imgslim 0.1.0

Reduce vision API costs by 60-80% with algorithmic image optimization

imgslim

Reduce your vision API costs by 60–80% before the image ever leaves your machine.

No AI. No external calls. Pure algorithmic compression tuned for how LLMs actually tokenize images.

pip install imgslim

from imgslim import VisionLite

result = VisionLite(model="claude").optimize("screenshot.png")
# {'output_path': '/tmp/imgslim_output/screenshot_optimized.png',
#  'savings_pct': 74.3,
#  'strategy_used': ['strip_exif', 'resize', 'grayscale']}

---

Why this exists

Vision APIs charge by image tokens. Claude, GPT-4o, and Gemini all slice images into tiles before processing — the larger the image, the more tiles, the higher the bill.

Most images sent to LLMs are massively over-sized for what the model actually needs to understand them. A 3000×2000 photo contains the same semantic information as a 1092×728 one, from the model's perspective.

imgslim exploits the exact tile thresholds of each model to resize images to the optimal resolution — plus strips EXIF metadata, removes whitespace margins, and converts to grayscale when color adds nothing.

---

Benchmark

Tested on 5 synthetic content types × 3 models. All transformations are algorithmic — no AI used.

Model	Sample	Input KB	Output KB	Saving %	Strategies
claude	diagram	95 KB	38 KB	59.4%	strip_exif, resize_to_tile_limit, convert_to_grayscale
claude	mixed	678 KB	225 KB	66.8%	strip_exif, resize_to_tile_limit, convert_to_grayscale, crop_whitespace
claude	photo	6877 KB	226 KB	96.7%	strip_exif, resize_to_tile_limit
claude	screenshot	98 KB	21 KB	78.4%	strip_exif, resize_to_tile_limit, convert_to_grayscale
claude	text_doc	934 KB	161 KB	82.8%	strip_exif, resize_to_tile_limit, convert_to_grayscale
gemini	diagram	95 KB	81 KB	14.9%	strip_exif, resize_to_tile_limit, convert_to_grayscale
gemini	mixed	678 KB	300 KB	55.8%	strip_exif, resize_to_tile_limit, convert_to_grayscale, crop_whitespace
gemini	photo	6877 KB	563 KB	91.8%	strip_exif, resize_to_tile_limit
gemini	screenshot	98 KB	35 KB	64.1%	strip_exif, resize_to_tile_limit, convert_to_grayscale
gemini	text_doc	934 KB	306 KB	67.3%	strip_exif, resize_to_tile_limit, convert_to_grayscale
gpt-4o	diagram	95 KB	36 KB	61.9%	strip_exif, resize_to_tile_limit, convert_to_grayscale
gpt-4o	mixed	678 KB	194 KB	71.4%	strip_exif, resize_to_tile_limit, convert_to_grayscale, crop_whitespace
gpt-4o	photo	6877 KB	190 KB	97.2%	strip_exif, resize_to_tile_limit
gpt-4o	screenshot	98 KB	18 KB	81.8%	strip_exif, resize_to_tile_limit, convert_to_grayscale
gpt-4o	text_doc	934 KB	138 KB	85.2%	strip_exif, resize_to_tile_limit, convert_to_grayscale

Average saving: 71.7% across all content types and models.

Benchmark script included: python benchmark/run_benchmark.py

---

How it works

1. Content detection (no AI)

OpenCV heuristics classify the image as photo, text, or diagram based on edge density, contour count, and line detection.

2. Strategy selection

All images   → strip_exif + resize_to_tile_limit
text/diagram → + convert_to_grayscale
text         → + crop_whitespace

3. Model-aware resize

Each model has documented tile thresholds. imgslim resizes to just below the optimal boundary — never upscales, always preserves aspect ratio.

MODEL_TILE_LIMITS = {
    "claude": {"max_size": 1568, "optimal_long_edge": 1092},
    "gpt-4o": {"max_size": 2048, "optimal_long_edge": 1024},
    "gemini": {"max_size": 3072, "optimal_long_edge": 1536},
}

---

Usage

from imgslim import VisionLite

# Default: optimized for Claude
v = VisionLite(model="claude")
result = v.optimize("invoice.jpg")

print(result["savings_pct"])      # 81.2
print(result["strategy_used"])    # ['strip_exif', 'resize', 'grayscale', 'crop']
print(result["output_path"])      # /tmp/imgslim_output/invoice_optimized.jpg

# Then pass output_path to your API call as usual

Supported models: claude, gpt-4o, gemini

---

Install

pip install imgslim

Dependencies: Pillow, opencv-python-headless, piexif

Python 3.10+

---

What imgslim does NOT do

• No quality assessment (does not verify the model still "understands" the image)
• No batch processing yet
• No async support yet
• Does not tune JPEG quality dynamically

These are on the roadmap. PRs welcome.

---

License

MIT