- name
- Image Enhancement
- slug
- image-enhancement
- version
- 1.0.0
- category
- image
- description
- Apply a bounded image-enhancement operation such as upscaling, background removal, extension, or cleanup while preserving the intended subject.
- status
- ready
- muapi_capabilities
- ["media.enhance_image","media.upscale","media.upload_file","media.check_result"]
- required_connections
- ["muapi"]
- permissions
- ["draft-only","workspace-write"]
# Image Enhancement
## Mission
Improve or prepare an existing image for delivery without confusing an
enhancement with a creative re-generation. Route the request to the narrowest
MuAPI operation that can perform it.
## Supported request types
- upscale resolution;
- remove or replace a background;
- extend/outpaint a frame when the provider operation supports it;
- restore, colorize, clean up, or apply a named transformation when the live
model/tool explicitly supports it;
- prepare a final selected image for a target platform or delivery size.
Use [MODELS.md](../../MODELS.md) and
[the MuAPI tool reference](../../references/muapi-image-tools.md) for current
operations. Do not invent a parameter or assume that every enhancement tool
supports every source format.
## Required inputs
- Source image URL or local file.
- One explicit operation and the desired output size or condition.
- Details that must remain unchanged.
- Target channel or delivery format when relevant.
## Workflow
1. Confirm the source, operation, target, preservation constraints, and whether
the user wants a concept or delivery candidate.
2. Check rights and privacy before uploading a local image. Upload it with
`media.upload_file` when the chosen operation requires a hosted URL.
3. Choose the operation-specific tool/model. Use `media.upscale` only after a
candidate is selected; it is not a substitute for repairing geometry,
spelling, or an incorrect product.
4. State the call count, model/operation, estimated cost signal, and retry
policy before execution when the account or cost is unknown.
5. Submit and poll with `media.check_result`. Preserve request ID, status,
output URLs, billing, and errors.
6. Review edges, fine detail, noise, halos, transparency, crop, face/product
identity, text, and dimensions. For background changes, check hair, glass,
shadows, reflections, and holes.
7. Return the enhanced output with a clear note about what changed and what
still needs manual post-production.
## Decision rules
- Enhancement should not silently change a person's identity, a product's
geometry, a label, or a brand mark.
- Use editing for an intentional scene/object change and enhancement for a
bounded preparation operation.
- Do not upscale every variant in a concept batch; select first unless the
user requests otherwise.
- Face swap or likeness transformation is a rights-sensitive edit, not a
routine enhancement. Require explicit consent and record the source roles.
## Output format
Return the source, operation, model/tool, parameters, output URL(s), dimensions,
status, QA observations, receipt link, and any unresolved rights or delivery
limitations.
## Failure and missing-data behavior
If the operation is unsupported or the result is partial, report the exact
boundary and preserve the failed call. Do not label a failed upscale or
background removal as completed because a request ID was created.
Auf GitHub ansehen