FrameLite's new demo lets a 0.8B model pick your photo crop

A small model with a concrete photography task

FrameLite now has a public Hugging Face demo that lets visitors ask a compact vision-language model to choose a photo’s framing. Upload an image or select an included example, press Find my crop, and the interface returns both a boundary overlay and the cropped image. The Space’s deployment history dates that interactive workflow to October 14.

The model is built on Qwen3.5-0.8B and predicts one crop through greedy decoding. Its authors say inference needs no candidate search or external scoring model. Their three-stage training first teaches spatial and composition preferences, then uses quality-guided supervision and optimization relative to the model’s current greedy answer.

Try the workflow, keep the download size straight

During editorial review, the public Landscape example completed successfully: the interface displayed a suggested frame and a downloadable crop, reporting a change from 768 by 1,024 pixels to 768 by 564 pixels. This confirms that one supplied example worked; it is not a benchmark replication or a guarantee for arbitrary photographs.

The model card’s approximately 224 MB download is the task checkpoint, containing a language LoRA and trained vision merger. Users running it locally also need the separate Qwen backbone. The documentation describes Python 3.10 and BF16-capable CUDA hardware, with a slower FP32 CPU option in the demo instructions. The hosted Space uses ZeroGPU and may queue.

FrameLite’s authors report outperforming the compared large vision-language croppers on public benchmarks. Those comparisons remain author-reported. Code and task weights use Apache 2.0. The arXiv submission and earlier author weight upload are dated October 13; today’s development is the interactive demo, rather than a newly invented model release date.