Start a new job
Regenerate a masked logo/detail on an uploaded image, using user-uploaded
reference shots as ground truth. Produces num_candidates variations, each a
standard ImageResult shown in the Gallery.
token_authAuthorizationBearer <token>Token-based authentication using access tokens.
application/json- body
image*ImageUUID of the uploaded image to fix (the detail/logo lives here).
references?array<string>UUIDs of reference images where the detail is clearly visible (ground truth). Optional — the lambda can fix some details from the prompt alone, but a reference makes the result far more accurate (and avoids hallucinating).
items <= 10mask?|The marked region: {bbox_norm: [x, y, w, h], strokes: [...]} in 0..1 coords.
Optional when mask_image is given (the API binary-mask path).
nullmask_image?|UUID of an uploaded binary mask image (white = the area to repair) at the
image-to-fix resolution. Pixel-precise alternative to mask for API pipelines —
no UI strokes needed. Resolved to its S3 key and forwarded to the engine, which
derives both the region shape and its bounding box from it.
nullproject_id*Project IdProject key where images are stored.
num_candidates?Num CandidatesHow many variations to generate (1-4).
1model?ModelGeneration engine (auto | nano_banana_2 | nano_banana_pro | seedream | gpt_image).
"auto"category?CategoryPrompt family / issue category. logo (default) and text/artifact/... use
the logo-optimized prompt; general switches to general-purpose inpainting (any
masked edit — hairstyle, garment colour, object removal, ...) driven by
reviewer_note, with references showing the desired result. Free-form.
"logo"image_notes?array<string>Optional brand/style hints forwarded to the prompt.
[]reviewer_note?|Optional targeted instruction (e.g. 'fix the first letter').
nulladd_ai_watermark?Add Ai WatermarkBake an 'AI-generated' disclosure mark into each output (irreversible).
falsemode?ModeGeneration strategy (enterprise only). guided-blend = replace only the masked area and blend the seam (default); guided = redraw the whole crop; annotated / marked = single-image variants. Ignored for non-enterprise.
"guided-blend"context_factor?|How much surrounding image the model sees (enterprise only): crop side = context_factor × mask size. None → engine default. Ignored for non-enterprise.
nullSuccessful Response
application/json- response
job_id*Job Idstatus*Statusmessage*Messagecurl -X POST "https://example.com/detail-repair" \ -H "Content-Type: application/json" \ -d '{ "image": "string", "project_id": "string" }'{ "job_id": "string", "status": "string", "message": "string"}