In image generation, when do you use mask inpainting instead of conversational editing?

## Two mechanisms, two guarantees The decision is not about which produces a nicer picture. It is about what each one *promises*, and the promises are categorically different. **Mask inpainting.** You supply the source image, a binary or feathered mask, and a prompt. The pipeline encodes the image, re-noises the masked region, denoises it with the surrounding context and the prompt as conditioning, decodes, and composites the result so that pixels outside the mask come from the original file. The guarantee is exact: the untouched area is unchanged, bit for bit, because it was never regenerated. Outpainting is the same mechanism with the mask outside the original frame. **Conversational multi-reference editing.** You supply the source, a set of reference images, and an instruction in natural language across a turn-by-turn thread. The model regenerates the entire frame while carrying subject identity and scene context forward from the references and previous turns. As of mid-2026 hosted models in this class accept on the order of a dozen reference images, support localized instructions in words rather than masks, and offer camera and lighting direction. The guarantee is *semantic* — that the chair is recognisably the same chair — not *pixel-level*. ## When the mask is mandatory Any time some part of the image is a matter of record. Approved product photography where the product geometry has been signed off. A photograph with a legal caption. A medical or insurance image where the untouched region is evidence. Regulated packaging where the label must be the real one. Compliance workflows where you must be able to say precisely which pixels changed, and diff the file to prove it. Masks are also right for small, well-localized swaps in a large frame — replace the rug, remove a light switch, change the vase — where regenerating the whole image would be a large amount of compute and a large amount of risk to buy a small change. ## When conversational editing wins When the change is inherently global. Relighting a scene, changing the season visible through a window, moving the camera, restyling a room around a fixed product: none of these decompose into a region you could draw a boundary around, and a mask would produce a hard edge between two mutually inconsistent lighting models. Also when *identity across a set* is the actual requirement. A catalogue that needs one armchair in six room styles is not six independent generations and not six masked edits — it is one subject carried through a series, which is exactly what reference-image conditioning and multi-turn threads are built for. Asking for 'the same chair, same room, now in oak' and getting a coherent variant is a different capability from region replacement. And when there is no one to author masks. Mask production is real labour: either a person drawing them, or a segmentation model plus a review step for its mistakes. For high-volume or self-serve workflows that pipeline may cost more than the generation. ## The failure modes to name *Mask inpainting.* Boundary seams where the regenerated region does not match the original in lighting, grain, colour temperature or focus. Context bleed, where the model continues the surroundings into the mask instead of following the prompt. Scale confusion when the mask is small — the model has little context and produces an object at the wrong size. Semantic incoherence, because nothing outside the mask can adapt: put a lamp in and no shadow appears anywhere else in the room, because those pixels are frozen by construction. Feathering the mask edge softens seams but blurs the guarantee. *Conversational editing.* Drift: each turn is a fresh generation, so details you never mentioned wander — a cushion changes weave, a wall shifts a shade, a logo deforms slightly. Over a long thread this compounds until the output no longer matches the approved original. Identity preservation is good but not exact, which is fatal for a product whose geometry is the point. Resolution and framing can shift between turns. And the fact that nothing is pinned means you cannot prove which parts are original. ## The hybrid, which is usually the right production answer Explore conversationally, finalise with masks. Use multi-turn reference editing to find the composition and lighting you want, then, once an image is approved, restrict all further changes to masked regions so the approved content is frozen. For a catalogue: generate the six room styles conversationally around a fixed product reference set, get sign-off, and thereafter make seasonal swaps — different rug, different throw — as masked edits on the approved frames. This keeps the exploratory flexibility where it is cheap and the guarantee where it is required. A second hybrid worth knowing: generate conversationally, then composite the approved product photograph back in over the generated scene, with a masked pass only to reconcile shadow and contact edges. That gives literal product fidelity with a generated environment. ## How to answer the question Lead with the guarantee, not the tooling. Mask inpainting is the only one of the two that promises untouched pixels, so it is the answer whenever some region must be provably unchanged; conversational multi-reference editing is the answer when the change is global or identity must carry across a set, and its cost is drift and the absence of any pixel-level promise.