Merge commit 'bf8f76505fdd25ea99770f262eb9371509bf226d' as 'vendor/dumas_h3_continuum'

This commit is contained in:
2026-10-20 13:43:38 +00:00
481 changed files with 131517 additions and 0 deletions
@@ -0,0 +1,14 @@
[Chunk 1]
An eye-level medium-wide camera tracks alongside the red-coated woman in @R1 and the navy-top woman in @R4 as they walk from screen left to screen right through a bright covered walkway. Preserve their distinct faces, hairstyles and clothes; adapt their poses to walking. Both remain clearly visible and keep walking as the chunk ends, with open space on the right.
overall_soundscape: Footsteps and quiet walkway ambience. No dialogue.
non_diegetic_music: N/A
[Chunk 2]
Continuation of Chunk 1. Keep the same tracking camera, lighting and walking direction. As the chunk begins, the teal-coated woman in @R9 enters from screen right and joins the red-coated woman in @R1 and the navy-top woman in @R4. Preserve all three distinct faces, hairstyles and clothes. Show all three together through the middle of the chunk. Then the navy-top woman turns into a side passage and walks out of view before the chunk ends. The red-coated and teal-coated women keep walking together.
overall_soundscape: Continuous footsteps and quiet walkway ambience. No dialogue.
non_diegetic_music: N/A
[Chunk 3]
Continuation of Chunk 2. The red-coated woman in @R1 and the teal-coated woman in @R9 continue walking side by side in the same direction. Preserve their distinct faces, hairstyles and clothes, and keep the same eye-level medium-wide tracking camera and lighting. Both remain clearly visible through the end. The side passage recedes behind them as their steady walk continues.
overall_soundscape: Continuous footsteps and quiet walkway ambience. No dialogue.
non_diegetic_music: N/A
@@ -0,0 +1,36 @@
# Structured reference format
Use only when requested or useful. Each chunk contains the following English fields:
```text
[Chunk 1]
subject_definitions:
<Subject 1> is the red-coated woman in @R1.
<Subject 2> is the navy-top woman in @R4.
summary:
[reference generation] Two women walk together through a covered walkway.
retention_analysis:
<Subject 1>: partially_preserved — preserve face, hairstyle and clothing; adapt pose and gaze to walking.
<Subject 2>: partially_preserved — preserve face, hairstyle and clothing; adapt pose and gaze to walking.
detailed_description:
[Shot 1] An eye-level medium-wide camera tracks alongside both women as they walk from screen left to screen right. Both remain clearly visible. They continue walking as the chunk ends, with open space on the right for a third person to join next.
overall_soundscape:
Footsteps and quiet walkway ambience. No dialogue.
non_diegetic_music:
N/A
```
Keep Subject 1/2 meanings in later chunks. A new person from R9 can be Subject 3; do not rename her Subject 2 because she becomes Picture 2. Re-establish only the needed source binding in each selected prompt. For a continuing person without an active reference, describe inherited identity/state without an inactive @R tag.
Use retention markers with a concrete scope: partially_preserved for preserved identity/clothing with changed motion/background; attribute_transfer for applying clothing or a prop from another source; fully_preserved only when the requested content actually warrants it. These are semantic instructions, not hard masks or guaranteed constraints.
For a bag reference: `<Subject 4> is the tan bag in @R2, worn by <Subject 1>.` In the action text keep this ownership. If the character's source image already contains another bag, explicitly replace that source accessory with the selected bag, while preserving the person's identity and clothes.
In Chunk 2+, start detailed_description with `Continuation of Chunk 1.` (use the actual preceding number) plus the inherited physical state. Keep the six-field order; the continuation sentence is not a parser requirement. Use new shot markers/times only for intentional cuts. Do not invent `<Video 1>` for preceding generated chunks.
Official MiniMax Reference guide uses these six fields and recommends substantial scene detail (normally 350–500 English words in detailed_description for generation). Compact per-chunk wording here is a Continuum adaptation to test, not an officially proven better length. A short simple prompt remains valid input to Continuum.
Sources checked for the design:
- https://github.com/MiniMax-AI/MiniMax-H3/blob/main/skills/h3-prompt-writing/references/ref-en.txt
- https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/docs/VIDEO_PROMPT_WRITING_GUIDE_ref_en.md
- Local V3.9: v3/reference_images_v39.py, v3/driving_nodes.py, v3/reference_runtime.py, v3/reference_effective_plan.py, v2/prompts.py (2026-09-28). V3.9 is a local pre-release implementation, not a promise about every public package named 3.9.