github BMB12d3/minimax-h3-prompt-composer V5.14.0

latest releases: V5.43.4, V5.40.2, V5.37.14...
one month ago

H3 Prompt Composer V5.14.0
Audit-Hardening / Correctness Release

OVERVIEW

V5.14.0 preserves the V5.13 multi-Subject scale/placement workflow while implementing the relevant findings from an independent full-source and Chromium runtime audit. The update focuses on correctness, project self-healing, generated-language quality, maintainability, and documentation alignment.

FIXES IMPLEMENTED

  1. Built-in camera composition enum repair
  • Fixed two built-in examples that stored the invalid camera composition value "thirds".
  • Full reference scene Shot 1 now uses the valid "left_third" enum.
  • Simple text-to-video uses the valid "centered" enum.
  • This prevents silent loss of composition language and UI/state desynchronization.
  1. Multi-Subject scale/placement grammar
  • Removed the broken possessive generated after a list of Subject labels.
  • Shot-level placement cues now use grammar such as:
    "...the approximate apparent size, left-to-right order, spacing, depth, floor contact, and blocking of <Subject ...> in this Shot..."
  • Fallback wording consistently uses "the assigned subjects".
  1. Camera enum self-healing
  • Project migration now validates stored Camera Builder enums against the current dropdown option sets.
  • Invalid legacy or hand-edited values are safely cleared rather than surviving as invisible state.
  • Auto-tidy applies the same sanitation to the active project.
  • Covered fields include start/end framing, focus area, viewpoint, angle, composition, stability, lens, depth of field, focus behavior, movement, range, speed, and timing.
  1. Multi-Subject scale mapping cleanup
  • Migration deduplicates stale scaleSubjectUids while preserving first-seen left-to-right order.
  • Auto-tidy removes invalid/environment Subject IDs and deduplicates valid mappings while preserving order.
  1. Custom retention-note validation
  • Prompt Check now flags speaker IDs such as (S1) or (S1,S2) when manually entered into a custom retention note.
  • Speaker IDs belong to dialogue/action output, not retention_analysis.
  1. detailed_description word-count accuracy
  • The advisory now counts the actual generated detailed_description section.
  • Style, edit preambles, scale guidance, continuity preambles, and Shot text are all included.
  1. Dead-code cleanup
    Removed the following unreferenced helpers identified by the audit:
  • appearanceStateAfterShot
  • embeddedAudioRefs
  • resolvedVoiceRef
  • allSpeakingSubjects
  • cameraViewpointLocation
  • wiredLabels
  • parenFor
  • loadDemo

This also removes the divergent dead scale-parenthetical implementation that could have become a future maintenance trap.

  1. Minor source cleanup
  • Removed the unused target parameter from angleSentence().
  • Removed a duplicated phrase-bank source comment.
  1. H3 duration floor updated
  • Target duration minimum changed from 5 seconds to 4 seconds.
  • This matches the current official MiniMax H3 output-duration specification of 4-15 seconds.
  • Existing ComfyUI effective-duration snapping remains based on the current 17k+5 frame grid at 24 fps.

DOCUMENTATION UPDATES

  • README updated to V5.14.0.
  • Full User Guide updated to V5.14.0.
  • Illustrated User Guide updated to V5.14.0.
  • Camera and overview screenshots regenerated from the corrected V5.14.0 interface.
  • Multi-Subject example wording now exactly matches the generated prompt.
  • Documentation explicitly explains that stable Subject label numbers do not need to increase from left to right.
  • Built-in example name standardized to "Simple text-to-video shot."

AUDIT OBSERVATIONS REVIEWED BUT INTENTIONALLY NOT CHANGED

  • Timed Action Beats continue to use the requested creative duration, while cut validation/effective duration can reflect the snapped model grid. This is intentional: beat entry remains human-friendly rather than exposing fractional grid tails.
  • Natural-English Camera Builder phrasing is retained instead of forcing canonical token ordering; the current prompt guides explicitly favor natural camera language.
  • Load-bearing edit/scale/continuity preambles remain before Shot 1 where needed even when that produces more than a single short style sentence.
  • The 17k+5 / 24 fps duration-grid behavior is retained after verification against current ComfyUI H3 documentation.
  • The current H3 reference limits were rechecked against MiniMax's current model card.

QA / REGRESSION TESTS

  • JavaScript syntax validation: PASS
  • Headless Chromium boot: PASS
  • All six built-in workflows generate with zero blocking Prompt Check errors: PASS
  • Every built-in camera enum resolves to a current enum option: PASS
  • Reference-scene valid left-third composition: PASS
  • T2VA valid centered composition: PASS
  • Four-Subject left-to-right scale/placement mapping: PASS
  • Correct multi-Subject placement grammar: PASS
  • Scale-list migration deduplication: PASS
  • Auto-tidy scale-list cleanup: PASS
  • Corrupt/legacy camera enum migration self-healing: PASS
  • Live Auto-tidy camera enum self-healing: PASS
  • Custom retention-note speaker-ID lint: PASS
  • Full detailed_description word-count extraction: PASS
  • V5.13.0 -> V5.14.0 migration: PASS
  • Runtime page errors in targeted regression suite: 0
  • External/network dependency static scan: PASS
  • Duplicate static element ID scan: PASS

SECURITY / LOCAL-ONLY CHECK

V5.14.0 remains a standalone local HTML application. Static checks found:

  • external script src: 0
  • external stylesheet href: 0
  • fetch(): 0
  • XMLHttpRequest: 0
  • WebSocket: 0
  • EventSource: 0
  • sendBeacon: 0
  • eval(): 0
  • new Function(): 0
  • iframes: 0
  • HTTP/HTTPS URLs: 0

Don't miss a new minimax-h3-prompt-composer release

NewReleases is sending notifications on new releases.