loyc/minimax-h3-anime-pipeline
Watch
1
0
Fork
You've already forked minimax-h3-anime-pipeline
 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions
Summary + free ComfyUI workflow pack (Mickmumpitz MiniMax H3 anime short pipeline) + model links
1 commit 1 branch 0 tags 147 KiB
Find a file
Exact
Repository files (latest commit first)
Filename Latest commit message Latest commit date
loyc 1e67f5d05f MiniMax H3 anime pipeline: video summary + free workflow pack (7 ComfyUI workflows)
2026-10-28 13:31:32 -07:00
workflows MiniMax H3 anime pipeline: video summary + free workflow pack (7 ComfyUI workflows) 2026-10-28 13:31:32 -07:00
README.md MiniMax H3 anime pipeline: video summary + free workflow pack (7 ComfyUI workflows) 2026-10-28 13:31:32 -07:00

README.md

Make Your Own ANIME With The BEST AI Tool — MiniMax H3 Pipeline (Mickmumpitz)

Summary + the free workflow files from Mickmumpitz's MiniMax H3 anime-film pipeline.

Video: Make Your Own ANIME With The BEST AI Tool [2026 MASTERCLASS] — Mickmumpitz, 12:35, published 2026-10-28 Workflow source: Free Patreon post with attachments · Example files post (Advanced patrons)

TL;DR

Mickmumpitz made a full 4K anime short film (SHIBUWAKA) using one single AI model — MiniMax H3 (reference version) — running entirely locally in ComfyUI, and released the entire production pipeline for free.

Instead of one giant node graph, the pipeline is a suite of small workflows, one per production step. They all share the same layout: references (text / image / audio / video) go in one side, you describe in the prompt what each reference is for, and the result comes out the other. Learn one → you know them all.

H3 is a reference-to-video model that understands text, audio, images and video references at once — and it also generates audio (voices, music, sound effects), so the whole film — images, voices, music, video — comes from one model.

The workflow pack (in workflows/)

Workflow What it does
H3-CHARACTER-SHEET Character sheets from a prompt, sketch, 3D layout, or reference photos. Also used for consistent environments and the cars' asset sheets. 3-part prompt (style / character / sheet layout), presets for extra views, optional reference images (un-bypass with Ctrl+B), face-crop second reference trick for consistency, optional upscaler.
H3-TURNAROUND Turnaround video + 8-view sheet from a reference image/sheet. Needs one extra LoRA (360° orbit LoRA, link below).
H3-IMAGE H3 as a reference-based image editor: start frames, end frames, style frames, consistent locations. Fully prompt-based or with references.
H3-VOICE Voice generation/cloning. Create a line once, then feed it back as a voice reference for consistency (turbo OFF for voice — use full quality).
H3-VIDEO Video generation. Pin timing to an existing voice line, or use an audio reference for new speech. Supports video references, drawn arrows, storyboards, Blender layouts (Union ControlNet needed for VIDEO REFERENCE mode). Turbo ON recommended for iteration.
H3-UPSCALE Re-renders a shot at higher resolution with the same prompt + references (adds crisp ink-line detail). This is how the whole film got to 4K. ~15 min/shot on RTX 5080; 5090 faster.
H3-UPSCALE-RTX4K RTX-video-nodes variant of upscaling (seconds instead of minutes, for newer RTX cards).

All files are standard ComfyUI workflow JSONs (verified — e.g. H3-VOICE = 52 nodes incl. MickmumpitzLabel, UNETLoader, SaveAudio). Drag & drop into ComfyUI.

Setup

  1. ComfyUI + the custom nodes it prompts for (ComfyUI Manager → search "Mickmumpitz" for his node pack: ComfyUI-Mickmumpitz-Nodes).
  2. Drag & drop a workflow from workflows/ into ComfyUI.
  3. Download the models below into the Comfy folder structure.

Core models (all workflows)

File Into Size Link
MiniMax H3 Ref2VA int8 models/diffusion_models/ 19.5 GiB Comfy-Org/MiniMax-H3
Qwen3-VL 32B NVFP4 text encoder models/text_encoders/ 14.6 GiB link
Video VAE fp16 models/vae/ 4.9 GiB link
Audio VAE fp32 models/vae/ 577 MiB link
Turbo LoRA v4 models/loras/ 592 MiB drbaph/MiniMax-H3-Turbo-Lora-ComfyUI

Workflow-specific

File Used by Size Link
RealESRGAN x2 → models/upscale_models/ upscale groups 64 MiB ai-forever/Real-ESRGAN
360° orbit LoRA → models/loras/ H3-TURNAROUND 148 MiB pablodawson/MiniMax-H3-360-Orbit-LoRA
Union ControlNet 2.0 → models/model_patches/ H3-VIDEO (VIDEO REFERENCE on) 4.2 GiB link

Low-VRAM / alternative options (swap in the loader)

  • 4-bit model, 11.0 GiB, ~2× slower than int8: Kijai/MiniMax-H3-experimental
  • Smaller video VAE int8 (Kijai), 3.2 GB: link
  • Text encoder int8 (no NVFP4, pre-Blackwell cards), 25.3 GiB: link

Key techniques from the video

  • One model, many jobs — images, video, voices, music, SFX all from H3; no clogged drive of a dozen models.
  • Face-crop trick — for character consistency, screenshot the face from a high-res reference and feed it as a second reference.
  • Voice consistency — generate a line once → reuse it as an audio reference so the voice never drifts (turbo off for voice).
  • Voice-pinned video — generate dialogue first, then the video workflow pins animation timing to the existing audio track.
  • 3-part prompting — style → world description → shot action; H3 follows very precise natural-language prompts.
  • 3D control — video references, drawn arrows, storyboard panels, or raw Blender layouts drive the animation (Union ControlNet).
  • Upscale last — same prompt + refs re-rendered at high res; this took the whole film to 4K.
  • Everything local: no subscriptions, no queues, private, iterate forever. (Video sponsored by NVIDIA; pipeline built on RTX 5080/5090.)

Links

About the workflow files in workflows/

Downloaded 2026-10-28 from the public (free) Patreon post above; each file verified to match the attachment size Patreon reports. Patreon retains ownership of Mickmumpitz's workflow content — these are copies for personal use; support him on Patreon for the example files and new releases: https://www.patreon.com/Mickmumpitz