One Free AI Model Just Built an Entire Anime Short Film on a Home PC

MiniMax H3 is a free, open-weight video model that runs locally in ComfyUI and handles almost every step of making an animated film: character sheets, turnarounds, environments, voices, sound effects, and final video shots. Creator Mickmumpitz used this single model to produce a full anime short called Shibuwaka, and shared the free ComfyUI workflows so you can build the same pipeline on your own computer.

Forget juggling a dozen separate models that eat your hard drive. The pitch here is simple: one download, one interface, one consistent workflow layout. Here is how the pipeline works and why it matters for anyone making animation, concept art, or short films with AI.

Why MiniMax H3 Is Different

The video centers on the reference version of MiniMax H3. In plain terms, it is a video generator that understands references. You hand it text, audio files, images, or video clips, tell it what to do with them, and it builds the shot around your inputs.

Every workflow in the project follows this same principle and shares the same layout. Learn one and you understand the whole pipeline. Because H3 also generates audio, it doubles as a voice cloner, a voice generator, and a sound effects tool.

Everything runs inside ComfyUI, the free node-based interface for AI models. Each step of the project has a ready-made template. You drag it into ComfyUI, install any missing custom nodes, download the model components, and place them in the right folders. The official ComfyUI documentation covers the built-in H3 templates and needs ComfyUI 0.30.0 or later, and the quantized components are hosted in the Comfy-Org MiniMax-H3 repository on Hugging Face. The model ships under its own community license, so read it before using results commercially. Mickmumpitz hosts the free workflows and the rest of the downloads in his original workflow post.

AI is moving into every stage of production, and our Artificial Intelligence (AI) coverage tracks the tools that actually matter to artists.

Step-by-Step: The Pipeline From Idea to Film

Step 1: Design characters with the character sheet workflow

• Name your character (this doubles as the output folder), then set resolution and seed.
• Leave turbo mode on for speed. Turn it off only if you want a quality bump.
• The prompt has three parts: the style, the character description, and a description of the sheet layout. Presets for extra views are built in.
• Add reference images by un-bypassing the reference nodes with Ctrl+B. For high-resolution references, a zoomed-in face crop as a second reference keeps details like faces and shoes consistent.
• Enable the upscaler for a sharper final sheet, and add a line like a three-quarter view to the prompt for another angle.

Despite the name, it is not limited to characters or anime. It can turn sketches into full sheets, build consistent environments, and create asset sheets, which the team used for the cars in the film.

Step 2: Spin up a turnaround

A separate workflow makes a turnaround video plus an eight-view sheet of any character or object. It needs one extra model, with the download link in the original post. Drop in your image, add a short subject description, and run it.

Step 3: Build environments and style frames

Here the video model becomes a reference-based image editor. Work purely from prompts or add references, then tell the model in natural language what to do with them. Use it for start and end frames, style frames, and consistent environments. Mickmumpitz suggests editing these images together early to lock the film’s timing, which helps when you write voice lines.

Step 4: Create consistent voices

You have three options: record your own voice and pin the video to it, generate voice lines first, or generate voices together with the video.

• For the voice workflow, name the voice, set the length, and turn turbo mode off for full quality.
• Each run can sound slightly different, so enable the voice reference and feed in a clip of the voice you want. In the demo, a clip of his own voice produced a startlingly close clone.

Step 5: Generate the video shots

The video workflow follows the same layout. You name the shot, set the length and seed, and keep turbo on for fast iteration. Then you can enable the upscale group at the end.

• Choose an existing voice line to pin the animation to, or use an audio reference to have the character speak in your reference voice.
• The prompt is split into style, the world of the scene, and the action. H3 understands precise direction, including multiple shots in one generation and actual time codes.
• References are numbered picture one, picture two, and so on, so you can point to them directly in the prompt.
• Optional start and end frame inputs sit at the bottom of the workflow.

Step 6: Steer the animation with extra control

Beyond text, the video workflow accepts three stronger forms of control:

• Video references: Film yourself acting out a scene, drop it in, and use the follow-acting option. It is a fun way to shoot a movie and can sharply improve shot quality.
• Storyboards: Feed in an entire storyboard with a detailed prompt, and the model follows your composition surprisingly well. If you plan films this way, our roundup of storyboard software is a good companion.
• 3D layouts from Blender: For a long shot with a very specific camera move, the team animated the move in Blender, loaded it as a reference, added a prompt, and it worked. Prompting alone could not have done it.

That last point is where 3D artists should pay attention. If you already block out scenes in Blender, our guide on blocking in Blender and finishing with AI shows a similar mindset, and you can go further by learning to render Blender animations with AI locally for free. Extra tools for your scenes live in our Blender addons collection.

Step 7: Add music and sound effects

You can generate music with the voice workflow, but Mickmumpitz says results are hit or miss, mostly miss, so the film uses stock music. Sound effects are a different story: type what you want and you get something very decent.

What About 4K and Hardware?

The team’s goal was 4K quality locally. They built a local H3 upscale workflow that rerenders a shot at high resolution using the same prompt and references, adding detail such as crisp ink lines. On a GeForce RTX 5080, the main GPU for the project, upscaling takes around 15 minutes per shot. A 5090 is faster.

This part of the video is sponsored by Nvidia, which highlights new RTX video nodes in ComfyUI that upscale shots to 4K in seconds, interpolate frames for slow motion, and convert footage to HDR. The upside of running locally is no subscription costs, no queues, and your ideas never leave your machine.

Watch the Full Breakdown

Frequently Asked Questions

Is MiniMax H3 really free and local? The weights are downloadable and the ComfyUI workflows run on your own GPU, so there are no per-shot fees. Check the model’s community license for usage terms.

Do I need a high-end GPU? The project ran mostly on an RTX 5080, and upscaling takes about 15 minutes per shot there. Faster cards cut that down.

Can I use it for styles other than anime? Yes. The character sheet workflow works for all kinds of styles, plus environments and asset sheets.

Where do I get the workflows? The templates are linked from Mickmumpitz’s post. Patreon supporters also get the example files made for the film.

Roll Camera on Your Own Machine

The takeaway is bigger than one model. A single reference-driven video generator, a consistent workflow layout, and a mid-range-to-high-end GPU can now carry a short film from first sketch to final cut. Start small: build one character sheet, then one shot, and iterate from there.

Tried the workflows or hit a snag? Drop a comment with your results and questions. For more pipelines like this, browse our tutorials, see how AI performance capture is changing animation, or read about Claude running a Blender and ComfyUI pipeline solo. If you want help bringing this into a real production, the team at Gachoki Studios does animation and VFX work, and you can get in touch anytime.