Skip to content

Photoshop beta: Rotate Object + Harmonize for 2D images (Adobe product announcement)

Adobe Evangelist Kris Kashtanova announces a Photoshop (beta) feature called Rotate Object that rotates 2D image content as if it were a 3D object, plus pairing with the existing Harmonize feature to re-light/re-shadow the rotated object to match the surrounding scene. Pitched as “Turntable [in Illustrator], but instead of vectors, it’s pixels in Photoshop.” A product demo, not a paper — no method details disclosed.

  • Rotate Object operates on rasterized 2D image content (not vectors), implying an implicit 3D-aware generative backbone that hallucinates novel views from a single image [tweet body].
  • Harmonize re-renders light and cast shadows on the rotated object to blend with the scene, suggesting a separate lighting/shadow re-composition pass on top of the novel-view step [tweet body].
  • The feature is shipping in Photoshop beta (March 2026); no technical report or model card is linked [tweet body].

Not disclosed. Inferred pipeline from the demo description: (1) segment / lift the target object from its background; (2) generate a novel-view rotation of the object (a single-image NVS step that likely depends on a video- or 3D-aware diffusion prior); (3) re-composite the rotated object onto the original background; (4) run Harmonize to adjust illumination, contact shadows and reflections so the composite reads as a single photograph. The composability story (Rotate → Harmonize as separate user-driven steps) is the only architectural hint.

No quantitative results. The tweet carries a video demo (1.3M views as of filing) showing rotation + harmonization on a single example. No comparisons, no failure cases shown.

Rotate-then-Harmonize is a productized version of the decomposition / re-compositing pipeline being studied in research artifacts like Generative Image Layer Decomposition with Visual Effects (LayerDecomp) and OmnimatteZero: Fast Training-free Omnimatte with Pre-trained Video Diffusion Models: lift an object off the background, manipulate it independently (here: a 3D rotation rather than a translation/swap), and re-render shadows + lighting to make the edit invisible. The fact that Adobe ships this as two separate UI primitives (Rotate Object, Harmonize) is consistent with Layered Image/Video Decomposition‘s recurring observation that effects (shadows, reflections, contact lighting) must be handled as a distinct stage from the object move itself — otherwise editing the foreground produces obvious compositing artifacts. The interesting open question is which prior powers the rotation step — a video diffusion model used for NVS, a multi-view 3D model, or an image-to-3D pipeline — none of which is disclosed here.