VDN H3 ComfyUI Text To Video | Faster MiniMax H3 VDN 8-Step Workflow wihout LoRA

VDN H3 MiniMax ComfyUI Text To Video#

VDN H3 ComfyUI Text To Video turns natural language prompts into short, cinematic videos with synchronized audio. It combines the MiniMax H3 family with Video DeltaNet hybrid attention and 8-step sampling to generate video and audio from a single prompt. It is well suited to storyboards, product clips, and stylized scenes such as cinematic fantasy, rally driving, and armored sword sparring.

The graph includes two independently selectable branches. The VDN-H3 branch applies the Video DeltaNet patch and was used to produce the included examples. The FastVideo branch is provided for comparing output and generation time with the VDN path. Both branches render MP4 with audio at your chosen frame rate.

Built on open components: VDN for MiniMax H3 GitHub and Hugging Face, VDN-H3 ComfyUI nodes by Saganaki22 GitHub, and base MiniMax H3 weights and VAEs Hugging Face.

Key models in Comfyui VDN H3 ComfyUI Text To Video workflow#

• MiniMax-H3 diffusion model (UNet). Drives the denoising process that turns noise into coherent spatiotemporal video and audio latents. Weights are provided in the MiniMax H3 bundle on Hugging Face.
• Qwen3-VL 32B MiniMax-H3 text encoder. Converts your prompt into conditioning for video and audio generation, tuned for the H3 family. Included in the MiniMax H3 release on Hugging Face.
• MiniMax-H3 Video VAE. Decodes video latents to RGB frames. Available in the VAEs section of Hugging Face.
• MiniMax-H3 Audio VAE. Reconstructs waveform audio from audio latents. Also provided on Hugging Face.
• VDN-H3 patch weights. Video DeltaNet applies hybrid attention to MiniMax H3. Get the checkpoints from Hugging Face.
• FastVideo UNet variant. The comparison branch uses minimax_h3_fastvideo_vsa_datafree_1300step_4step_int8_convrot.safetensors from Kijai/MiniMax-H3-experimental.

How to use Comfyui VDN H3 MiniMax ComfyUI Text To Video workflow#

Start by setting your prompt, resolution, duration, and frame rate. Select the VDN-H3 branch or the FastVideo branch for an individual run. Queue All may execute both enabled branches and produce two MP4s; separate graph branches do not guarantee simultaneous GPU execution. Both branches are configured for 8 sampling steps.

User Input#

Use the (input:prompt) Text Prompt panel to describe both visuals and sound in plain language. For audio, add a final line beginning with Audio: to request specific Foley or ambience, for example “Audio: rhythmic hoofbeats, wind, distant roar.” Control visual size with (input:Resolution) Megapixels which feeds the ResolutionSelector to produce a width and height that are compute-friendly. Set (input:Duration) Seconds and (input:FPS) to define how long and smooth the clip should feel. The workflow converts duration and FPS to a valid frame count and quietly snaps it to a sequence length that plays well with H3’s internals.

Models#

The CLIPLoader provides the MiniMax-tuned text encoder used for conditioning. Two VAELoader nodes bring in the MiniMax-H3 Video VAE and Audio VAE so decoded frames and sound match the model’s latent space. These loaders are shared by both branches to ensure apples-to-apples A/B results.

Conditioning#

MiniMaxH3ImageToVideo (#219) turns your prompt plus size and length into the initial video-and-audio latent along with the positive conditioning. Optional image inputs can be used to lock the first or last frame when you need a specific entrance or exit pose. This stage is where your narrative decisions matter most: camera motion, subject action, lighting, and the Audio: line all become guidance for the sampler downstream.

VDN-H3 patch#

The VDN branch applies the Video DeltaNet upgrade with ApplyVDNH3Advanced (#6126). It applies the selected hybrid attention patch. MiniMaxChunkFeedForward and ModelPatchTorchSettings configure memory and execution settings. MiniMaxH3SigmaShift adjusts the noise schedule for video and audio before sampling begins.

Sampling (VDN-H3)#

VDN sampling starts from RandomNoise (#6141), merges guidance with BasicGuider (#6134), chooses the solver in KSamplerSelect (#6137), and schedules a compact number of diffusion steps in BasicScheduler (#6136). SamplerCustomAdvanced (#6135) performs the denoising over the full sequence to produce refined video and audio latents. Keep the seed fixed for repeatability or change it for fresh variations with the same prompt.

Decoding and create video (VDN-H3)#

VAEDecode (#6146) reconstructs frames while VAEDecodeAudio (#6145) rebuilds the soundtrack. VHS_VideoCombine (#6170) muxes frames and audio into an MP4, honoring the frame rate you set earlier and saving a preview-ready file. The VDN output is saved with a distinct filename prefix so it is easy to compare against the FastVideo result.

FastVideo comparison branch#

The FastVideo path loads a speed-tuned MiniMax H3 UNet and applies a lightweight attention patch before sampling. It reuses your prompt, size, and frame count so you can directly compare timing and look. Sampling mirrors the structure of the VDN branch and is also configured for 8 steps. Frames and audio are decoded and combined with VHS_VideoCombine (#6104) to produce a second MP4. Use these side-by-side outputs to decide which branch best fits your scene.

Key nodes in Comfyui VDN H3 ComfyUI Text To Video workflow#

• ApplyVDNH3Advanced (#6126). Enables Video DeltaNet on MiniMax H3 using the selected checkpoint from the VDN-H3 release. Adjust only when you want to change the VDN checkpoint or switch attention backends; keep the hybrid attention defaults for general use. Reference implementation: Saganaki22/ComfyUI-VDN-H3.
• MiniMaxH3ImageToVideo (#219). Central control point for story and timing, accepting the text prompt, computed width and height, and the derived frame count. For audio, append a single Audio: line to the prompt to steer Foley and ambience while avoiding speech unless explicitly requested.
• MiniMaxH3SigmaShift (#6131, #6007). Rebalances the sigma schedule for video and audio streams, which can help reduce jitter and preserve detail at low step counts. Consider small adjustments only if you change model variants or notice motion instability.
• VHS_VideoCombine (#6170, #6104). Muxes decoded frames and audio into MP4 with your chosen frame rate and filename prefix. Useful options include trimming to audio length and selecting the desired H.264 pixel format. Project page: ComfyUI-VideoHelperSuite.

Optional extras#

• Prompting tips: lead with a single sentence that sets subject, action, and camera move, then add look-and-feel cues and one Audio: line for foley.
• For longer clips, prefer modest megapixels over high resolution to keep motion stable and memory usage in check.
• Fix the seed in RandomNoise when you want consistent iterations; change it to explore variations.
• Compare the two branches on your own scene to assess output quality and generation time.
• If you hit memory limits, reduce megapixels or duration first; the chunked feed-forward already helps with longer sequences.

Acknowledgements#

This workflow implements and builds upon the following works and resources. We gratefully acknowledge OpenVDN for the vdn-minimax-h3 model, weights, and repository, Saganaki22 for the ComfyUI-VDN-H3 nodes, and Benji (AI Future Tech) for the VDN-H3 workflow for their contributions and maintenance. For authoritative details, please refer to the original documentation and repositories linked below.

Resources#

• OpenVDN/vdn-minimax-h3
  • GitHub: OpenVDN/vdn-minimax-h3
  • Hugging Face: OpenVDN/vdn-minimax-h3
• Saganaki22/ComfyUI-VDN-H3
  • GitHub: Saganaki22/ComfyUI-VDN-H3
• Benji (AI Future Tech)/VDN-H3 Workflow
  • Docs / Release Notes: AI Future Tech Patreon post

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