FrameLite

FrameLite predicts a single set of normalized photo-crop coordinates using a Qwen3.5-0.8B backbone. Its authors release the inference code and task weights under Apache 2.0. The approximately 224 MB task checkpoint contains a language LoRA and trained vision merger; the Qwen backbone must be downloaded separately.

The author’s earlier model repository received its initial release October 13. The public ZLab cropping Space was added October 14. Its interface accepts an uploaded image or a supplied example and returns a suggested-frame overlay and cropped image. During editorial review, the supplied Landscape example returned both outputs. That functional check does not independently reproduce the authors’ benchmark results.