SpatialEdit-16B: Geometry and Evaluation Contract¶
The SpatialEdit paper presents a geometry-driven image-editing benchmark, synthetic supervision, and a SpatialEdit-16B baseline for object-centric and camera-centric edits. The official release publishes code, checkpoints, benchmark assets, and separate external prerequisites. This supports evaluation of the named release; it does not make a plausible image proof that the requested geometry was followed.
Bind the release and edit contract¶
For every evaluated or published edit, retain:
- repository revision, model/checkpoint identifier and digest, code and artifact terms, dependency/runtime record, and every external prerequisite used by the chosen configuration;
- source-asset digest, ownership or authorized-edit record, subject/object definition, protected regions, and any prohibited semantic changes;
- task class: object translation, scaling, canonical rotation, or camera viewpoint/framing change; source and target coordinate systems; target box, pose, crop, resolution, and orientation convention;
- input preprocessing, VAE/model paths, seed, scheduler/settings, output digest, and mapping from output pixels back to the source frame; and
- geometric-fidelity result, non-target preservation result, failure examples, reviewer decision, and permitted-use conclusion.
Code availability or a permissive repository license does not establish terms for every checkpoint, base model, dataset, or source image. Keep each authority record separate.
Evaluate geometry and preservation separately¶
Use held-out, source-disjoint assets that cover the required subject classes, occlusion, reflections, fine boundaries, background complexity, and target camera/object transformations. Score the requested transformation in its declared coordinate system, then independently inspect object identity, background continuity, lighting/shadow consistency, text, logos, and factual product detail.
Do not collapse a visual-preference score into geometric correctness. A result may look coherent while its object location, scale, orientation, or view change is wrong. Compare against a declared baseline and retain both successful and failed cases rather than selecting only attractive outputs.
Gotchas¶
- Issue: A prompt can yield a semantically plausible scene with the wrong spatial result -> Fix: validate the requested box, pose, framing, or camera relation in the stated source coordinate system.
- Issue: The official release includes placeholder/internal paths and external checkpoint prerequisites -> Fix: replace them only in a versioned local configuration and record every resolved artifact before treating a run as reproducible.
- Issue: A general editing result is assumed to transfer to a new object/camera operation -> Fix: evaluate that operation on an operation-specific held-out set; do not infer it from the model family.
Failure boundary¶
If source authority, exact release/terms, coordinate definition, preprocessing record, or geometry-and-preservation review is missing, keep the output in review state. Do not silently substitute another model, reinterpret a semantic-only edit as metric control, or publish it as a verified view or factual reconstruction.