Open Source
Open Source

ai-manju: A Local Production Pipeline for AI Short Dramas - ComfyUI + MiniMax-H3 from Chinese Draft to Video with Audio

catiseyeqaq/ai-manju-shengcheng-xitong: a local production pipeline for realistic AI films/short dramas. ComfyUI 0.31.0 orchestration + MiniMax-H3 joint audio-video generation (BF16 weights ~129GB, 8x PPU-ZW810E) + Qwen3.6-35B-A3B Chinese prompt polishing. 33 template workflows, T2V/I2V/R2V, majicFlus + PuLID cross-shot face lock + H3 first/last-frame continuity + 48k audio alignment. Full-chain analysis, hardware bar, 5 FAQs. Early personal project (4 stars); verify via Releases.

Published August 15, 20268 min read
<!-- ai-manju-resource | open-source | ai-manju: A Local Production Pipeline for AI Short Dramas - ComfyUI + MiniMax-H3 from Chinese Draft to Video with Audio -->

The AI comic-drama (manju) niche is hot, but most players stall at the same gate: single images are easy, coherent footage is hard. The GitHub repo catiseyeqaq/ai-manju-shengcheng-xitong (AI Film & Short Drama Generation System) offers a complete local answer: a realistic AI film/short-drama production pipeline built on ComfyUI + the MiniMax-H3 joint audio-video model + Qwen3.6 prompt polishing, carrying a Chinese creative draft all the way to video segments with stereo audio - and the author dogfoods it: finished clips are downloadable in Releases. This analysis covers the architecture, capabilities, hardware bar, and pitfalls.

Boundaries: repo facts from the GitHub API and README (as of 2026-08-15; 4 stars, Python, pushed the same day); output quality per the author's Releases - this is not a hands-on deployment. For methodology see our Xiaohongshu viral content SOP; for the tooling landscape see the AI video tools comparison.

1. What the Pipeline Solves: from "Can Generate an Image" to "Can Ship an Episode"

Every single capability exists in consumer AI tools; short-drama production is a pipeline problem: Chinese idea -> English cinematic prompt -> node-graph inference -> cross-shot character consistency -> first/last-frame continuity -> audio-visual aligned stitching. ai-manju packages the whole chain as reusable workflow assets:

StageSolutionNotes
Prompt polishingQwen3.6-35B-A3B (SGLang)Chinese draft -> MiniMax-H3-friendly English prompts with camera/light/SFX cues; integrates MiniMax's official H3 prompt-writing skill
Video generationMiniMax-H3 (T2V/I2V/R2V)Text-to-video, image-to-video, reference-to-video; natively joint-generates stereo audio
Character consistencymajicFlus portraits + PuLID face lockSame face across shots
Quality chainFaceDetailer + USDU repairFace and frame detail rescue
ContinuityH3 first/last-frame stitchingAction and scene transitions between shots
Final cut48k audio-visual aligned stitchingComplete segments with audio

The repo ships 33 ready-to-import template workflows: T2V/I2V/R2V, cinematic keyframe/storyboard/coherent-film pipelines, the "Masters" six-pack, the "Wanling Scroll" guqin-calligraphy-painting series, plus QwenImage/FLUX2 image generation and interior-render pipelines. Stack in one sentence: ComfyUI 0.31.0 for orchestration and UI (port 8188), SGLang for the polishing service (port 8030), one-command start/stop for both.

2. The Hardware Bar and Deployment Shape

This is the section to read first: loading MiniMax-H3 BF16 weights locally takes about 129GB. The author's deployment is an 8-card PPU-ZW810E cluster with ComfyUI acceleration parameters and attention-backend switching scripts. Practical implications:

  • Without 8-card-class hardware, don't expect to run full BF16 H3 locally; consider quantization or cloud inference - the workflow logic remains reusable
  • The repo positions itself as deployment configs + template workflows + ops scripts + docs; it does not include the 100GB-scale weights or the ComfyUI upstream source (fetch per docs)
  • Finished clips (1080p with audio) live in GitHub Releases, not Git history - "story · rain day v2" and "Wanling Scroll · guqin/calligraphy/painting" are downloadable for verification

3. Who It's for and How to Use It

Typical uses (per README plus our judgment):

  1. AI comic-drama studios: storyboard previews and batch shot production - PuLID face lock + frame stitching are exactly the two most painful steps in manju
  2. Product ads / character PVs / concept films: R2V reference-to-video keeps product and character consistency
  3. Secondary developers: ComfyUI API + queue scheduling + multi-card batching to wire into business systems
  4. Learners: 33 workflows are 33 dissectible textbooks on how cinematic coherent-film pipelines are assembled

Caveats: at 4 stars with a single maintainer this is an early personal project, not a community-grade engineering effort; CI runs but docs and issue response are limited; "cinematic realism" should be judged by your own runs - README frames land nearer "café/rainy-street realism" than theatrical quality; check copyright and likeness compliance before commercial use (the README notes assets belong to the author, no unauthorized commercial use).

FAQ

Q1: What is ai-manju as an open-source project? A1: A deployment engineering repo for a "realistic AI film/short-drama local generation pipeline": ComfyUI 0.31.0 for orchestration, MiniMax-H3 for joint audio-video generation, Qwen3.6-35B-A3B for Chinese-to-English cinematic prompt polishing, covering T2V/I2V/R2V plus character consistency, first/last-frame continuity, and audio-visual stitching. It ships 33 template workflows and start/stop scripts; model weights are not included.

Q2: What hardware do I need to run it locally? A2: MiniMax-H3 BF16 weights are ~129GB; the author deploys on an 8-card PPU-ZW810E cluster. Laptops can't run full BF16 - consider quantization or cloud inference. The ComfyUI (port 8188) + SGLang polishing (port 8030) front/back architecture can be split and deployed as needed.

Q3: How is cross-shot character consistency achieved? A3: A combo: majicFlus grounds the base look, PuLID locks the face across shots, R2V reference-to-video constrains character/product consistency from a reference image, topped by the FaceDetailer/USDU quality chain and H3 first/last-frame stitching for shot transitions.

Q4: Compared with online video tools (Jimeng/Kling-style), what's the value? A4: Three things: pipeline completeness (polish -> generate -> face lock -> stitch in one system, versus manual relaying between web tools per step); controllability (every ComfyUI node parameter editable, workflows savable); cost shape (local deployment pushes marginal cost to power and depreciation, suited to batch production). The price is deployment friction and hardware investment.

Q5: How mature is it - can it be used commercially today? A5: It's an early personal project (4 stars, single maintainer, still pushing daily); CI runs but docs and responsiveness are limited. Download the Releases clips to judge quality and reproduce a workflow at small scale first. For commercial use mind two things: copyright/likeness compliance of generated assets, and the README's note that assets belong to the author with no unauthorized commercial use.


References

  • GitHub: catiseyeqaq/ai-manju-shengcheng-xitong (README and repo structure, 2026-08-15, 4★/Python/pushed same day)
  • Project docs: docs/production-asset-inventory.md (realistic short-drama full-chain asset inventory)
  • Showcase: GitHub Releases "showcase-videos" ("story · rain day v2", "Wanling Scroll", 1080p with audio)
  • Related components: Comfy-Org/ComfyUI, MiniMax-H3, Qwen3.6-35B-A3B (SGLang deployment)

Assembled from public repo information (2026-08-15); not a hands-on reproduction; quality and hardware per actual verification. Related: AI video tools comparison | Xiaohongshu viral content SOP | ComfyUI local workflow SOP

This article is AI-assisted and human-edited. Last updated: 2026-08-15

FAQ

What is ai-manju as an open-source project?
A deployment engineering repo for a "realistic AI film/short-drama local generation pipeline": ComfyUI 0.31.0 for orchestration, MiniMax-H3 for joint audio-video generation, Qwen3.6-35B-A3B for Chinese-to-English cinematic prompt polishing, covering T2V/I2V/R2V plus character consistency, first/last-frame continuity, and audio-visual stitching. It ships 33 template workflows and start/stop scripts; model weights are not included.
What hardware do I need to run it locally?
MiniMax-H3 BF16 weights are ~129GB; the author deploys on an 8-card PPU-ZW810E cluster. Laptops can't run full BF16 - consider quantization or cloud inference. The ComfyUI (port 8188) + SGLang polishing (port 8030) front/back architecture can be split and deployed as needed.
How is cross-shot character consistency achieved?
A combo: majicFlus grounds the base look, PuLID locks the face across shots, R2V reference-to-video constrains character/product consistency from a reference image, topped by the FaceDetailer/USDU quality chain and H3 first/last-frame stitching for shot transitions.
Compared with online video tools (Jimeng/Kling-style), what's the value?
Three things: pipeline completeness (polish -> generate -> face lock -> stitch in one system, versus manual relaying between web tools per step); controllability (every ComfyUI node parameter editable, workflows savable); cost shape (local deployment pushes marginal cost to power and depreciation, suited to batch production). The price is deployment friction and hardware investment.
How mature is it - can it be used commercially today?
It's an early personal project (4 stars, single maintainer, still pushing daily); CI runs but docs and responsiveness are limited. Download the Releases clips to judge quality and reproduce a workflow at small scale first. For commercial use mind two things: copyright/likeness compliance of generated assets, and the README's note that assets belong to the author with no unauthorized commercial use.

Related

Open Source

Open Video Model LTX-2.5: Diffusion Decoder, 12GB VRAM Entry

Lightricks shipped LTX-2.5 on August 11, 2026, replacing LTX-2.3. Repos (dual form): huggingface.co/Lightricks/LTX-2.5 (HF API verified October 9, 2026: gated=auto, meaning log in and accept terms for automatic access; license tag "other" = the custom LTX-2.x Community License; 1.689M downloads; the Diffusers variant Lightricks/LTX-2.5-Diffusers is also gated=auto), code at github.com/Lightricks/LTX-2. Architecture: a 22B audiovisual backbone plus a custom Gemma 4 12B text encoder and a diffusion video decoder - the VAE decoder is itself a diffusion model, so aggressive latent animation compression and high fidelity no longer trade off; that is this generation's headline architectural change, plus DFR adaptive compute allocation. Capabilities: 4K (3840x2176) HDR, up to 50fps, 24kHz stereo generated in the same pass, native multi-shot (characters/environment/lighting/sound held across cuts), RAW/EXR workflows. Duration conflict, both cited: pinggy says up to 20 seconds per pass (Extend pipeline can lengthen), innfactory says 121 frames at 24fps (~5s) - not settled. VRAM: 80GB+ full, distilled/FP8 variants from 12GB on one card, about 66GiB total download. License: free commercial use under $10M annual revenue, above that requires a Commercial Use Agreement. VFX positioning: HDR/restore/colorize/day-to-night LoRA family, AlphaGen matte extraction, Tiled Fusion 4K/8K upsampling. Ecosystem: ComfyUI (int8_convrot builds), Diffusers 0.40.0, ltx-trainer fine-tuning.

Oct 10, 202610 min read
Open Source

ComfyUI: Why Pros Don't Use Web Image Generators

ComfyUI is the most-starred open-source AI image-generation project on GitHub (122k stars, GPL-3.0), turning diffusion-model generation into a visual node graph. Pros choose it for four reasons: control (ControlNet/LoRA/IPAdapter wireable together), reproducibility (workflows are JSON, re-runnable identically), free/local execution (data stays on your machine), and a custom-node ecosystem. Steep curve, GPU-hungry, GPL-3.0 copyleft - suited for artists and devs who need fine control and a reproducible backend.

Jul 30, 202610 min read