PhotoFramer: Multi-modal Image Composition Instruction

TL;DR

PhotoFramer is a CVPR 2026 project page for a multi-modal composition-instruction model that, given a poorly composed photo, returns both a textual guidance describing how to recompose it and a generated example image of the improved framing. The framework decomposes composition guidance into three photographer-motivated sub-tasks — Shift, Zoom-in, and View-change — corresponding to subject placement, focal choice, and vantage point. Training data is curated by sampling Shift/Zoom-in pairs from existing image-cropping datasets and synthesizing View-change pairs from multi-view datasets via a learned degradation model that maps well-composed expert photos to plausible poor compositions. The model is a unified text+image generator fine-tuned on this dataset; experiments report that textual instructions effectively steer composition and that pairing text with example images improves over exemplar-only baselines. Data is not yet released.

Key claims

Composition guidance can be hierarchically decomposed into three sub-tasks — Shift (subject placement / border distractions), Zoom-in (tighter crop simulating longer focal length), View-change (new vantage point or camera pose) — corresponding to the three photographic factors of vantage point, focal choice, and subject placement [project page §“three tasks”].
Shift and Zoom-in training pairs are sourced from existing image-cropping datasets, while View-change pairs are constructed by a two-stage pipeline that first samples viewpoint pairs from multi-view datasets and then trains a degradation model to turn well-composed expert photos into poorly composed ones [project page §“dataset”].
A single model that jointly processes and generates text and images is fine-tuned on this dataset to produce textual guidance plus an example image as joint outputs [project page §“PhotoFramer Model”].
Textual instructions effectively steer image composition, and coupling textual guidance with exemplar images yields consistent improvements over exemplar-only baselines [project page §“Experiments”].

Method

PhotoFramer is set up as a multi-modal composition-instruction system. Input: a single poorly composed photo. Output: (a) a natural-language description of how to improve composition, and (b) a generated image illustrating a better composition. Three sub-tasks are defined — Shift, Zoom-in, View-change — chosen to match the basic photographic factors (subject placement, focal choice, vantage point).

Training data is assembled task-by-task. Shift and Zoom-in pairs are mined from existing cropping datasets, where the original frame supplies the poorly composed example and the human-selected crop supplies the well-composed one. View-change is harder because real “same-scene, different-pose” pairs in the wild rarely come with a quality label. The authors handle this with a two-stage synthesis pipeline: (1) sample pairs with varying viewpoints from multi-view datasets and use them to train a degradation model that learns to map a well-composed photo to a plausibly poorly composed one of the same scene; (2) apply this degradation model to a separate corpus of expert-taken photos, turning each into a synthetic poor/good pair. The model is then a unified text-and-image generator fine-tuned on the resulting dataset of (poor image → text instruction + good image) triples.

Results

The project page reports qualitative samples for each of Shift, Zoom-in, and View-change and an “Composition Assessment Model” plus the “PhotoFramer Model” sections; quantitative numbers are not surfaced on the page beyond the claim that “textual instructions effectively steer image composition” and “coupling them with exemplars yields consistent improvements over exemplar-only baselines” [project page §“Experiments”]. Data is not yet released at filing time.

Why it’s interesting

PhotoFramer is the rare paper that treats composition — not subject identity, not motion, not 3D geometry — as the first-class generative target, and frames it as joint text+image output. It contrasts cleanly with CameraBench: Towards Understanding Camera Motions in Any Video: CameraBench taxonomizes camera motion in video for evaluation, while PhotoFramer taxonomizes camera framing in stills for instruction — the same cinematographer-prior axis applied to two different output spaces. The View-change sub-task is also a natural client for the camera-control adapter cluster (Camera-Controlled Video Diffusion, Kling-MotionControl Technical Report), which would supply a stronger generator for the “what would this shot look like from a different pose” sub-problem than image-level inpainting alone. The training-data construction is a clean example of the Synthetic Training Data recipe — a learned degradation model used to manufacture in-distribution poor/good pairs from expert photos when the natural pairing doesn’t exist — and the unified text-image-generation backbone places it alongside the Unified Multimodal Models cluster.