024",
    pag_scale=2.0,        # > 0 → switches to SanaPAGPipeline
    steps=15,
    seed=42,
)

How `pag_scale` interacts with `guidance_scale`¶

`pag_scale`	`guidance_scale`	Pipeline
0	any	`SanaPipeline` (regular CFG)
> 0	any	`SanaPAGPipeline` (PAG + CFG)

Recommended: pag_scale=2.0, keep guidance_scale=4.5 (default).

Demo¶

PAG output

_{Sana-1.5 1.6B + PAG (`pag_scale=2.0`) · 15 steps · 62s on Thor}

When to use PAG¶

✅ Subjects with intricate detail (jewelry, glass, fur)
✅ Off-distribution prompts where CFG alone overshoots
❌ Sprint models (already distilled)
❌ Speed-critical pipelines (~2× compute)