MiniMax H3 Infinite Video: Make 2-Hour Movies in ComfyUI (0-Shot)
Full tutorial link > https://www.youtube.com/watch?v=1580ZDX-60Q
Turn one folder of scene prompts into a long, coherent AI video with MiniMax H3 - locally, 0-shot and without babysitting every clip. This ComfyUI walkthrough shows how to match references, queue scenes, generate clips and automatically merge everything into one movie.
The opening is the raw workflow result. Then we rebuild it from installation to playback: models, presets, VRAM modes, prompt creation, folder batching, reference syntax, draft settings, troubleshooting, selective regeneration and consistency. It can scale to very long projects, including the 2-hour movie shown here.
COMFYUI INSTALLERS + PRESETS:
SWARMUI MODEL DOWNLOADER:
DISCORD HELP & SUPPORT:
Zip File Content Screenshots
Some Other Screenshots From Workflows
Ultra advanced multiple reference upload and mention system
Optional Low VRAM optimization with just a toggle enable disable
Trim and upload references feature
Automatically download necessary files ultra fast and SHA256 verified
Sana Labs published 4x speed up implementation
Video Chapters
[00:00:00] 0-shot generated movie showcase
[00:00:58] MiniMax H3 local workflow reveal
[00:01:10] Entire movies, prompts and scenes automated
[00:01:22] Audio references and folder-batch strategy
[00:01:40] Local desktop vs cloud deployment
[00:01:50] Download ComfyUI and the model downloader
[00:02:01] Fresh install and recommended Python 3.12
[00:02:17] Verify installation and install bundle 100
[00:02:37] Troubleshooting and support
[00:02:47] Final checks before launch
[00:02:57] Launch ComfyUI with run_gpu.bat
[00:03:11] 6 GB GPU support and speed tradeoffs
[00:03:22] Logs, errors and starting the preset
[00:03:33] Find the MiniMax H3 presets
[00:03:44] References-to-video 4x workflow
[00:03:54] Compilation, 20 steps and speed LoRA
[00:04:12] Change logs and Windows requirements
[00:04:24] Frequent workflow and custom-node updates
[00:04:36] Missing models and downloader setup
[00:04:48] Launch downloader and share model paths
[00:05:03] Core vs low-VRAM MiniMax H3 bundles
[00:05:16] INT4 vs recommended INT8 quality
[00:05:29] Options for 6-12 GB GPUs
[00:05:43] Select the correct models path
[00:05:55] Folder structure and model download
[00:06:07] Fix path warnings and select both VAEs
[00:06:23] Reference manager and default interface
[00:06:35] The enhanced prompt helper file
[00:06:47] Improve prompts with any major LLM
[00:06:59] ChatGPT setup with five voice attachments
[00:07:17] Example prompts and downloadable ZIP
[00:07:28] Prompt construction and easy referencing
[00:07:42] Upload the audio references
[00:07:52] Reference syntax and voice samples
[00:08:08] Listen to the supplied voice references
[00:08:18] Audio and image reference options
[00:08:28] Reorder image references by drag and drop
[00:08:38] Batch folder mode needs no manual prompt
[00:08:50] Set the folder path and draft resolution
[00:09:01] Aspect ratios and divisible dimensions
[00:09:15] Recommended 1344x768 for 16:9
[00:09:25] 30-second limit and 15-second sweet spot
[00:09:35] Best parameters: ready to run
[00:09:47] Run a quick 0.4 MP draft
[00:09:57] Automatic merging and queue monitoring
[00:10:08] Folder prompts and reference matching
[00:10:20] Load many references; use only the matches
[00:10:30] Per-generation reference limits
[00:10:40] Batch power for full videos and animations
[00:10:50] Iterate prompts with your favorite GPT
[00:11:00] Draft low resolution, review, then refine
[00:11:13] The key file for better prompts
[00:11:23] Single-clip mode and included presets
[00:11:33] Image, text and references-to-video presets
[00:11:46] Lightricks speed-up LoRA implementation
[00:11:59] LoRA version notes and future updates
[00:12:17] SwarmUI support and future advanced tutorial
[00:12:28] Read the docs and enable low-VRAM mode
[00:12:38] Save 15-40%+ VRAM
[00:12:54] Built-in node help and first output
[00:13:08] Regenerate only a weak section
[00:13:26] Play the generated result
[00:13:36] Current audio-reference limitation
[00:13:46] Improve voice and character consistency
[00:13:56] Match image IDs to characters
[00:14:06] Included reference and consistency guide
[00:14:16] 0-shot proof: ChatGPT authored all prompts, the movie script and rough draft
[00:14:35] Automated setup and broad GPU support
[00:14:45] Final requirements reminder and outro
MiniMax H3 supports 30-second segments; 15 seconds is the sweet spot. This demo uses 10-second scenes for easy review and regeneration.
For 16:9, use 1344x768. Draft around 0.4 MP, review the sequence, improve prompts, then render final quality.
Load up to 99 references; each prompt uses only matched IDs. Per generation: up to 3 audio, 3 video and 9 image references. Extra audio/video refs are discarded.
Use INT8 for speed and near-BF16 quality; INT4 is for lower VRAM. It runs on 6 GB GPUs, but more slowly. Low-VRAM mode saves about 15%, reaching 40%+ in some cases.
Audio references are less reliable than image references. If voice quality drops, try no audio reference. Map image IDs clearly for consistency.
Covers local Windows, Massed Compute, RunPod and SimplePod. Use a fresh install with Python 3.12 where recommended. Read requirements and changelogs; nodes and workflows update often.
If one scene fails, regenerate only that section. Join the Discord above for setup help.
#MiniMaxH3 #ComfyUI #AIVideo #LocalAI #GenerativeAI