Merge remote README guidance into V3.5.0 release

This commit is contained in:
ukr8b3g-cmyk
2026-08-24 14:54:04 +09:00
2 changed files with 89 additions and 7 deletions
+1 -1
View File
@@ -69,7 +69,7 @@ b470819f16fc179654dc1600721a63f0a5a14e777f4c004d574b499da6886ba5 public_api.py
16bc48b9ef6127af7e0e3ddd96791640cedf5159cd7c06452d72604f41998a52 pyproject.toml
94fdc2d545b2d4077a4d70229fd22ba3ccac21d49d1097d2873f76d04e9a3c7a pytest.ini
365f1b1ecc89f8f2ced338ce467d02384ef612441202a15064e5d1a5c6cd1bcf README_JA.md
6a48d52e23aa594d79e5788265f17daa9eb3501beea1ae849a82a8bb2ff1ab12 README.md
18d3e7302db1bfe4f4417f6c520d2fcc616dceb23f703eb9e733147276b9c9a2 README.md
5d89c366b4f076fa73a2695557e0af749bf95df1dad0d55dbb2fe5b1c2519ad7 reference_audio.py
8f1744dc99f1a3a222b7d6e986f6dac1edcc0abbbe8b61a1d2606b1ca3569b3d reference_video.py
2995924fcb7bac27b548b01187d8276d2ab103c6f43416cf85400de3846c6297 reference.py
+88 -6
View File
@@ -317,15 +317,97 @@ The supplied templates use Video Helper Suite for this split. Any compatible IMA
## Prompt formats
Timeline:
### Timeline
For **Timeline** mode, each time-range header must be written on its own line. Write the scene description on the following line or lines.
Recommended format:
~~~text
[0-5s] First section: action, camera movement, and scene progression.
[5-10s] Continue from the exact final state of the previous section.
[10-15s] Continue naturally without resetting the scene.
[0-5s]
First section: action, camera movement, and scene progression.
[5-10s]
Continue from the exact final state of the previous section.
[10-15s]
Continue naturally without resetting the scene.
~~~
List prompts use --- separators. Fixed reuses one prompt for every chunk. Auto detects the available structure. If a timeline cannot be parsed, V3.4 reports a warning and uses an applicable fallback rather than rejecting an otherwise runnable workflow.
For 10-second chunks:
~~~text
[0-10s]
Describe the first scene.
[10-20s]
Describe the next scene.
[20-30s]
Continue the sequence.
[30-40s]
Describe the final section.
~~~
**Important:** do not place the prompt text on the same line as the time header.
Correct:
~~~text
[0-5s]
Describe the scene.
~~~
Not recommended:
~~~text
[0-5s] Describe the scene.
~~~
The second form is not recognized as a Timeline header. When **Prompt Format = Auto**, it may therefore be interpreted as a Fixed prompt instead, causing the complete text to be reused across chunks.
The time ranges should normally match the configured `chunk_seconds`.
- `chunk_seconds = 5` → `[0-5s]`, `[5-10s]`, `[10-15s]` ...
- `chunk_seconds = 10` → `[0-10s]`, `[10-20s]`, `[20-30s]` ...
- `chunk_seconds = 15` → `[0-15s]`, `[15-30s]`, `[30-45s]` ...
#### Global preamble and per-chunk prompting
Text placed before the first Timeline header is treated as a **global preamble** and is automatically included in every chunk prompt. This is a good place for information that should apply throughout the sequence, such as subject definitions, overall style, persistent environment details, or other global instructions.
However, the Timeline sections do not need to contain only the changing action or scene description. For complex scenes, multiple characters, or longer chunks such as 10–15 seconds, it can be useful to repeat important details about the subjects, appearance, location, camera, and current scene state inside each section. More self-contained chunk prompts may sometimes give H3 stronger continuity and reduce cut-like transitions or drift.
Continuum carries visual/audio context from the previous chunk, but that does not guarantee that H3 will preserve every semantic detail automatically. Instead of relying only on phrases such as "continue from the previous chunk," describe the intended next action and important continuity details explicitly.
For example:
~~~text
Subject: A woman in a red jacket.
Style: cinematic nighttime photography.
[0-15s]
A woman in a red jacket stands at the nightclub bar, speaking with two people beside her. Medium shot, slow handheld camera movement, crowded dance floor visible behind them.
[15-30s]
The same woman in the same red jacket continues the conversation at the same bar with the same two people. The camera remains close and slowly moves around the group as one person replies.
[30-45s]
The woman turns toward the dance floor while remaining beside the bar. Keep the same clothing, nightclub environment, nearby characters, and continuous camera style.
~~~
The global preamble is optional. If fully self-contained prompts inside every Timeline section give better continuity for a particular model or scene, that is also a valid approach.
### Other prompt formats
**List** prompts use `---` separators.
**Fixed** reuses one prompt for every chunk.
**Auto** detects Timeline, List, or Fixed formatting from the supplied text.
If Timeline syntax cannot be parsed, V3.4 reports a warning and falls back to an applicable prompt mode rather than rejecting an otherwise runnable workflow.
## Run Storage
@@ -456,4 +538,4 @@ No OOM was observed in the cited recent local V3.4 checks, including two-chunk 8
## License
See [LICENSE](LICENSE).
See [LICENSE](LICENSE).