Use this page as a field guide. The model, workflow, and directory facts below are linked to official sources. Hardware advice is a starting point, not a cross-GPU benchmark.
Choose the first run
Start with T2V. Move to image or frame references once the local install is clean.
Text to Video
Describe the subject, evolving action, camera, and audio in one structured prompt.
- Input: text prompt
- Best first check: official baseline
- Template: Comfy-Org T2V
Image to Video
Use an image as the first frame and describe what changes after the reference frame.
- Input: first-frame image
- Prompt: preserve then animate
- Guide: reference alignment
First + Last Frame
Describe an observable transition from the first frame toward the final frame.
- Input: first and final frame
- Prompt: transition path
- Guide: H3 prompt structure
Download the files
Download the official model files, then place each one in the matching ComfyUI folder. The current official T2V and I2V template manifests list all four files.
ComfyUI/ └── models/ ├── vae/ │ ├── minimax_h3_video_vae_fp16.safetensors │ └── minimax_h3_audio_vae_fp32.safetensors ├── diffusion_models/ │ └── minimax_h3_fl2va_pruned_int8_convrot.safetensors └── text_encoders/ └── qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
minimax_h3_video_vae_fp16.safetensorsminimax_h3_audio_vae_fp32.safetensorsminimax_h3_fl2va_pruned_int8_convrot.safetensorsmodels/diffusion_modelsUseRequired for the current official T2V and I2V templates.Download file qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors- 01Download all four files from the official model links above.
- 02Copy each file into the exact
ComfyUI/models/subfolder shown. - 03Keep every filename unchanged; restart or rescan ComfyUI after copying.
- 04Open the official T2V or I2V template only after all four files are visible.
These download URLs point to the Comfy-Org MiniMax H3 repository on Hugging Face. After the files are visible, open the official T2V template or official I2V template . If nodes or model selectors are missing, update ComfyUI before changing random settings.
Start from a known baseline
The official template gives a clean starting point. The other two cards are editorial starting points, not benchmarks.
20 steps · res_multistep sampler · simple scheduler. Use this as the first comparison point before tuning for your hardware.
A smaller official-size starting point for a first local check.
Recommended starting pointA middle size from the official resolution reference before the full baseline.
Recommended starting pointThe current T2V template baseline with 20 steps, res_multistep, and simple.
There is no universal minimum VRAM promise here. If you hit OOM, lower resolution and duration first, then check ComfyUI version, attention/offload, and model precision.
Treat these as compatibility paths, not magic switches. Update ComfyUI, confirm the workflow loads, then test one attention or offload change at a time.
Fix the first failure
Classify the error before changing settings. One clean run is more useful than ten random tweaks.
Model selector is empty
- Check the exact filename.
- Move it into the matching folder.
- Restart or rescan ComfyUI.
Workflow cannot load
- Update ComfyUI.
- Open the current official template.
- Compare the node names again.
Run stops in sampling
- Lower resolution and duration.
- Check attention/offload options.
- Change one setting, then retry.
For version and workflow-template issues, start with the ComfyUI update guide . For model and node bugs, use the official support record .
Build a structured H3 prompt
This formatter stays in your browser. It uses curated director variations instead of random word fragments.
integrated_multimodal_description: Cinematic realism. medium shot of a woman in a red raincoat in a neon-lit city street after dark. a woman in a red raincoat walks through the rain and turns toward the camera. The camera uses slow push in, keeping the motion clear and coherent. Neon night. End on the woman pauses beneath a glowing sign. no text, no logos, keep the face consistent overall_soundscape: rainfall, footsteps on wet pavement, distant traffic non_diegetic_music: N/A
Prompt syntax is based on the H3 Base Prompt Guide . Full Reference / Ref2VA uses a separate Reference Prompt Guide and is not simplified here.
Read the shape
These examples remain in the HTML so the guide is useful even before the builder loads.
Cinematic T2V / rain and neonCopyable text
integrated_multimodal_description: Cinematic realism. Medium shot of a woman in a red raincoat in a neon-lit city street after dark. She walks through the rain and turns toward the camera. The camera uses a slow push in, keeping the motion clear and coherent. Neon night. End on the woman pausing beneath a glowing sign. no text, no logos, keep the face consistent. overall_soundscape: rainfall, footsteps on wet pavement, distant traffic non_diegetic_music: N/A
I2V / first-frame continuityCopyable text
<Picture 1> is the exact first frame at 0.00 seconds. Preserve the subject, composition, clothing, and environment from the reference; describe only what changes after the first frame. integrated_multimodal_description: A close-up of the ceramic cup remains on the wooden table as steam slowly rises and the camera performs a restrained push in. Warm window light stays consistent. overall_soundscape: quiet room tone, soft kettle hiss non_diegetic_music: N/A
Dialogue / bound speakerCopyable text
integrated_multimodal_description: Cinematic realism, medium shot of a woman at a train platform at dusk. The woman turns toward the arriving train and says: <d>[English] I get off at the next station.</d> Keep the line associated with the named speaker and keep the camera on her readable facial performance. overall_soundscape: station ambience, rail vibration, footsteps non_diegetic_music: N/A
Questions after the first run
Short answers for the queries that usually appear when a new workflow meets an old install.
How do I run MiniMax H3 in ComfyUI?
Update ComfyUI, place the official model files in the matching folders, open the official H3 workflow template, and start with the 1344 × 768 / 5s baseline.
Where should MiniMax H3 model files go?
VAE files go in models/vae, the diffusion model goes in models/diffusion_models, and the text encoder goes in models/text_encoders. Use the exact filenames shown in the Models section.
How much VRAM does MiniMax H3 need?
This guide does not publish a universal minimum. Hardware, precision, resolution, duration, and offload choices change the result. For OOM, lower resolution and duration first, then check the install and attention path.
What is the best starting setting?
Use the current official T2V baseline: 1344 × 768, 5 seconds, 20 steps, res_multistep sampler, and simple scheduler. Starter and Balanced cards are editorial starting points for local testing.
Does the Prompt Builder send my prompt anywhere?
No. The builder runs in the browser, does not call an AI API, does not save text, and does not send your free-text fields to analytics.
Does this page support Full Reference / Ref2VA?
The page links the official Reference Prompt Guide, but the first builder supports T2V, I2V, and First + Last Frame only. It does not pretend to format the stricter multi-reference syntax.
Check the source
This page is an independent editorial guide. Verify fast-moving details against the official pages before a production run.
- MiniMaxH3 release and model overview
- MiniMax APIVideo generation API reference
- ComfyUIH3 support PR #15224
- Workflow templateOfficial H3 T2V workflow · Template repository
- Model filesComfy-Org H3 model repository
- Prompt guideBase prompt guide · Reference prompt guide
MiniMax and ComfyUI are trademarks and projects of their respective owners. This site is an independent guide and is not affiliated with either project.