findpics-writer-2b

A 2B vision-language model that writes one line of search facts about a photo. It runs on an iPhone (MLX) inside the find pics app, which searches a person's photo library on the device: each photo is described once while the phone charges, and searches then read the lines instead of looking at every photo.

Format: MLX, language layers affine 4-bit (group 64), vision tower bf16; one 1.8 GB safetensors file.

Input and output

• Input: one photo, long side 448 px, and the prompt Search facts for this photo, one line.
• Output: one line of field: items groups joined by ; (scene, objects, colours, people, activity, text, look, event). About 74 tokens on average, 90th percentile about 100.
• Settings used in the app: greedy (temperature 0), at most 120 new tokens, repetition penalty 1.05 over a 4,096-token window. Decoding stops when a list item repeats for the third time inside a field.

How it was made

• Base model: Qwen/Qwen3-VL-2B-Instruct (Apache-2.0).
• LoRA fine-tune on about 359k (photo, line) pairs for 2 epochs, merged into the base, then converted to MLX.
• Target lines were written by Qwen/Qwen3.5-27B-FP8 (open weights).
• Training photos, all public: DISBench training libraries (Flickr / YFCC100M, Creative Commons licences), Open Images V4 (CC BY 2.0) and Wikimedia Commons iPhone photos (free licences). No private photos were used.

Use with mlx-vlm

pip install -U mlx-vlm
python -m mlx_vlm.generate --model rezashamji/findpics-writer-2b --max-tokens 120 --temperature 0.0 \
  --prompt "Search facts for this photo, one line." --image <path_to_image>

Limits

The lines are written for retrieval, not for display. They can miss small objects, get colours wrong, or misread text in the photo. Do not use them to make decisions about people.