DIRcreative v0.8.0
DIRcreative v0.8.0 repairs the transition from professional planning to real image execution and adapts instruction handling to GPT-6 Astra.
- Retains the selected specialist during execution, removes the first-image circular dependency, and supports standalone image plans without a fictional film timeline.
- Binds execution to project scope, asset roles, source dialogue, provider inputs and reference bytes. Rejected or stale inputs stop before generation.
- Defaults character identity portraits to a frontal view and rejects unintended transparency in neutral-background masters. Original image bytes are preserved for review.
- Keeps short Chinese edits lightweight. Sepia is selected for demonstrated structural or discourse needs, with voice and factual constraints preserved.
- Covers optional image/schema backends and dependency-free fallback behavior across Python 3.10, 3.12 and 3.14.
Validation: full source and installed-package checks, exact-commit artifact verification, ADCO bilateral checks, independent reviews, real-provider compile/replay tests, and Linux/macOS byte-identical packaging. The tag and archive bind commit 3517875b9462eb720ad5f87497fa086428c5dad7.
The separate five-call image workflow verified actual generation, frontal/opacity fixes and four-body detection. Its final character candidate still missed the requested 3:1 canvas and portrait crop; it remains unapproved for downstream master use. This release does not claim final-film quality, video-generation acceptance or an exact-release media certificate.
Install using the release archive and SHA256SUMS, verified by the exact tagged source as described in the README. Sepia, Jingzao and Seedance specialist Skills remain external dependencies; they are not bundled or silently upgraded.