ComfyUI-MiniMaxH3-FlowDirector Introduction
License: MIT (https://img.shields.io/badge/License-MIT-yellow.svg) ComfyUI (https://img.shields.io/badge/ComfyUI-Custom_Node-blue.svg)
ComfyUI-MiniMaxH3-FlowDirector is a visual timeline director for MiniMax Hailuo H3 in ComfyUI. It helps you build extended videos from shorter, connected scenes, with support for image-guided transitions and synchronized audio.
Rather than generating an entire long video in one operation, the extension works through your timeline in smaller sections called blocks or chunks. Each block continues from the final image of the previous one. This gives you a way to develop a longer sequence without treating every scene as an unrelated clip.
You can use it to:
- Extend a scene over multiple blocks, changing the prompt as the action develops.
- Guide a sequence toward particular images, using them as visual destinations.
- Combine text-driven and image-guided sections in the same timeline.
- Work with audio alongside video, including waveform displays, trimming, and stereo mixing.
- Reduce the memory pressure of long-video generation by processing blocks sequentially instead of all at once.
The author recommends approximately 5–10 seconds per block. You can chain many blocks to build multi-minute sequences.
About video length and memory: The author describes the approach as supporting indefinitely extendable timelines with a low, constant VRAM footprint. This refers to processing one block at a time rather than loading an entire long sequence into a single generation. It is not a guarantee that every workflow will avoid memory errors or other hardware limitations.
How ComfyUI-MiniMaxH3-FlowDirector Works
Think of the timeline as a storyboard in which each panel becomes a short video. The extension connects these panels by handing the final image of one section to the next.
1. Divide your idea into short sections
Instead of asking for one very long video, organize the action into smaller blocks.
For example:
| Block | Creative direction |
|---|---|
| 1 | A traveler walks along a misty forest path. |
| 2 | The traveler approaches a clearing as the light grows warmer. |
| 3 | The camera reveals a sunlit garden beyond the trees. |
This structure lets you change the direction of the story over time without generating the whole sequence in one large operation.
2. Generate and decode one block at a time
Each block is generated and decoded separately. Decoding means turning the model’s internal video representation into visible frames.
Because the extension handles blocks sequentially, increasing the timeline length does not require generating all of its frames simultaneously. This is the main principle behind its reduced VRAM requirements for long sequences.
3. Pass the final frame forward
After a block finishes, its final decoded frame becomes the next block’s opening image through the first_frame input.
For example, if one block ends with the traveler standing beside a tree, the next block starts from that image rather than inventing a completely unrelated opening scene.
This gives neighboring blocks a shared visual handoff. The generated motion and interpretation of your prompts still determine how naturally the sequence develops.
4. Guide transitions toward target images
When you drag an image onto a later block—identified in the original documentation as Block N > 0—the extension treats it as a destination transition frame through last_frame.
A useful way to think about these two controls is:
first_frame: Where the motion begins.last_frame: Where the transition is intended to arrive.
For example, a garden image can serve as a visual destination for a sequence that begins in a forest. The model generates the transition toward that target rather than simply cutting to an unrelated still image.
5. Remove repeated boundary frames
The shared frame between consecutive blocks can otherwise appear twice: once at the end of one block and once at the beginning of the next.
The extension automatically removes redundant boundary frames to avoid a small pause or stutter caused by showing the same frame twice.
6. Send the results to a video-combining node
The director outputs:
images— the generated video frames.audio— the audio output.fps— the playback frame rate.
The recommended workflow connects these outputs to VHS Video Combine to assemble the final video.
ComfyUI-MiniMaxH3-FlowDirector Features
Visual timeline direction
The extension provides a WYSIWYG, or “what you see is what you get,” visual timeline for arranging your sequence.
Instead of thinking only in terms of a single prompt, you can plan a progression of short sections. This is useful when your video needs changing actions, environments, or visual goals.
Sequential generation for longer videos
Each timeline section is sampled and decoded in an isolated block.
Why it matters: Generating a very large number of frames at once can place heavy demands on GPU memory. Sequential generation reduces that pressure by breaking the work into smaller pieces.
Creative starting point: Follow the author’s recommendation of roughly 5–10 seconds per block, then chain additional blocks to extend the sequence.
Automatic last-to-first frame chaining
The final decoded image of one block automatically becomes the starting image of the next.
Why it matters: You do not need to manually export the final frame of every section and reconnect it as the next section’s opening image.
Example: A character’s pose at the end of a walking sequence becomes the starting pose for the following section.
Image-guided destination transitions
Images placed on later blocks act as transition targets.
Why it matters: You can direct the sequence toward a specific composition or visual state rather than relying entirely on text.
Example: Use text to describe a camera moving through a landscape, then provide an image of the destination to guide where the transition should lead.
The author highlights using additional images as destinations as a preferred way to direct longer sequences.
Mixed text and image guidance
The documented workflow supports chaining text-driven sections with image-guided sections and then continuing with further text prompts.
Why it matters: You can reserve image guidance for important visual moments while using text to describe the action between them.
For example:
- Establish a scene with a text prompt.
- Guide the sequence toward a target image.
- Continue the action with another text-driven section.
Automatic seam cleanup
Duplicate frames at block boundaries are removed automatically.
Why it matters: This addresses the brief playback hesitation that can occur when the same handoff frame appears at both the end of one block and the start of the next.
This feature removes repeated boundary frames; it should not be interpreted as a guarantee that every generated transition will be visually perfect.
Integrated audio and motion tracks
The documented timeline features include:
- Audio waveform displays.
- Automatic audio trimming.
- Stereo mixing.
- Native MiniMax audio decoding.
- Integrated audio and motion tracks.
Why it matters: Audio can be handled alongside the video timeline rather than treated only as a separate finishing step. The waveform display also gives you a visual reference for where sound occurs.
ComfyUI-MiniMaxH3-FlowDirector Models and Workflow Connections
The original documentation provides a recommended set of model components and connections, rather than a catalog of alternative model versions.
These components have different roles in the workflow:
| Component | Director input | Role in the documented setup |
|---|---|---|
| FL2VA Model | model | Connects the FL2VA model used by the director. |
| REF2VA Model | model_ref2va | Connects the REF2VA model used by the director. |
| MiniMax Text CLIP | clip | Provides the text-encoding component that turns prompts into information the generation system can use. |
| MiniMax Video VAE | vae | Provides the component used to decode video into visible frames. |
| MiniMax Audio VAE | audio_vae | Provides the component used to decode MiniMax audio. |
| KSamplerSelect | sampler | Supplies the sampling method used during generation. |
| BasicScheduler | sigmas | Supplies the progression of noise levels used during sampling. |
| RandomNoise | noise | Supplies the starting noise used during generation. |
The provided documentation does not specify model download links, model-size comparisons, or detailed quality and performance differences between model variants.
Recommended output connections
Connect the director’s outputs to VHS Video Combine:
MiniMax H3 Flow Director
images ──> VHS Video Combine
audio ──> VHS Video Combine
fps ──> VHS Video Combine
For a concrete reference, open the included example workflow:
MINIMAX_H3_FLOW_DIRECTOR_ULTRA_TURBO.json
Use it to understand how the model components, generation controls, director, and video output fit together.
Troubleshooting ComfyUI-MiniMaxH3-FlowDirector
The following checks are based on the documented workflow and features.
Generation stops with a GPU memory error
Sequential generation reduces the memory demands of long timelines, but it does not eliminate the memory needed for an individual block.
- Check the workflow against the included example.
- Test a short timeline before attempting a multi-minute sequence.
- If your blocks are much longer than the author’s recommendation, try approximately 5–10 seconds per block.
- If the error continues, review the memory requirements of the connected models and the rest of your ComfyUI workflow.
Neighboring sections do not flow naturally
- Confirm that the sections belong to the same director timeline so automatic frame chaining can apply.
- Check whether the next prompt asks for a sudden, major change from the previous scene.
- Try a simpler transition to evaluate the handoff between blocks.
- Where a particular destination matters, add a target image to a later block.
Shared boundary frames provide a connection between sections, but complex changes in scene, subject, or action can still be challenging for video generation.
An image is not behaving like a new opening still
An image placed on a later block is documented as a destination frame, not simply a replacement opening frame.
- Check where you placed the image.
- If it is on Block
N > 0, interpret it as a target for the transition. - Adjust the surrounding prompts so the action leads toward that image.
Playback appears to pause at a block boundary
The extension is designed to remove duplicate boundary frames automatically.
- Check that you are using the director’s assembled frame output.
- Inspect any additional concatenation or frame-processing steps after the director.
- Confirm that the director’s
fpsoutput is connected to the video-combining node. - Compare your connections with the included example workflow.
The exported video has no audio
- Check that MiniMax Audio VAE is connected to
audio_vae. - Check the audio content and arrangement in the timeline.
- Confirm that the director’s
audiooutput is connected to VHS Video Combine, alongsideimagesandfps.
Frequently asked questions
Do I need an image for every block?
No. The author describes sequences that combine text-driven and image-guided sections. Images can be used where you want stronger control over a destination.
What block duration should I start with?
The author recommends approximately 5–10 seconds.
Can I make a video longer than one generation block?
Yes. Chaining blocks is the extension’s central purpose, including building multi-minute sequences.
Does “infinitely long” mean there are no practical limits?
No. It describes an extendable, sequential approach rather than a promise of unlimited runtime or resources. Your system and surrounding workflow still have practical limits.
Does frame chaining guarantee perfectly consistent motion or identity?
It provides a shared starting frame between blocks. It does not by itself guarantee every aspect of the generated content.
Learn More about ComfyUI-MiniMaxH3-FlowDirector
-
Extension repository
Explore the project files and its documentation. -
Included example workflow
A practical reference for understanding the intended node connections. -
ComfyUI
Learn more about the node-based application in which the extension runs. -
MIT License
Review the extension’s licensing terms.
