MiniMax H3 in ComfyUI: native workflows, hosted generation and ControlNet

Choose native or hosted generation, download a first-frame workflow and file map, and inspect nodes, canvas and references before a real run.

This guide helps you choose and prepare a workflow. It was checked against the linked official documentation; no installation, local generation or hardware benchmark was performed here.

Reviewed 2026-10-05 · Original teaching drafts have not been independently generated. No speed or quality benchmark was run.

01. Choose native, hosted or controlled conditioning

H3 workflow responsibilities
Route	You maintain	Use it when
Native ComfyUI H3	Installation, model files, node graph and compute	You need to inspect or customize the native workflow.
Hosted H3 in this Studio	Authorized inputs, form settings and task budget	You want a supported form without maintaining local weights.
H3 Fun ControlNet	Its separate patch and control-video workflow	You need the documented pose/depth/edge conditioning path.

The official ComfyUI H3 overview distinguishes native open-weight workflows from hosted access. Its native 768-pixel short-edge canvas and later 2K processing are different from selecting 2K in this Studio. Do not transfer a native node parameter into a hosted form that does not expose it.

02. Inspect environment, storage and license first

Open the current official instructions and inspect the selected template before downloading. A native workflow requires a suitable compute environment, substantial model storage and compatible ComfyUI nodes. This tutorial does not claim that a particular Mac, GPU or notebook service can run it.

Update ComfyUI as required by the current template; an older installation may lack H3 nodes.
Read model downloads and their intended folder locations from that same template/version.
Review the current model license and any patch or custom-node license at the publisher source. Open weights do not by themselves establish commercial permission for every use.
Keep your input-media permission separate from software and model licenses.

Local compute still has time, storage and infrastructure cost. Hosted generation has the selected service quote. Neither route is automatically cost-free.

03. Load a coherent template and its matching files

The official native tutorial provides text-to-video, image-to-video and reference-to-video templates. Start with one current template and keep its loaders, weights, text encoder and audio/video decoding components matched. Obtain exact filenames and paths from that source instead of combining a random node graph with older files.

Template component responsibilities
Component	What to verify
Model loader	The expected H3 model variant and file path
Text conditioning	Matching encoder and intended prompt fields
Visual/audio encode and decode	Required matching components for the selected media
Sampling and output	Template canvas, frame settings and save destination

For a first run, keep the graph close to the supplied template. Add one reference or change one prompt requirement only after the basic path loads and produces a readable output.

04. Download a pinned first-frame exercise

Download one first-frame exercise derived from a pinned official I2V template. The task is one small curtain movement while a lamp stays still. The package contains JSON and instructions, not input photos, weights or a generated result.

Use one coherent version

This graph was inspected against ComfyUI v0.33.0 source. All 21 backend types in the authored graph have declarations there, and its serialized links pass static checks. The docs list 0.30.0+ for the base family; that does not establish this graph’s runtime minimum or maximum. Frontend 1.49.6 is only the template writer version.

ComfyUI/models/diffusion_models/: minimax_h3_fl2va_pruned_int8_convrot.safetensors
ComfyUI/models/text_encoders/: qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
ComfyUI/models/vae/: minimax_h3_video_vae_int8_convrot.safetensors
ComfyUI/models/vae/: minimax_h3_audio_vae_fp32.safetensors
ComfyUI/models/loras/: minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors

Publisher metadata totals about 40.07 GB for the four core files, or 42.03 GB including the referenced LoRA. These are stored-file sizes, not VRAM requirements. Turbo is off, but its loader remains in the graph; verify file references before importing.

Follow the connections

1. LoadImage → first_frame
Select your authorized lamp-and-curtain photo. Leave last_frame empty.
2. ResolutionSelector → width / height
The outer selector overrides the subgraph widgets. Check the connected values first.
3. MiniMaxH3ImageToVideo → BasicGuider
The prompt and first frame produce the positive conditioning.
4. MiniMaxH3ImageToVideo → SamplerCustomAdvanced
Follow the latent path into sampling, not the preview-image path.
5. Video / audio VAE decode → CreateVideo → SaveVideo
Decode video and audio separately, combine, then save under output/video with the h3_lamp_curtain prefix.

Import and prepare the first shot

Prepare your own 1344 × 768 lamp-and-curtain photo. The package filename is a placeholder; select a real file in LoadImage after import.
Drag the JSON into ComfyUI and expand the I2V subgraph. Inspect each model loader, first_frame and the empty last_frame.
Keep the authored 16:9 selector, 0.98 MP, multiple 32, duration 5, seed 42 and turbo_mode=false. The checked formula produces 1344 × 768. Selecting 1.0 MP would instead produce 1376 × 768.
Keep the paired curtain prompt. After environment and file checks, run once and record the actual output. The template aligns duration 5 to 124 frames; 124/24 is about 5.17 seconds, not a measured saved duration.

Static inspection does not establish frontend import, weight loading, memory, inference or playable audio/video. No ComfyUI or GPU run was performed here. Review the lamp, curtain movement, ending and audio after an actual run. The retained MIT workflow notice is separate from model, media and output terms.

Pinned official template · Publisher files · Static validation scope

05. Assign references and a first-shot acceptance plan

An appearance reference, a first/last frame and a timeline guide perform different jobs in native workflows. Follow the selected template slots and labels. An image connected to a timeline anchor is not automatically a text-visible reference in every prompt field; the dedicated Add Guide reference explains that distinction.

Use an authorized image of a lantern beside a curtain. Your first task is one small curtain movement while the lantern remains still. This is a short intent summary of the downloaded exercise; adapt it to the actual template fields. Neither version has been run here.

The reference defines the lantern, curtain and room appearance. A fixed camera keeps the lantern and curtain visible. The curtain sways gently once, then settles. The lantern remains still with steady warm light. No new object, new camera move or scene change. End on a readable resting state.

Check motion, object shape and ending separately. Choose frame count, canvas and saving options in the native template; set hosted duration and native frame count separately.

06. Treat Fun ControlNet as a separate workflow

The official Fun ControlNet page describes a separate H3 patch for pose, depth or edge control videos. Use its documented control preparation and patch connections. This is different from uploading an ordinary action video to the hosted H3 reference form.

If all you need is a general action reference, the extra control workflow may not be necessary. If you require its conditioning path, first verify that the source and selected patch fit; writing “pose control” in a hosted prompt does not install or activate that patch.

ComfyUI: H3 Fun ControlNet

07. Locate the failure before replacing the workflow

Native first-run checks
Symptom	First check	Next action
Missing H3 node	Core version and startup import errors	Follow the template update requirement; do not guess an unrelated package.
Model not found	Filename, download completion and loader path	Match the current official file list and folder.
Encoding/decoding error	Matching components and input types	Return to a coherent template before adding more media.
Memory or loading failure	Chosen weights, canvas and compute environment	Review the documented requirements; consider a hosted route if maintaining this environment is unsuitable.
Output does not follow a reference	Input role, conflicting prompt and actual guide connections	Simplify to one role and one action, then inspect the output.

Save the template version, settings and error message. Do not classify a failed local load as evidence that the hosted model cannot perform the task, or a hosted success as proof that your local environment is configured correctly.

08. Choose one next step instead of downloading everything

Read the current official template first if your goal is native execution. For exact guide-node inputs, use Add Guide; for the separately maintained community director package, inspect its own source and version. The publisher prompt-writing Skill is a writing aid, not a renderer.

For the hosted alternative, select supported materials in the Studio form and inspect its quote. This site does not execute your ComfyUI graph or expose its timeline-node controls.