vision-squeezer 0.6.0

LLM-native image optimization — reduces vision model token consumption by snapping to tile boundaries

vision-squeezer (Python)

Python bindings for vision-squeezer — LLM-native image optimization. Reduces vision model token consumption by snapping to tile boundaries.

Install

pip install vision-squeezer

From source (requires Rust + pip install maturin):

cd python && maturin develop --release

Usage

import vision_squeezer as vs

# Optimize an image on disk and write next to it.
report = vs.optimize_image(
    "screenshot.png",
    model="claude",
    quality=75,
    output_path="screenshot.optimized.jpg",
)
print(f"Saved {report['tokens_saved']} tokens, {report['size_reduction_pct']:.1f}% smaller")

# Or work with raw bytes.
with open("screenshot.png", "rb") as f:
    raw = f.read()
report = vs.optimize_image(raw, model="gpt4o", format="webp")
# report["bytes"] holds the optimized image
# report["base64"] is the same image, base64-encoded

# Estimate tokens without optimizing.
est = vs.estimate_tokens(4096, 3072, model="claude")
print(est)  # {'model': 'claude', 'tokens': ..., 'tiles': ...}

# Optimal send dimensions for a given model.
print(vs.optimal_dimensions(4096, 3072, model="gpt4o"))

Supported models

• claude (3.5 / 4.x — area-based token estimation)
• gpt4o (GPT-4o / 4.5 — 2048 fit then 768 short-side, 512px tiles)
• gpt5 (GPT-5/5.5 — 6000px max, 10.24M pixel cap, 1536 token cap)
• gemini (Gemini 2.0/3.0 — flat 258 ≤384×384, else 768px tiles)

Output formats

jpeg (default) and webp. Format affects bytes, not tokens.