v0.3.5-alpha.5 MTV Singing H3 升级audio Separator

https://www.youtube.com/watch?v=V-Dh2e5Xgvo

这一次加入audio separator, 会自动把声音分成vocal only & music only, 然后自动添加如A1 & A2

v0.3.2 MiniMax H3 Director Cut Studio with Qwen TTS

“https://github.com/karuvanan/MiniMax-H3-Director-Cut-Studio/issues/8”

MiniMax H3 Director Cut Studio v0.3.1 Long-Form Pipeline

Full Changelog: https://github.com/karuvanan/MiniMax-H3-Director-Cut-Studio/blob/main/CHANGELOG.md
code: minimax h3 reference r2v v3.1 alpha 2.zip

v0.3.1 Budget-Aware Long-Form Pipeline

This release introduces a budget-aware long-form production pipeline designed for longer AI video projects, especially workflows that need to expand short videos into complete 120-second episodes while keeping generation cost, disk usage, and GPU time under control.

Highlights

• Added a project-based long-form workflow for 120-second timeline production.
• Introduced Shot-level planning, preview, approval, reuse, and final rendering.
• Added support for low-cost Storyboard Preview before running full H3 generation.
• Added 0.2MP Shot Preview for faster visual checking.
• Added 1.0MP final rendering for accepted Shots.
• Added Shot-level Accept / Reject logic.
• Added Seed and Prompt locking for approved Shots.
• Added support for reusing approved Shots instead of regenerating the full video.
• Added a structure for expanding an existing 45-second video into a 120-second production.
• Improved handling of generated assets, takes, cache, audio, and final renders.
• Added foundation for future Chapter / Scene / Shot organization for longer projects.

Production Workflow

flowchart LR

Why This Release Matters

Instead of treating a 120-second video as one huge generation job, this version treats the full timeline as a managed production made of independent Shots, Takes, audio, references, and cache.

This makes the workflow more practical for limited hardware. The goal is to avoid unnecessary regeneration, reduce wasted GPU time, prevent project folders from becoming messy, and make longer AI video production easier to control.

Suggested Production Strategy

Use different quality levels depending on the importance of each Shot:

Shot Type	Use Case	Suggested Handling
Hero	Opening, twist, action, ending	Test at 0.2MP, then upscale accepted Shot to 1.0MP with the same Seed
Standard	Dialogue, investigation, reaction Shots	Stable framing, moderate movement, avoid frequent regeneration
Economy	Phone screen, doorplate, empty room, evidence close-up	Image animation, simple camera push, short hold, or compositing
Reuse	Previously approved Shots	Do not regenerate, use directly
Locked	Final approved Shots	Prevent recalculation unless manually unlocked

Recommended 120-Second Structure

• 16 to 20 Shots total.
• 3 to 5 Hero Shots.
• 7 to 10 Standard Shots.
• 3 to 5 Economy or static-composite Shots.
• Final Hold does not need complex regeneration.
• Transitions should preferably be handled during FFmpeg assembly instead of regenerating video just for fades or cuts.

Budget-Aware Generation Rules

• Each Shot should receive at most one initial 0.2MP preview.
• A new Seed should only be used after the user rejects the current Shot.
• Accepted Shots should reuse the same Seed for 1.0MP generation.
• Automatic retry should be limited.
• Minor non-critical defects should not automatically trigger regeneration.
• Studio should show estimated GPU time, disk usage, and Shot count.
• If the project exceeds the generation budget, user confirmation should be required.

Project Workspace

This release moves toward a cleaner “one project, one workspace” structure:

example/
└─ project_name/
   ├─ project/
   │  ├─ project_name.h3director.json
   │  └─ project_manifest.json
   ├─ design/
   │  ├─ requirement.txt
   │  ├─ design_plan.json
   │  └─ revisions/
   ├─ media/
   │  ├─ imported/
   │  ├─ generated_references/
   │  └─ audio/
   ├─ shots/
   │  ├─ S001/
   │  │  ├─ approved.mp4
   │  │  └─ takes/
   │  ├─ S002/
   │  └─ S003/
   ├─ renders/
   │  ├─ preview/
   │  ├─ scene_masters/
   │  └─ final/
   ├─ proxies/
   ├─ cache/
   └─ logs/

Reuse and Expansion

A major goal of v0.3.1 is to support low-cost expansion from an existing 45-second production into a full 120-second short drama.

Recommended behavior:

• Mark the existing 45-second section as APPROVED / LOCKED / REUSE.
• Expand the timeline to 120 seconds.
• Only generate the new 45s–120s section.
• Keep all approved Shots unchanged.
• Extract the final frames of the 45-second section as optional continuity reference.
• Stitch the old and new sections together with unified audio, color, and encoding.

Preview Modes

This release encourages three preview levels:

Storyboard Preview

• Does not run H3.
• Uses reference images, generated stills, text layers, TTS, and simple camera movement.
• Allows fast review of story rhythm, dialogue, Shot length, and asset placement.

Motion Preview

• Runs only selected Shots.
• Uses 0.2MP preview quality.
• Does not upscale.
• Best for testing motion, framing, and scene continuity.

Final Render

• Processes only accepted Shots.
• Reuses locked Preview Seed.
• Uses 1.0MP output for accepted Shots.
• Prioritizes high-quality rendering for Hero Shots.

Editable Shot Data

Each Shot should preserve enough metadata to support selective editing:

shot_id
timeline_range
prompt
seed
references
incoming_state
outgoing_state
render_profile
approved_take
prompt_hash
media_hash
audio_hash
output_path

This allows changes to remain local:

• Dialogue changes only update related TTS and affected Shots.
• Prompt changes only regenerate that Shot.
• Transition changes should only rebuild the assembly when possible.
• Music changes only require remixing.
• Locked reuse ranges should not be affected by later timeline expansion.

Future Long-Form Foundation

v0.3.1 also prepares the foundation for future 90-minute projects through a Chapter / Scene / Shot structure:

Feature Project
├─ Chapter
│  ├─ Scene
│  │  ├─ Shot
│  │  ├─ Shot
│  │  └─ Shot
│  └─ Scene Master
└─ Final Assembly

The full 90-minute editor is not the goal of this release, but the file structure, Shot IDs, manifest, cache, and locking system should remain compatible with that future direction.

Suggested Development Milestones

v0.3.1-alpha.1

• Single Project Workspace.
• Shot Take and Approved status.
• Storyboard / 0.2MP / 1.0MP quality modes.
• Disk and GPU time estimation.

v0.3.1-alpha.2

• Protected Reuse Range.
• Import existing 45-second project or MP4.
• Generate only the new 45s–120s section.
• Mixed Storyboard and generated Shot playback.

v0.3.1-alpha.3

• Shot-level Render Queue.
• Seed locking.
• Reject / Accept flow.
• Local rerender and cache reuse.
• Rejected Take cleanup.

v0.3.1-alpha.4

• Project Manifest.
• Basic Chapter / Scene structure.
• Lazy loading foundation for 90-minute projects.
• Resume interrupted generation.
• Disk quota handling.

Release Goal

The main goal of v0.3.1 is to make “45 seconds to 120 seconds” production practical, affordable, editable, and reusable, while preparing the project architecture for much longer AI video workflows in the future.

Title: MiniMax H3 Director Cut Studio v0.1.0-alpha.1 — First Public Preview

Full Changelog: https://github.com/karuvanan/MiniMax-H3-Director-Cut-Studio/commits/v0.1.0-alpha.1

MiniMax H3 Director Cut Studio v0.1.0-alpha.1

First public preview of MiniMax H3 Director Cut Studio.

Highlights

• Premiere-inspired multi-track Timeline
• MiniMax H3 Ref2VA prompt generation
• 9 image, 3 audio and 3 video reference slots
• AI-assisted Director Design through LM Studio or Online GPT
• Z-Image reference generation through ComfyUI
• BLIP, Whisper, VAD and beat analysis
• Optional AI Semantic Enrichment
• Automatic Timeline Prompt Reconcile
• Shot-aware Smart Long Render beyond 15 seconds
• Selective Shot re-rendering and reusable segment cache
• Program Monitor source/output comparison
• Project save, Undo/Redo and automatic output archiving

Runtime Download

Complete Windows runtime:

https://drive.google.com/file/d/1mC_GpmCuYw7zaQPfkaqtQVXTSt6DlRsM/view?usp=drive_link

Download and extract the runtime into the project folder. The final structure must contain:

ai_libraries_common/python_env/
ai_libraries_common/engine_ffmpeg/
ai_libraries_common/models/

Verification

• Project format: version 14
• Python: 3.11.15
• PySide6: 6.11.2
• Torch: 2.12.1+cu126
• CUDA runtime: 12.6
• Transformers: 5.12.1
• Tests: 156 passed

Notes

ComfyUI checkpoints are not included. Prepare all required models listed in README before running the Studio.

This is a pre-release intended for testing and feedback.