This is a customized version of the DaSiWa MiniMax H3 workflow:
It keeps the original T2VA, I2VA, FL2VA and REF2VA modes, while adding quality-of-life tools such as faster previews, optional acceleration, clip continuation, seed management, latent upscale, automatic aspect-ratio handling, and color matching.
## Added features
- Spectrum acceleration
Optional acceleration for H3 denoising. It predicts steps based on previous ones to reduce generation time.
Disable it for the final version if you notice weaker motion, face detail, or audio/video consistency.
Repository:
https://github.com/xmarre/ComfyUI-Spectrum-MiniMax-H3
- Continue From Clip (NativeAddGuide)
Lets MiniMax H3 use a previous clip as audio/video context, creating more natural transitions while preserving character, scene, motion direction, ambience, and audio continuity.
Connect one video to the Clip input and enable Continue From Clip (NativeAddGuide).
Recommended AddGuide Context Frames:
- 16: lighter and faster continuation
- 22: recommended balance
- 32+: stronger continuity, but slower generation
This uses the native MiniMaxH3AddGuide node included in recent ComfyUI master builds. Update ComfyUI from the master branch if the node is missing.
- Masked AV continuation (Recommended)
An alternative continuation method that preserves a selected prefix of a previous video/audio clip and generates the remaining section. In my tests, it can be more precise than the native guide option, but both methods are worth testing for your scene.
Recommended settings:
- Use a source video with properly synchronized audio and video.
- Keep the source clip close to the desired aspect ratio when possible.
Repository:
https://github.com/seitanism/ComfyUI-H3-Motion-Context-MultiRef
- Automatic clip-aspect handling and color matching
When Continue From Clip (NativeAddGuide) or Masked AV continuation is enabled, the workflow derives an output resolution close to the source clip’s aspect ratio while respecting the selected megapixel preset and MiniMax H3’s size constraints.
It also automatically matches the generated clip’s color to a reference frame from the source clip, sampled shortly before its end. This helps reduce visible changes in temperature, saturation, brightness, and overall color between consecutive clips.
Color Match Strength controls how strongly the generated clip is adjusted:
- 0.10: subtle correction; preserves more of the new generation’s original color
- 0.20: recommended balance
- 0.30–0.40: stronger match; useful when the generated clip looks noticeably warmer, yellower, cooler, or more saturated than the source clip
Color matching is enabled automatically only during continuation modes.
- LBH 3D Video Latent Upscaler
Optional high-resolution refinement. It first generates at a lower resolution, upscales the video latent in latent space, then refines it at the target resolution.
Unlike a normal upscale such as RTX, this does not upscale the already decoded final video. Instead, it upscales the latent before final decoding, allowing MiniMax H3 to refine the result at the larger resolution.
The workflow uses the selected number of steps minus 4 for the lower-resolution generation. The last 4 steps are used after latent upscaling at the final target resolution.
This can reduce generation time if you start from a very small resolution, but using a resolution that is too low can reduce movement quality and fine detail. Make your own tests.
Node repository:
https://github.com/LBH-123-AI/Comfyui_Minimax_h3_latent_Upscaler
Model download:
https://huggingface.co/LBH-123-AI/Minimax_h3_latent_Upscaler
- Fast video preview
Uses KJNodes ModelPreviewOverride for quick previews without waiting for the complete standard decode/export path.
Repository:
https://github.com/kijai/ComfyUI-KJNodes
Model:
https://huggingface.co/Kijai/MiniMax-H3-TAE
- rgthree Comfy Seed
Makes seeds easier to randomize, lock, copy, and reuse when comparing generations.
Repository:
https://github.com/rgthree/rgthree-comfy
- External prompt textbox
A simple external text input is included for storing or pasting a complete prompt. It can be used instead of filling every field in the MiniMax H3 Director prompt builder.
Description
FAQ
Comments (10)
u need the https://github.com/madebyollin/taehv/blob/62f7591f59dfbb4c3c02b7a621d180a9eeaba26c/safetensors/taeh3.safetensors
for the preview node to get a flawless preview. put the 21mb "taeh3" file into the vae_approx folder and select it in the preview node(tiny vae).
thx, I forgot to leave it selected!
It has Ref also! I can't wait to try.
EDIT: Be sure to update using "C:\ComfyUI_windows_portable\update\update_comfyui.bat" or wherevers you have it.
Works amazing <3
getting a SamplerCustomAdvanced failed when trying to use continue from clip, but t2va and all the rest work just fine
Have you pulled the latest changes from the Comfy master branch?
yes, the latest was throwing errors on any fl2va run, as well as not letting me run nativeaddguide. Roll back to 33.1 fixed the fl2va but the nodes for nativeaddguide are not updated anymore. ill wait for the rollout i guess
Sorry for asking, but I have the prompts field greyed out and says "External prompt detected — builder fields are disable", how do I disable this external prompt so I can use the "original" prompt builder from Dasiwa prompts fields?
You can turn on the normal builder fields again by disconnecting the wire in the external_prompt input in the H3 director node
@NegativeMeta You`re the best , that worked ! thanks!
Anyway to make it faster?