Skip to main content
FastVideo FastH3 is a DMD2-distilled MiniMax H3 checkpoint from the FastVideo team that generates video with synchronized native audio in 8 sampling steps instead of the base model’s full schedule. It reuses the H3 text encoder, video VAE, and audio VAE, so output matches the base model’s setup while sampling finishes in a fraction of the steps. The FastH3 preview work was developed in collaboration with Nuva Lab and the NVIDIA FastGen team; read the FastH3 announcement for distillation details. ComfyUI ships with the 8-step V2 checkpoint, repacked by Comfy-Org in the FastVideo-FastH3-Comfy repository. FastH3 trades some motion and audio fidelity for speed. Use it for drafts, iteration, and fast turnaround; use the base MiniMax H3 workflows when maximum quality matters.
The distilled checkpoint supports text-to-video and first/last-frame image-to-video only. Ref2VA (multi-reference conditioning) was not distilled; use the base MiniMax H3 workflows for reference-based generation.

​
Requirements

  • ComfyUI 0.36.0 or later
  • The MiniMax H3 text encoder and VAEs (shared with the base model, listed below)

​
FastH3 Text to Video

Generate synchronized video and audio from a text prompt in 8 steps.

Download Workflow

Download JSON or search “FastVideo FastH3: Text to Video” in Template Library

​
Prompting tips

  1. Describe the whole scene: State the overall scene first (location, character, what is happening), then break it into timed shots
  2. Shots, camera, and audio: Describe the shots, camera moves, and the accompanying audio (dialogue, SFX, music) in one prompt block
  3. Resolution: H3’s native canvas is a 768px short edge, capped at 768x1344, and resolutions are rounded to a multiple of 32
  4. Duration: The duration input snaps to the model’s 17-frame-per-block (17k+5) grid at 24fps
  5. Steps are fixed at 8: The distilled checkpoint is trained for exactly 8 steps; changing the step count in the scheduler degrades quality
For prompt writing resources, see the prompt guide.

​
FastH3 Image to Video

Animate a still image with synchronized audio, with optional first/last-frame control.

Download Workflow

Download JSON or search “FastVideo FastH3: Image to Video” in Template Library
Input material Upload this file to the matching LoadImage node, or use your own image:

red_line_barrier.png

Example first frame for the Image to Video workflow

​
First/last-frame mode

Connect images to first_frame and/or last_frame on the MiniMaxH3ImageToVideo node to generate the motion between two keyframes. Leave both disconnected for text-to-video.

​
Prompting tips

  1. Describe the motion and audio: With an image providing the look, the prompt should focus on motion, camera, and the accompanying audio
  2. Duration: The duration input snaps to the model’s 17-frame-per-block (17k+5) grid at 24fps
  3. Steps are fixed at 8: Keep the scheduler at 8 steps; the checkpoint is distilled for that schedule

​
Model downloads

The diffusion model comes from the FastVideo repository; the text encoder and VAEs are shared with the base MiniMax H3 model.

Diffusion Model: fastvideo_fasth3_8step_v2_pruned_int8_convrot

Place in ComfyUI/models/diffusion_models/

Text Encoder: qwen3vl_32b_minimax_h3_nvfp4_awq

Place in ComfyUI/models/text_encoders/

VAE: minimax_h3_video_vae_fp16

Place in ComfyUI/models/vae/

VAE: minimax_h3_audio_vae_fp32

Place in ComfyUI/models/vae/

​
Model storage

​
Workflow structure notes

Both FastH3 workflows share these nodes with the base H3 setup, with distillation-specific settings:
  • BlockSparseAttention: Runs Video Sparse Attention (VSA) with keep_percent 10 starting at 20% of the schedule, reducing attention cost while preserving quality
  • MiniMaxH3SigmaShift: Applies the H3 sigma shift (10 for video, 3 for audio) that matches the distilled schedule
  • ComfyMathExpression: Converts the duration input into a valid frame length on the 17k+5 grid at 24fps
  • Sampler: res_multistep sampler with the simple scheduler at 8 steps