nodes/exportAnything-ComfyUI-DiffusionGemmaPromptBuilder
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Add separate MiniMax Music 3 advertisement workflow

Browse Source
This commit is contained in:
danrh
2026-08-24 02:48:59 -05:00
parent aaeec80037
commit db71a81524
41 changed files with 26505 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/DEFERRED_AGENT_SKILL_BRIEF.md
+18 -3
View File
@@ -1,21 +1,22 @@
# Deferred DiffusionGemma Agent Skill Brief
Status: **Pinned for post-hardening implementation.** This document is intentionally outside `.agents/skills`, so Codex and other compatible agents do not discover or invoke an operating contract that is not finished yet.
Status: **Pinned for post-hardening implementation; inactive.** This document is intentionally outside `.agents/skills`, so Codex and other compatible agents do not discover or invoke an operating contract that is not finished yet. The separate Advertisement workflow expands the future scope but does not activate or install this Skill.
Pinned: 2026-08-23
Last reviewed: 2026-08-24
## Canonical skill identity
```yaml
name: diffusiongemma-director
description: Configure, run, validate, and troubleshoot DiffusionGemma Director LTX and MiniMax-H3 workflows in a user's local ComfyUI through an available local operations interface. Use for this package's target profiles, reference preparation, soundtrack and performance routing, native-shot and generation-lane planning, identity relay, contract failures, queueing, or output verification; do not use for unrelated ComfyUI graphs or generic prompt writing.
description: Configure, run, validate, and troubleshoot DiffusionGemma Director LTX, MiniMax-H3 music-video, and separate Advertisement workflows in a user's local ComfyUI through an available local operations interface. Use for this package's target profiles, typed campaign and reference contracts, soundtrack source and performance routing, native-shot and generation-lane planning, identity/product continuity, deterministic finishing, media QA, contract failures, queueing, or output verification; do not use for unrelated ComfyUI graphs, generic prompt writing, or ungoverned advertising claims.
```
Treat that `name` and `description` as the canonical starting point when the active skill is created. Revise them only if the hardened public surface materially changes the skill's scope or trigger boundary.
## Intended outcome
The eventual skill should translate an ordinary-language production request plus user-supplied images, video, or audio into a correctly configured DiffusionGemma workflow; operate it through an available local ComfyUI interface; diagnose contract or runtime failures; and verify actual branch media output rather than treating a completed queue as proof of generation.
The eventual skill should translate an ordinary-language production request plus user-supplied images, video, or audio into a correctly configured DiffusionGemma workflow; operate it through an available local ComfyUI interface; diagnose contract or runtime failures; and verify actual branch media output rather than treating a completed queue as proof of generation. For Advertisement requests it must also preserve exact approved copy and claims, distinguish performer and product authority, separate the H3 motion guide from final voice-over, and report delivery evidence honestly.
The skill supplies operating knowledge. It does not itself grant filesystem, browser, model-download, or ComfyUI execution access. Those capabilities must come from explicitly available tools, ideally a local ComfyUI MCP server declared as a plugin dependency.
@@ -27,6 +28,13 @@ The skill supplies operating knowledge. It does not itself grant filesystem, bro
- Treat Natural, Dance, and Lyrics as mutually exclusive performance authorities. Written lyrics are not a timing schedule without verified timed-lyrics evidence.
- Keep clean identity references authoritative. Previous-lane relay is optional continuity evidence, not a replacement identity source.
- Preserve the selected soundtrack source, pristine final audio, exact retained-frame clock, and user-selected camera intent.
- Keep the known-good V6 workflow and generic contracts separate from Advertisement contracts. Never migrate or overwrite V6 to satisfy an ad request.
- For Advertisement references, preserve Picture 1 as performer hero, Picture 2 as same-performer sheet, Picture 3 as independent product/package sheet, and Picture 4 as retained-tail continuity only. Never make Picture 4 a new identity or product authority, and never combine performer and product sheets to save a reference slot.
- Distinguish the three Advertisement soundtrack origins—MiniMax Music 3, Upload song, and Legacy ACE-Step—from the `Instrumental` / `Vocal` / `Auto` content policy. Do not require vocal evidence for a governed instrumental.
- Send only the selected soundtrack to H3 as the motion guide. Optional non-diegetic VO belongs only in the final delivery mix and must not become dialogue or lip-sync conditioning.
- Keep exact brand, CTA, and legal copy in the deterministic end card. Never invent or strengthen a product claim, and never represent generated in-scene typography as exact.
- Treat the 30-second master plus rendered 15-second and 6-second temporal cutdowns as media outputs. Treat 1:1 or 16:9 adaptations as unrendered until an actual semantic reframe renderer produces them.
- Preserve `pass`, `fail`, and `not_measured` as distinct QA states. A completed queue, valid graph, or successful Music 3 canary is not evidence that performer identity, product identity, copy legibility, sync, or the complete H3 campaign passed.
- Require fresh authorization before downloads, installs, external publication, destructive changes, or other actions not already placed in scope.
## Hardening exit criteria
@@ -39,6 +47,12 @@ Promote this brief to `.agents/skills/diffusiongemma-director/SKILL.md`, or pack
- validation errors, cache refresh/reuse behavior, lane assembly, and media-output checks have repeatable tests;
- required versus optional models and custom-node packs are expressed without machine-specific paths;
- representative Natural, Dance, Lyrics, upload-song, identity-relay, long-duration, and failure-recovery requests have been forward-tested.
- the Advertisement workflow has clean-install coverage on ComfyUI `>=0.33.1`, including the locally verified MiniMax Music 3 INT8/ConvRot text-encoder path and all three soundtrack origins;
- Picture 1/2 performer authority, Picture 3 product authority, Picture 4 retained-tail relay, soundtrack-only H3 guidance, optional final-mix VO, deterministic end cards, actual 15-second/6-second cutdowns, and honest QA states have repeatable integration tests;
- at least one complete Advertisement H3 render—not only an audio canary or serialized fixture—has been reviewed for performer/product retention, copy legibility, audio sync, and exact delivery clocks;
- model-generated package text, stochastic reference drift, missing semantic reframing, and unmeasured QA evidence remain documented as limitations rather than silently repaired or marked passed.
Current hardening evidence: the exact release API graph has completed a full two-lane 30-second Advertisement render with retained-tail relay, deterministic end card, exact 30/15/6 media clocks, and all diagnostic roots. A sampled visual contact-sheet review found stable performer/wardrobe and recognizable product continuity. The active Skill is still deferred because evidence-backed listening/audio-sync approval, representative mode/source coverage, and the remaining clean-install matrix above are not complete; the workflow QA result remains `not_measured`, not promoted to `pass`.
## Expected active-skill structure
@@ -48,6 +62,7 @@ diffusiongemma-director/
|-- agents/
| `-- openai.yaml
`-- references/
|-- advertisement-workflow.md
|-- workflow-modes.md
|-- node-contracts.md
`-- troubleshooting.md