A video VAE (e.g. the new Sol-H3-VAE) wired into audio_vae failed deep in
Core's decode with 'tuple index out of range'. Classify both slots by Core's
MiniMaxH3VideoVAE/MiniMaxH3AudioVAE classes before decoding, in both output
modes, so the error names the wiring and new H3 VAE variants keep passing.
As in PromptForge: picking a number on Shots shows one box per shot under
it (none on Auto or in a continuation). The node folds filled boxes into
the idea as "Shot N:" lines before anything reads it, so easy mode's
Character names in a shot box are handled like the idea's. Empty boxes are
skipped, boxes past the count dropped, and the boxes alone are enough to
generate. Typed text is kept per node and saved with each draft; drafts
saved before this still match when the boxes are empty.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The h3-forge-easy work, rebuilt on the reworked reference panel instead
of merged over it. All of 0.4.73 stays: per-picture instructions and
keep/ignore, Pose and Custom, saved references, continuity references,
the output canvas, reference validation and media citation checks.
Picture labels: in REF2VA each picture gets one label, picked in two
steps - Character, Several characters (who stands where, left to
right), Place, Style, First frame, Last frame, Pose, Custom. Pictures
with the same Character number are one subject. The label is saved on
the timeline item (forge_label, next to forge_role) and also sets the
role and group, so continuity and the Director read what they did
before. A picture never labelled gets one from its role and group: a
group becomes one Character number, an ungrouped subject the next free
one.
For a new draft, labelled REF2VA runs the "REF2VA easy" prompt: code
numbers the subjects, writes Subject definitions and Retention analysis,
and no picture goes to the writer (H3 sees them). The idea's
"Character 1", "the place" and "picture 3" become tags. Instructions,
keep and ignore go on each cast line; Pose and Custom pictures get a
line of their own; a saved reference keeps its usual reference line.
Also carried over: the Shots control (Auto, 1-5) with the count checked
afterwards, cut times at or past the end moved into the clip, music only
when the idea asks for it, a draft that ran away refused, presence
penalty 0, and the llama-swap / Ollama unload fixes.
Bundle: 0.4.73's detail and creativity rules kept as they are; the music
wording re-applied to every mode; "REF2VA easy", music_words and
shot_counts added. The easy prompt gets the same canvas and shot-policy
block and Summary wording as REF2VA.
test_h3_forge_groups_ui.mjs failed on main (forgeReferences is imported
and the sandbox stripped it); it now imports the state module directly.
Measured on llama-swap, 10 s clips, a 1280x720 canvas, two characters
(one in two pictures) and a place, then a pose picture: a 9B 9/9 in
13-18 s, a 4B 8/9 in 8-12 s (one draft without a [Shot 1] marker,
flagged by the shot-count warning). Every picked shot count matched.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Adds .github/scripts/sync_issue_form.py (AST-parses __init__.py, no
runtime deps) and a workflow that regenerates the dropdown and
auto-commits to main when __init__.py or the form changes.
Darksidewalker