English · العربية · Español · Français · 日本語 · 한국어 · Tiếng Việt · 中文 (简体) · 中文（繁體） · Deutsch · Русский
Maximum-quality, local MiniMax H3 video generation for a dual-RTX 4090 workstation—native picture, sound, references, and careful resource ownership.
LocalVideoGen is a reproducible operating layer around a pinned external ComfyUI installation and the official aligned MiniMax H3 model package. It provides a loopback-only H3 Studio webapp, checksum-gated model acquisition, T2V/I2V/R2V workflow presets, native video-and-audio generation, persistent job history, and conservative lifecycle controls tuned for two 24 GiB RTX 4090 GPUs with 128 GiB host RAM.
| Donate | PayPal | Stripe |
|---|---|---|
 |
 |
 |
The light theme keeps reference setup, quality controls, and render status readable in one local workspace. The default Single Clip setup is intentionally beginner-friendly: choose Start with words, describe the scene and sound, then press Create H3 video. H3 Studio starts with a 2-second, 864×480 fast preview; final-quality and repeatability controls remain available when needed.
Choosing Use references in the Single Clip UI applies the separate Long reference · 24 GiB safe preset: 14 seconds, 25-step quality_int8_offload, reference matching, and a 1248×704 or 704×1248 canvas. The raw POST /api/renders compatibility default remains five seconds when an R2V caller omits duration; requests with a visual video also default to INT8/offload and 1248×704, while image/audio-only R2V keeps the earlier BF16 and 1344×768 defaults. Every visual video needs server-trusted normalized dimensions at every duration and is fitted to H3-native 32-pixel axes within 1024 px and 589,824 pixels (576×1024 for 9:16). Match-mode admission uses aligned_frames × (output_pixels + sum(video_reference_pixels)) + max(0, matched_still_count − 1) × output_pixels: each still after the calibrated first costs one output canvas once, not once per frame. The API rejects a base video-frame load above 510,000,000 at any duration; from 243 aligned frames (about 9.5 requested seconds), visual-video R2V also requires quality_int8_offload and match. The validated 14-second portrait case is 704×1248 + one 576×1024 reference + one matched still, 345 frames, or 506,603,520 proxy pixels. Additional stills or visual/continuity videos may require a shorter clip or smaller canvas; use audio-only guidance when pictures are unnecessary. The 510M/24 GiB claim applies to this measured match-mode route—explicit shorter BF16/max mixes remain available but are outside that guarantee.
Every creation appears in Recent sessions with a plain-language status, prompt title, size, and duration. Opening a session restores its player, and Reuse settings copies its prompt and settings into a new attempt. Reference files must be selected again because browser upload handles are deliberately not persisted as reusable file access.
Switch between Single Clip and Series without leaving H3 Studio. Series mode offers LALACHAN Series, a quality-first World Travel preset, and a neutral My Movie preset for any cast or visual style.
- Upload shared cast, world, voice, and motion references once, then arrange 2–12 editable shot cards with individual prompts, durations, and seeds.
- Pre-clean every visual video reference before upload and run the fail-closed subtitle-free reference guard on the exact 2–15-second segment. A prompt that says “no subtitles” cannot reliably remove caption pixels already present in the reference; use audio-only input when only voice or music continuity is needed.
- World Travel keeps seven canonical logical character/prop pictures at P1–P7, gives every shot its own destination plate at P8, and reserves P9 for the preceding accepted shot's exact final frame. Later shots default to excluding the opening-only Words card, LightMind glasses, and Patchwork notebook, with visible per-shot overrides and an effective H3 map before start. H3 Studio compacts physical slots and safely remaps authored tags while always retaining Zhuangzi Robot, Rara Xia, Aya Chan, and Sasa Kun. Earlier episodes may guide identity or voice only; they cannot steer the new country, plot, blocking, palette, or composition.
- A single admission gate keeps every H3 render strictly sequential. When continuity is enabled, each validated non-final shot supplies its exact final frame and configured 2–4-second tail to the next shot. World Travel defaults to 2 seconds to reduce replay; LALACHAN Series and My Movie retain the 3-second default.
- Pause after the current shot, resume after a restart, retry one shot or everything that follows, and retry post-processing or final stitching without spending GPU time on an already valid MP4.
- Every render attempt remains preserved. The final movie is assembled by lossless stream copy only after checks for the expected frame count, reported average 24 fps, stereo-audio alignment within AAC tolerance, full decode, and SHA-256; a validation manifest is retained beside it.
- Other local projects and Codex sessions can use the dependency-free Series client. It uploads bounded references, verifies server capabilities, writes an atomic durable-ID receipt before generation, survives paused review and polling interruptions, and verifies artifact size plus SHA-256 before installing a download.
The workflow is inspired by the clarity of Xiaoyunque's storyboard experience, but all generation and project state stay on this workstation over loopback. It does not call Xiaoyunque or any paid cloud generation service.
See the Series workflow guide for continuity and recovery behavior, the cross-project Series API guide for the stdlib CLI/client and full HTTP contract, and the smooth long-video options review for the trusted native H3 baseline, experimental continuity projects, optional interpolation, and quality gates.
Review one paused or completed accepted shot without overwriting retained evidence:
./scripts/review_series_shot.sh \
/absolute/path/to/series-receipt.json \
0 \
runtime/private/quality-reviewThe zero-based helper downloads the verified accepted MP4 and, for a non-final shot, its current continuity tail and exact final frame. It fully decodes and probes the media, records hashes, creates a full contact sheet and outgoing-boundary strip, and prints—but never runs—an optional Whisper large-v2 command.
- Highest-fidelity preset for a short visual-video reference or image-only R2V: pruned BF16 Ref2VA/FL2VA DiT, the selected aligned or Heretic NVFP4 Qwen3-VL conditioner, FP16 video VAE, FP32 audio VAE, and 25 full-model steps. In Series, image-only means Continuity handoff: Off; otherwise successor shots gain a visual continuity tail. Long visual-video references default to the measured INT8/offload safety envelope above.
- Shared-workstation stage placement: GPU 0 runs the DiT, denoiser, Qwen/reference conditioning, and final video/audio decode by default, leaving GPU 1 available for LocalLLM or another protected workload. For BF16 stage-placement profiles, this keeps the same model, BF16 precision, sampler, resolution, and 25-step quality while using GPU-0/CPU offload. On an exclusive workstation, deliberately set
H3_AUX_DEVICE=gpu:1to move Qwen plus reference-conditioning VAE work to GPU 1; final decode still remains on GPU 0. No PCIe peer-to-peer path is assumed. - Local T2V, I2V, and multi-reference R2V, including a max-identity reference preset and native synchronized audio.
- Quality, single-GPU fallback, and low-resolution INT8 Turbo preview profiles.
- Local H3 output up to a 768-pixel short edge at 24 fps. MiniMax's separate 2K regeneration stage is API-only and is not presented as a local feature.
flowchart LR
B[Browser] -->|loopback :8190| S[H3 Studio]
S --> V[Upload and graph validation]
V --> J[(Private job registry)]
V -->|loopback :8188| C[Pinned ComfyUI]
C --> G0[GPU 0: all H3 stages by default]
G1[GPU 1: reserved for LocalLLM by default]
C -. optional H3_AUX_DEVICE=gpu:1 .-> G1
C <--> R[Host RAM: DynamicVRAM + async offload]
M[SHA-256 verified model bundle] --> C
The webapp never starts a second ComfyUI process. Services bind to 127.0.0.1; uploads are normalized into bounded media, browser-visible identifiers are opaque, and output paths are allowlisted.
| Path | Purpose |
|---|---|
webapp/ |
Responsive H3 Studio, local API, upload normalization, job registry, and tests |
workflows/dual_gpu/ |
BF16/INT8 T2V, I2V, and R2V stage-placement workflows |
workflows/quality/ |
Single-GPU 25-step quality workflows |
workflows/preview/ |
Small INT8 Turbo preview workflows |
scripts/ |
Download, verification, lifecycle, resource, smoke-test, and workflow tools |
config/model-manifest.sha256 |
Exact nine-file model allowlist and SHA-256 checksums |
config/runtime-versions.txt |
Pinned external runtime versions and commits |
The large ComfyUI/, workflow_templates/, model weights, outputs, uploads, databases, and runtime receipts are local installations or private/generated state; they are deliberately not committed.
This repository is the public orchestration layer, not a universal installer. Before using these commands, install external ComfyUI/ and workflow_templates/ working copies at the exact commits in config/runtime-versions.txt, create .venv with the recorded Python/PyTorch/CUDA stack, and install the upstream ComfyUI requirements. Those upstream trees are intentionally not vendored.
The complete official aligned model queue is 147,804,799,439 bytes (137.65 GiB). Downloads are resumable, but plan for the model data plus at least 32 GiB of free-space headroom.
git clone https://github.com/lachlanchen/LocalVideoGen.git
cd LocalVideoGen
# After installing the pinned external runtime:
./scripts/download_models.sh
./scripts/verify_models.sh
./scripts/status.sh
./scripts/start_comfyui.sh
./scripts/start_webapp.sh
# Optional exclusive-workstation speed mode: use both GPUs for H3 stages.
H3_AUX_DEVICE=gpu:1 ./scripts/start_webapp.shOpen http://127.0.0.1:8190. ComfyUI remains private at http://127.0.0.1:8188.
H3 Studio is designed for beginners as well as repeat use:
- Always remember saves optional requirements such as “no subtitles” and adds them to every new single clip and series shot.
- Clip length remembers the last selected duration on this workstation and restores it for a new session or page reload.
- Recent sessions lets you reopen or reuse a render. A finished standalone session has a Delete button with an irreversible confirmation; it removes both the session row and its allowlisted generated video. Active renders and series-owned shots are protected from deletion there.
- If the disposable session database is removed, the running studio initializes a fresh empty history automatically. Remembered requirements and clip length are kept in a separate private settings file.
All of this usage data is local-only. Prompts, settings, sessions, uploads, generated videos, browser profiles, logs, and screenshots are excluded from Git and must never be committed or pushed. Only the reusable tool source, generic tests, and documentation belong in the repository.
The local CLIPLoader selection is reversible and applies to H3 Studio plus all generated editable workflows. The selector verifies the chosen weights before changing anything:
# Use the optional local Heretic NVFP4 encoder.
./scripts/set_text_encoder.py heretic
# Restore the official aligned Comfy-Org encoder.
./scripts/set_text_encoder.py aligned
# Show the current selection without changing it.
./scripts/set_text_encoder.py statusIf H3 Studio is already running, restart only the webapp after switching. ComfyUI discovers
both files from ComfyUI/models/text_encoders/; neither model is renamed or deleted.
Stop only this project's verified processes when no render is active:
./scripts/stop_webapp.sh
./scripts/stop_comfyui.sh- Startup requires at least 48 GiB available RAM, 20,000 MiB free on every requested GPU, and swap use at or below 75%.
- Partial model downloads and changed model size/mtime fingerprints block startup; all nine files are structurally checked and SHA-256 verified.
- PID, boot identity, command line, listener ownership, service marker, and queue state are checked before lifecycle actions.
- GPU cleanup is dry-run by default. Explicit cleanup uses exact pidfds and protects process trees rooted in
LocalLLMandAgenticApp; foreign processes are never stopped automatically. - Swap is emergency headroom, not a substitute for the RAM launch gate. Only one render engine and one lightweight studio are expected for this project.
Run static workflow generation/validation and the webapp tests without submitting a render:
./scripts/prepare_workflows.py
./scripts/validate_workflows.py
.venv/bin/python -m unittest discover -s webapp/tests -v
./scripts/verify_models.shWith verified models and an idle GPU 0, run the tiny native audio-video smoke graph:
H3_CUDA_DEVICES=0 ./scripts/start_comfyui.sh
./scripts/smoke_test.sh
H3_SMOKE_MODEL=minimax_h3_fl2va_pruned_bf16.safetensors ./scripts/smoke_test.sh
./scripts/stop_comfyui.shThe five-frame smoke output checks Qwen conditioning, joint video/audio sampling, both VAEs, MP4 muxing, and stream decodability; it is not a visual-quality benchmark.
The original LocalVideoGen code and documentation are released under the MIT License. That license does not relicense ComfyUI, workflow templates, MiniMax H3 weights, Qwen components, FFmpeg, or any other upstream dependency or generated asset. Each retains its own terms.
No jailbroken or abliterated conditioner is included: the manifest accepts only the aligned Comfy-Org qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors file at its recorded checksum. The MiniMax H3 Community License excludes the EU, UK, Republic of Korea, and USA from its applicable territory and adds redistribution obligations; review the upstream model card, license, applicable law, and every dependency license before downloading or using the weights. This project does not provide legal advice.
If you use LocalVideoGen in research, cite the repository. GitHub reads CITATION.cff and shows a Cite this repository panel on the repo page.
@software{chen_localvideogen_2026,
author = {Chen, Lachlan},
title = {LocalVideoGen: Resource-Aware Local MiniMax H3 Video Generation},
year = {2026},
version = {0.1.0},
url = {https://github.com/lachlanchen/LocalVideoGen}
}Version 0.1.0 is a workstation-focused research release, validated on Linux with two RTX 4090 GPUs and 128 GiB RAM. It favors reproducibility, model integrity, visual quality, and safe coexistence with other long-running projects over broad hardware support or one-click setup. Results remain generative and should be reviewed before publication.
Project: github.com/lachlanchen/LocalVideoGen · Homepage: lazying.art