Guide / Targeted-edit evaluation / editorial
Seedance 2.5 targeted editing: define the change before testing.
Seedance 2.5 targeted editing begins with an authorized source, one declared change, and frozen visual and audio invariants; the term does not prove that a named edit control or preservation behavior is currently available.
Evaluate a Seedance 2.5 targeted-editing claim with authorized source media, a declared change mask, preserved invariants, matched review, and no inferred capability.
Write the intended delta and frozen invariants
Start with an authorized source file and a rights record for every visible person, voice, brand, location, and supplied asset. Describe one requested change in observable terms: the object, region, time range, spoken segment, or identity attribute that may change. Then freeze everything that should remain stable, including duration, shot order, framing, camera motion, background geometry, unaffected people, wardrobe, lighting, lip timing, ambience, music, captions, and delivery encoding as relevant. Draw a spatial or temporal mask when words alone are ambiguous. This contract makes leakage review possible. The sanitized the public discussion signal only records circulation of the phrase; it does not establish that Seedance 2.5 or SEELE can perform the edit.
- Authorized immutable source
- One observable intended change
- Spatial and temporal edit scope
- Explicit visual and audio invariants
Verify the product surface before creating a branch
Open the current product context and record the visible provider, exact model label, account, region, date, accepted inputs, available edit controls, limits, and terms. No dated first-party evidence already in this repository establishes a Seedance 2.5 targeted-editing feature, so this page supplies no specification, access, or behavior claim. If the interface does not expose a relevant, authorized method, stop and record non-availability in that observed context; do not relabel regeneration, manual compositing, or another model's function as targeted editing. If a test can proceed, preserve the source untouched, create a separately identified branch, save instructions and settings, and keep rejected or failed outputs rather than selecting only a favorable example.
Review the requested change and collateral drift separately
Compare source and candidate at matched timecodes, normal playback, slow playback, and frame steps. Score whether the requested delta is present, absent, partial, or ambiguous. In a separate pass, inspect all frozen invariants for spatial spill, temporal flicker, geometry changes, identity drift, altered expression, wardrobe or prop mutation, background reconstruction, camera shift, unintended dialogue edits, lip-sync displacement, changed ambience, new artifacts, and timing or encoding differences. Report the exact source, interface, date, instructions, branch, and reviewer. A conclusion belongs only to that run and edit contract; it cannot be generalized into quality, repeatability, adoption, campaign, availability, or commercial-performance claims.