Video generation
MiniMax-H3 · Single device
A single-device MiniMax-H3 FL2VA setup for video generation with native audio and optional 1080p upscaling. The performance results cover the image-to-video workflow.
01 / Model parameters
Model architecture & capabilities
MiniMax-H3 generates video with native audio from text, images, or multimodal references. FL2VA accepts text alone, a single image, or first and last frames. Ref2VA supports combinations of reference images, video, and audio. Each variant and workflow has its own resource requirements and generation times.
- Short videos from text or images
- Image-to-video with audio
- Higher resolution through upscaling workflows
- Publisher / model
- MiniMax · MiniMaxAI/MiniMax-H3
- Parameter count / architecture
- 33B Omni-Transformer dense backbone; approx. 13B AdaLN branches can be precomputed and pruned, plus a text encoder and audio/video VAEs
- Input / output
- Text, image, and multimodal reference input; video and 32 kHz stereo audio output
- Official generation specifications
- 4–15 seconds, 24 FPS, base resolution of 768 pixels on the shorter side; actual local output depends on workflow settings
- Generation variant
- FL2VA: 0 images for text-to-video, 1 for image-to-video, 2 for first/last frames; Ref2VA: multimodal reference generation
- License
- MiniMax-H3 Community License
- Quantization / variant used here
- FL2VA; Comfy-Org INT8 pruned diffusion weights approx. 20.97 GB
- Text encoder & VAE
- Qwen3-VL 32B NVFP4 AWQ approx. 15.69 GB; video VAE in FP16, audio VAE in FP32
FL2VA supports text-to-video and image-to-video. Single-device MegaCube tests on this page cover only image-to-video and upscaling workflows. 1080p is upscaled output, not the native generation resolution.
Model information checked on 2026-09-20
02 / Deployment
MegaCube deployment configuration
- Compute devices
- 1 × MegaCube · NVIDIA GB10 · 128 GB
- Software environment
- ComfyUI 0.35.0；Python 3.12.3；PyTorch 2.13.0+cu130
- Generation settings
- Turbo LoRA v4; strength=1.3; simple scheduler; current workflow uses 10 steps
- Output
- 1376×768 native audio/video; RealESRGAN x2 post-processing to 1920×1080
- Execution method
- One local task at a time, port 8188
Prepare the model components
Download the diffusion model, text encoder, both VAEs, and Turbo LoRA; verify SHA256 against the pinned versions.
Configure ComfyUI
Prepare ComfyUI and its nodes in an ARM64 environment, then load the validated model combination.
Import a reference image
Set the image, prompt, frame count, and steps through the MiniMaxH3ImageToVideo workflow.
Save the output
Generate video with native audio, then optionally upscale and encode the final video; save the original and upscaled outputs separately.
Find and download this model in the PINVOU Model Store, then follow its model card to deploy and enable the service.
Manage models in PINVOU →Model components & pinned versions
| Component | Repository & pinned revision | Actual filename | Precision / current settings | Size |
|---|---|---|---|---|
| Diffusion model | Comfy-Org/MiniMax-H3 a98869194787969724c7425d95d0ed73ce9202af | minimax_h3_fl2va_pruned_int8_convrot.safetensors | INT8 pruned UNETLoader weight_dtype=default | 20,970,379,616 bytes 20.97 GB |
| Text encoder | Comfy-Org/MiniMax-H3 a98869194787969724c7425d95d0ed73ce9202af | qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors | Qwen3-VL 32B NVFP4 AWQ，type=minimax | 15,687,142,551 bytes 15.69 GB |
| Video VAE | Comfy-Org/MiniMax-H3 a98869194787969724c7425d95d0ed73ce9202af | minimax_h3_video_vae_fp16.safetensors | FP16 video decoding | 5,207,808,496 bytes 5.21 GB |
| Audio VAE | Comfy-Org/MiniMax-H3 a98869194787969724c7425d95d0ed73ce9202af | minimax_h3_audio_vae_fp32.safetensors | FP32 audio decoding | 605,254,808 bytes 0.61 GB |
| Turbo LoRA | larryvrh/MiniMax-H3-Turbo-Lora 43a74557ac3f6539db8e0f2a959d03feb7a81480 | minimax_h3_turbo_v4_step600_ema.safetensors | v4 step600 EMA strength=1.3，low_vram=false | 779,849,816 bytes 0.78 GB |
| AI upscaling | Real-ESRGAN official release v0.2.1 | RealESRGAN_x2plus.pth | 2× general-purpose image upscaling then cropped to 1920×1080 | 67,061,725 bytes 67.06 MB |
03 / Performance tests
MegaCube performance test results
2026-09-09 to 2026-09-11. The table reports an earlier 6/8-step benchmark. The 12 successful tasks are a snapshot from a later 10-step workflow, so timings from these workflows must be interpreted separately. Client-observed time includes polling at approx. 5-second intervals.
| Configuration | Total client time | Server | Peak memory | Result |
|---|---|---|---|---|
| 640×384 / 6 | 95.84 s | 93.44 s | 87.69 GiB | Passed |
| 832×480 / 6 | 116.37 s | 114.98 s | 90.29 GiB | Passed |
| 832×480 / 8 | 141.66 s | 137.59 s | 91.36 GiB | Passed |
| 1376×768 / 6 | 399.06 s | 396.44 s | 96.81 GiB | Passed |
- Memory usage is calculated as MemTotal − MemAvailable, sampled every 5 seconds. It measures system memory use, not GPU memory allocated to an individual process.
- 1080p is an upscaled output, not native 1080p generation by the model.
- No comparable single-device text-to-video performance data is currently available.
Plan model deployment and application integration
Evaluate device configurations and inference services for your model size, workloads, and concurrency needs.