MiniMax H3 ComfyUI Text-to-Video 4-step Turbo

MiniMax H3 ComfyUI Text-to-Video 4-step Turbo: prompt-to-clip with native audio in one pass#

This workflow turns a single cinematic prompt into a short video with jointly generated stereo audio using MiniMax H3. It applies a Turbo 4-step acceleration LoRA so you can iterate on dialogue, singing, and ambient scenes rapidly while keeping motion and sound tightly coupled.

Built for creators who want prompt-only drafts, the MiniMax H3 ComfyUI Text-to-Video 4-step Turbo workflow produces clean speech or vocals, cohesive environmental sound, and consistent cinematography without manual audio assembly. The graph handles length alignment, latent sampling, decoding, and muxing automatically, so you can focus on storytelling and performance cues.

Key models in Comfyui MiniMax H3 ComfyUI Text-to-Video 4-step Turbo workflow#

• MiniMax H3 diffusion model (FL2VA UNet). Core audiovisual generator that learns a shared latent for picture and sound, enabling synchronized motion and audio from the same sampling pass. Weights: Comfy-Org/MiniMax-H3.
• Qwen3-VL 32B text encoder for H3 (AWQ/NVFP4 variant). Encodes rich, multi-sentence prompts into conditioning that drives scene layout, performance, and audio events. Packaged in Comfy-Org/MiniMax-H3.
• MiniMax H3 Video VAE. Decodes the sampled video latent into frames with filmic tone and detail. File: minimax_h3_video_vae_fp16.safetensors.
• MiniMax H3 Audio VAE. Decodes the shared latent into time-aligned stereo audio for speech, singing, and ambience. File: minimax_h3_audio_vae_fp32.safetensors.
• MiniMax H3 Turbo 4-step LoRA. Applies the acceleration recipe that preserves scene fidelity while enabling very fast sampling. File: minimax_h3_turbo_4step_pruned_comfyui.safetensors.

How to use Comfyui MiniMax H3 ComfyUI Text-to-Video 4-step Turbo workflow#

At a high level, your prompt is encoded, a Turbo-augmented H3 model samples a single audiovisual latent, and the video and audio VAEs decode that latent before the clip is muxed to a playable file.

1) Write the cinematic prompt and set duration#

Use the easy positive (#147) text field to describe look, action, and performance. To guide voice or music, include clear cues like “Audio: middle‑aged male voice delivers the line... rain ambience... low thriller underscore.” The Float (duration) controller (#133) sets clip length in seconds. A math node converts your duration to frames and quietly aligns the count to H3’s required cadence, so you do not need to manage frame math.

2) Load H3, text encoder, VAEs, and Turbo LoRA#

UNETLoader (#127) loads the MiniMax H3 diffusion model, and CLIPLoader (#128) loads the Qwen3‑VL text encoder tailored for H3. Two VAELoader nodes bring in the video VAE (#119) and the audio VAE (#120). LoraLoaderModelOnly (#153) applies the MiniMax H3 Turbo 4-step LoRA to the model so the scheduler and sampler operate in a fast regime without losing audiovisual coherence.

3) Build conditioning and the initial audiovisual latent#

MiniMaxH3ImageToVideo (#131) turns your prompt into positive conditioning and prepares an initial latent aligned to your chosen width, height, and frame count. Optional first_frame and last_frame inputs exist for anchoring a starting or ending image if you choose to extend the graph later. The node outputs the conditioning and latent that downstream sampling will refine.

4) Sample with Turbo#

BasicGuider (#126) pairs the model with your conditioning, KSamplerSelect (#123) picks the sampling algorithm, and BasicScheduler (#124) sets the step schedule that works with the Turbo LoRA. RandomNoise (#129) seeds the generation for reproducibility. SamplerCustomAdvanced (#125) performs the denoising pass to produce the final shared latent that encodes both video and audio.

5) Decode and export the final clip#

VAEDecode (#122) reconstructs the image frames from the latent while VAEDecodeAudio (#121) reconstructs synchronized stereo audio. CreateVideo (#130) muxes frames and audio into a single stream, then SaveVideo (#92) writes the file to disk using the filename prefix you set.

Key nodes in Comfyui MiniMax H3 ComfyUI Text-to-Video 4-step Turbo workflow#

MiniMaxH3ImageToVideo (#131)#

Produces the positive conditioning and initializes the audiovisual latent from your prompt, resolution, and length. Adjusting width, height, or length changes both motion scale and how much on-screen action can fit. If you later extend the graph, first_frame and last_frame let you lock continuity across shots.

LoraLoaderModelOnly (#153)#

Applies the Turbo 4-step LoRA to the loaded H3 model. Keep the scheduler in a low-step regime to benefit from Turbo; substantially raising steps shifts the look away from the intended fast iteration behavior. This LoRA is purpose-built for MiniMax H3 and lives alongside the main weights in the H3 repository.

BasicScheduler (#124)#

Controls the denoising schedule that the sampler follows. Use it to balance speed and fidelity while staying compatible with the Turbo LoRA. If you change the sampler algorithm, revisit the schedule choice to maintain stable motion and clean audio.

SamplerCustomAdvanced (#125)#

Runs the actual denoising given noise, guider, sampler, sigmas, and the initial latent. It is the final arbiter of texture, timing, and audiovisual sharpness. Use the same seed when comparing prompt tweaks so you can attribute differences to text changes rather than noise.

PathchSageAttentionKJ (#150)#

Applies an attention optimization before LoRA injection to improve throughput on large resolutions. Provided by KJNodes, which offers performance helpers for ComfyUI. Repository: ComfyUI-KJNodes.

CreateVideo (#130)#

Combines decoded frames with the decoded audio to a synchronized clip. You can change frame rate or bit depth here if you need to match delivery specs. The node keeps audio-video sync derived from the shared latent.

SaveVideo (#92)#

Writes the rendered clip to disk with your chosen filename prefix, container, and codec. Use consistent naming to track iterations for the same scene and seed.

Optional extras#

• Prompting for speech: put the exact line after “Audio:” and describe voice age, gender, tone, and pace. For singing, specify style and tempo. For ambience, list sources, intensity, and whether you want underscore only.
• Resolution and duration tradeoffs: higher pixel counts and longer clips cost more sampling time. If you need many takes, start smaller, then upscale or lengthen once timing and delivery feel right.
• Reproducibility: keep the same seed when only tweaking text so you can A/B small prompt edits.
• Length alignment is automatic: the graph converts seconds to frames and aligns the count to H3’s internal cadence for stable audiovisual synthesis.
• Continuity options: the model node exposes optional first_frame and last_frame. If you extend the workflow later, connect images there to match shot starts or ends across edits.
• Safety: avoid copyrighted characters, real-person likenesses without consent, and on-screen logos or subtitles in your prompt.

Acknowledgements#

This workflow implements and builds upon the following works and resources. We gratefully acknowledge MiniMax for the MiniMax H3 multimodal generation model and announcement, Comfy.org for the ComfyUI MiniMax H3 workflow tutorial, and Comfy-Org for the MiniMax-H3 model weights for their contributions and maintenance. For authoritative details, please refer to the original documentation and repositories linked below.

Resources#

• MiniMax/MiniMax H3: An Open Model Breaking the Boundaries Between Tasks and Modalities
  • GitHub: MiniMax-AI/MiniMax-H3
  • Hugging Face: MiniMaxAI/MiniMax-H3
  • Docs / Release Notes: MiniMax H3 official announcement
• Comfy.org/MiniMax H3: ComfyUI Workflow Examples
  • GitHub: Comfy-Org/docs
  • Docs / Release Notes: Comfy.org MiniMax H3 tutorial
• Comfy-Org/MiniMax-H3
  • Hugging Face: Comfy-Org/MiniMax-H3

Note: Use of the referenced models, datasets, and code is subject to the respective licenses and terms provided by their authors and maintainers.