MiniMax H3 · ComfyUI

MiniMax H3 ComfyUI workflow: T2V, I2V and R2V

ComfyUI ships the MiniMax H3 ComfyUI workflow templates natively from version 0.30.0 — no H3-specific custom node is required. Open Template Library › Video › MiniMax H3 and you get three.

Everything below was read out of the template files themselves — the three MiniMax H3 ComfyUI workflow JSONs in github.com/Comfy-Org/workflow_templates — rather than paraphrased from the tutorial page. That is why the node names, the shipped defaults and the duration grid here disagree with most write-ups of this workflow.

The three official MiniMax H3 ComfyUI workflows

Each MiniMax H3 workflow is one JSON file, and each one loads a specific checkpoint. Picking the wrong checkpoint for the workflow is the most common way a first run fails.

T2V

Text to Video

Generates a 5–15 second clip with native stereo audio from a prompt alone.

Inputs: One prompt, plus a duration in seconds and a resolution.
Checkpoint: minimax_h3_fl2va_pruned_int8_convrot.safetensors
Nodes: ResolutionSelector · MiniMaxH3ImageToVideo · SaveVideo
JSON: video_minimax_h3_t2v.json

I2V

Image to Video

Turns a single still into a coherent clip with synchronized stereo audio, with optional first/last frame control.

Inputs: A first frame, an optional last frame, and a prompt describing the motion between them.
Checkpoint: minimax_h3_fl2va_pruned_int8_convrot.safetensors
Nodes: LoadImage · ImageScaleToTotalPixels · MiniMaxH3ImageToVideo · SaveVideo
JSON: video_minimax_h3_i2v.json

R2V

Reference to Video

Locks character, style or motion from reference material, then generates a new shot.

Inputs: Up to 9 reference images, 3 reference videos — each of which can carry its own soundtrack — and 3 standalone audio references.
Checkpoint: minimax_h3_ref2va_pruned_int8_convrot.safetensors
Nodes: LoadImage · MiniMaxH3ReferenceToVideo · VAEDecodeAudio
JSON: video_minimax_h3_r2v.json

There is no MiniMaxH3TextToVideo node

Open the T2V and I2V template files side by side and they reference the same subgraph — id 4c314f31-ecda-4b08-ae98-faaba1bf613f, named "Image to Video (MiniMax H3)". The generation node inside it is MiniMaxH3ImageToVideo in both. Text-to-video is that same graph with nothing wired into the image input, which is also why both templates load the same fl2va checkpoint.

If a guide hands you a spec sheet for a node called MiniMaxH3TextToVideo, it was not read off the workflow. Only two H3 nodes exist in the official templates: MiniMaxH3ImageToVideo and MiniMaxH3ReferenceToVideo.

Source: Comfy-Org/workflow_templates · video_minimax_h3_t2v.json + _i2v.json

What is inside the MiniMax H3 workflow graph

Every MiniMax H3 ComfyUI workflow runs the same six-stage chain. In the T2V and I2V templates it is collapsed into a single subgraph node, so you only see it after you expand it; the R2V workflow ships with the same chain laid out flat.

01
Load
UNETLoader · CLIPLoader · VAELoader ×2
One fl2va or ref2va checkpoint, the Qwen3-VL text encoder, and both VAEs — video and audio are separate files and both are mandatory.
02
Size
ResolutionSelector · ComfyMathExpression
Aspect ratio and megapixels become width and height; a duration in seconds becomes a frame count on H3's 17-frame grid.
03
Condition
MiniMaxH3ImageToVideo / MiniMaxH3ReferenceToVideo
Prompt, dimensions, frame count and any reference material go in; conditioning and an empty latent come out.
04
Sample
BasicGuider · KSamplerSelect · BasicScheduler · SamplerCustomAdvanced
res_multistep with the simple scheduler at 20 steps. Video and audio latents advance together in one pass.
05
Decode
VAEDecode · VAEDecodeAudio
Two decodes, two VAEs. Miss the audio branch and you get a silent clip from a model that generated sound.
06
Write
CreateVideo · SaveVideo
Frames plus the decoded audio track are muxed at 24fps and written out.

The settings every MiniMax H3 workflow ships with

These are the widget values baked into the template files. They decide what your first run looks like, and one of them — megapixels — is set to a draft value that surprises almost everyone.

Sampler: res_multistep; KSamplerSelect — Same in all three templates. H3 advances a video and an audio latent together, and this is the sampler the templates were tuned against.
Scheduler: simple; BasicScheduler — Paired with 20 steps at denoise 1.0.
Steps: 20; BasicScheduler — The baseline cost of one clip. Cutting this without the turbo LoRA degrades the audio branch first, not the picture.
Megapixels: 0.4; ResolutionSelector — This is a draft setting — 864×480 at 16:9. The single most common reason a first H3 render looks soft is that nobody raised it.
Multiple: 32; ResolutionSelector — H3's resolution grid. Leave it alone; every dimension gets rounded to a multiple of 32 anyway.
Frame rate: 24 fps; CreateVideo — Fixed. The duration you ask for is converted to frames at 24fps before it hits the model.
Text encoder type: minimax; CLIPLoader — Not a cosmetic label. Rebuild the graph by hand with the wrong loader type and the encoder loads but conditions nothing.

Source: Comfy-Org/workflow_templates · widget values read from the template JSON

Resolution: what the workflow gives you per megapixel

Max resolution: 2K
Frame rate: 24 fps
Max duration: ~15s
Short edge cap: 768px short edge, capped at 768×1344
Reference images: up to 9
Reference video / audio: 3 / 3

Source: docs.comfy.org / MiniMax release notes

You do not type a width and a height. The ResolutionSelector node takes an aspect ratio and a megapixel target and derives both, rounded to a multiple of 32. The templates carry this conversion table themselves:

Megapixels	Output at 16:9	Note
0.2	608 × 352
0.3	736 × 416
0.4	864 × 480	Template default — draft quality
0.6	1056 × 608
0.8	1216 × 672
0.98	1344 × 768	H3's native canvas, exactly
1.0	1376 × 768	Past the 1344 cap
2.0	1920 × 1088	Well past it

Source: Table shipped inside the official MiniMax H3 templates

The row you want is 0.98, not 1.0. H3’s native canvas is a 768px short edge capped at 768×1344, and 0.98 megapixels lands on 1344×768 exactly. Every row above it hands you dimensions past that cap — you pay for the pixels without the model having been trained to fill them. Compose inside the cap and upscale afterwards.

Duration: the number you type is not the number you get

The duration input is seconds, but H3 generates in blocks of 17 frames. A Math Expression node inside the workflow converts one to the other and always rounds up:

max(5, round(a * 24)) + (5 - (max(5, round(a * 24)) % 17)) % 17

You ask for	Frames	You get
4 s	107	4.46 s
5 s	124	5.17 s
6 s	158	6.58 s
8 s	192	8.00 s — lands exactly
10 s	243	10.13 s
12 s	294	12.25 s
15 s	362	15.08 s

Ask for 6 seconds and the workflow renders 158 frames — 6.58 seconds, and you pay for all of them. Across the whole 5–15 second range there is exactly one duration that divides cleanly into 24fps: 8 seconds, 192 frames. If you are cutting to music or stitching shots, build on 8.

Computed from the expression shipped in the templates' Math Expression node — arithmetic on a published formula, not a measurement

Common MiniMax H3 workflow errors and what causes them

Most reports of a broken MiniMax H3 ComfyUI workflow are not bugs — they are a default doing exactly what it was set to do, or a checkpoint that belongs to a different template. In rough order of how often they come up:

MiniMax H3 doesn't appear in the Template Library

Cause · ComfyUI is older than 0.30.0. The H3 templates arrived with that release.

Fix · Update ComfyUI to 0.30.0 or later, restart, then check Template Library › Video again. You can also load the template JSON straight from the workflow_templates repo.

The clip is sharp but tiny, or looks soft after upscaling

Cause · All three templates ship with ResolutionSelector at 0.4 megapixels — 864×480 at 16:9. That is a draft setting, not the model's ceiling.

Fix · Raise megapixels to 0.98 for the native 1344×768 canvas. Going higher hands you dimensions past H3's stated cap rather than more detail.

Video renders but there is no audio

Cause · The audio branch is a separate path: a second VAELoader for minimax_h3_audio_vae_fp32, a VAEDecodeAudio node, and an audio input on CreateVideo. Rebuilt or trimmed graphs lose it easily.

Fix · Check that both VAEs are loaded and that VAEDecodeAudio's output reaches CreateVideo. If the graph is intact, the cause is the prompt — see the audio block note below.

Reference images are ignored in the R2V workflow

Cause · References are bound by tag in the prompt text, not by wire order alone. The official template's own prompt says "Use <Picture 2> and <Picture 1> as references".

Fix · Name every reference explicitly in the prompt with <Picture 1>, <Picture 2> and so on, and give each one a job — one for the character, one for the style.

The R2V template errors on the checkpoint you loaded

Cause · T2V and I2V want an fl2va checkpoint; R2V wants ref2va. They are two separate downloads of comparable size, not one file with a switch.

Fix · Load minimax_h3_ref2va_pruned_int8_convrot.safetensors for reference-to-video and keep the fl2va file for the other two workflows.

Node error when loading Q8_CR or U16G weights

Cause · Those builds target a specific ComfyUI custom node rather than the official workflow — they are not general-purpose GGUF files.

Fix · Either install the custom node the publisher specifies, or switch to a more broadly compatible build such as Q4_0 or Q8_0.

Weights load fine, then generation OOMs

Cause · File size is not peak VRAM. The VAE, text encoder, latents and audio branch all allocate on top, and they grow with resolution and duration.

Fix · Establish a working baseline at the stock 0.4 megapixels and 5 seconds first, then raise resolution and length one step at a time until it breaks.

The clip came back with music you never asked for

Cause · H3 generates picture and audio together. With no audio block in the prompt, the model chooses one for you.

Fix · Write an explicit audio block. For a silent clip, state it: "Audio: silent. No music, no room tone."

Skip the ComfyUI workflow entirely

If you’re here because the workflow won’t load or your card won’t hold the weights, H3Video’s hosted text-to-video generator avoids the local model setup.

Compare hosted pricing →

Still set on running the MiniMax H3 workflow locally? Start with the VRAM requirements breakdown to pick a quant your card can actually survive at your target resolution. Once it’s running, the prompt guide will save you more time than any node tweak — most disappointing outputs are prompt problems, not workflow problems, and the audio block that stops H3 inventing a soundtrack lives there.

For the files themselves — which safetensors belongs in diffusion_models versus text_encoders, and why both VAEs are mandatory — see the local setup guide. And once the stock workflow runs, the turbo LoRA drops sampling from the stock 20 steps to 4. It replaces the sampler the template ships with, because at that step count the audio branch needs its own schedule rather than sharing the video’s.

Frequently asked questions

Which ComfyUI version supports MiniMax H3?

ComfyUI 0.30.0 or later. Every official MiniMax H3 ComfyUI workflow ships with that release under Template Library › Video › MiniMax H3, and no H3-specific custom node is required to run them.

What MiniMax H3 workflows does ComfyUI include?

Three: Text to Video, Image to Video (with optional first and last frame control), and Reference to Video, which accepts up to nine reference images, three reference videos and three standalone audio references.

Where can I download the MiniMax H3 ComfyUI workflow JSON?

All three templates live in the Comfy-Org/workflow_templates repository as video_minimax_h3_t2v.json, video_minimax_h3_i2v.json and video_minimax_h3_r2v.json. On ComfyUI 0.30.0 or later you do not need to download anything — the same files are already in your Template Library.

Is there a MiniMaxH3TextToVideo node?

No. The text-to-video and image-to-video templates reference the same subgraph, and the generation node inside both is MiniMaxH3ImageToVideo — text-to-video is that graph with nothing connected to the image input. The only other H3 node in the official workflows is MiniMaxH3ReferenceToVideo.

What sampler and step count does the MiniMax H3 ComfyUI workflow use?

The templates ship with the res_multistep sampler, the simple scheduler, 20 steps and denoise 1.0, rendering at 24fps. Those are the values baked into the template JSON, so 20 steps is the baseline cost of one clip unless you add the turbo LoRA.

Why does my MiniMax H3 workflow output look low resolution?

Because the ResolutionSelector node ships at 0.4 megapixels, which is 864×480 at 16:9 — a draft setting, not the model's ceiling. Raise it to 0.98 for H3's native 1344×768 canvas. Going above that produces dimensions past the model's 768×1344 cap rather than more detail.

Why won't my quantized H3 weights load in ComfyUI?

Some community builds — notably the Q8_CR and U16G variants — are made for specific ComfyUI custom nodes rather than the official workflow, and will fail to load elsewhere. Either install the node the publisher specifies, or use a more broadly compatible build such as Q4_0 or Q8_0.

What is the maximum output resolution?

Up to 2K at 24fps and about 15 seconds, with the short edge capped at 768px and an overall cap of 768×1344. Requesting more than that will be clamped back down, so choose an aspect ratio inside the cap and upscale afterwards if you need more.