H3 Infinite Video Continuation and automatic Stitching with Fl2Va Model

With these Workflows (you'll also need to download my Custom Nodes via the ComfyUI Manager) you can create MiniMax H3 videos of pretty much infinite length by chaining shorter clips together.

I specifically built the Nodepack around the MiniMax H3 FL2VA model.

The main idea is to keep the visual quality and control of FL2VA while allowing you to continue a video over and over again without having to rely on only the last rendered frame.

Starting with v1.4, continuation now uses ComfyUI's native Masked AV Latent Continuation. Instead of asking H3 to recreate the beginning of the next clip from guidance, a safe part of the previous video + audio latent is copied directly into the next generation and protected from denoising.

In my testing this significantly improved continuity between clips and, especially, fixed the small brightness / color changes that could sometimes happen with the older continuation method.

v1.4 also supports flexible H3 conditioning and multiple Qwen Reference Images, so you can use the workflows with different combinations of First Frames, Last Frames and References.

Included Workflows

The following Workflows are included:

1. Start Video

Herrgotts_H3_Infinite_v1.4_01_Start.json

Use this Workflow to create the first Clip of your video.

v1.4 supports flexible H3 conditioning, so depending on what you connect you can use it similar to:

• T2VA

• I2VA

• L2VA

• FL2VA

You can provide:

• Prompt

• First Frame

• Last Frame

• Up to 9 Qwen Reference Images

First and Last Frames are optional, although for most projects I'd recommend using both.

After generation the Workflow automatically saves the complete Video + Audio Latent that will be used for the next Clip.

2. Continue Video into infinity

Herrgotts_H3_Infinite_v1.4_02_Continue.json

Use this Workflow for every Clip after the first one.

Load the saved AV Latent from your previous Clip, enter your new Prompt and optionally provide a new Last Frame and Qwen Reference Images.

The Workflow automatically analyzes the previous Clip, removes the typical unusable FL2VA tail and selects a safe latent context before it.

That context is then inserted directly into the next generation using the new Native Masked AV continuation method.

After the new Clip finished generating, the new complete AV Latent is saved again and can be used for Clip 3, Clip 4, Clip 5 and so on.

Just make sure you're loading and saving the correct Clip Index.

Using the same Save Index as an existing latent will overwrite that latent, which can actually be useful if you intentionally want to regenerate an individual Clip.

Net New Content

Net New Content is now the default Duration Mode.

This means that if you select roughly 5 seconds, the Workflow tries to give you roughly 5 seconds of new usable content, instead of counting the 39-frame continuation context as part of those 5 seconds.

This requires H3 to generate a somewhat longer latent and therefore increases sampling time, but I think it makes the Workflow much more intuitive to use.

You can still switch back to Total Generation if you prefer the old behavior.

3. 3-Clip Showcase / Auto Stitch

Herrgotts_H3_Infinite_v1.4_03_3Clip_Showcase_AutoStitch.json

This Workflow creates:

Start Video → Continue Video → Continue Video → Final Stitch

all inside one graph.

You'll mainly have to provide:

• Starting Frame

• Prompts for the individual Clips

• Optional Last Frames

• Optional Qwen Reference Images

The graph is intentionally structured so you can duplicate the final continuation block for Clip 4, Clip 5 and beyond.

For very long projects I'd still recommend generating individual Clips separately using Workflows 1 and 2, since it's much easier to regenerate only one Clip if something goes wrong.

4. Stitch Saved Chain

Herrgotts_H3_Infinite_v1.4_04_Stitch_Saved_Chain.json

Use this Workflow to combine a longer chain of Clips that you've generated separately.

It loads and processes one saved AV Latent at a time and writes the final sequence directly to MP4.

Because the entire video doesn't need to exist decoded in memory at the same time, memory usage during final stitching stays roughly tied to the size of an individual Clip instead of continuously increasing with the total video length.

This makes it much more practical to stitch very long projects without eventually running into a length-dependent OOM during the final stitch.

Installation

To install the Nodepack, open one of the Workflows and use Missing Custom Nodes, or search for:

Herrgotts-H3-Infinite-Continuation-Suite

or simply:

Herrgott

inside ComfyUI Manager.

Important

v1.4 uses ComfyUI's relatively new native H3 Masked AV support, so you'll need a recent version of ComfyUI.

If the required H3 mask functionality isn't available in your installation, the Node will tell you to update ComfyUI instead of silently falling back to another continuation method.

After updating ComfyUI I'd recommend doing a full restart before using the v1.4 workflows.

Main Features

• Easy to use Workflows with an intentionally simple structure so you can rearrange, duplicate and modify Nodes easily.

• Native Masked Video + Audio Latent Continuation using ComfyUI's H3 denoise masks.

• No decode / re-encode handover between Clips.

• Freeze- and brightness-aware Auto Handover.

• Safe pre-freeze latent context selection instead of blindly using the final frames of the previous Clip.

• Recommended 39-frame Native Masked AV context for stable continuation.

• Independent Audio Tail Carryover to improve dialogue continuity across Clip boundaries.

• Net New Content duration mode so continuation context doesn't have to reduce the amount of new visible content you requested.

• Total Generation mode still available if you prefer shorter / faster generations.

• Flexible T2VA / I2VA / L2VA / FL2VA-style starting conditions.

• Up to 9 Qwen Reference Images.

• Automatic Picture Map for Qwen prompt references.

• Repeated Last Frame anchors for visual control and regular quality resets.

• Freeze detection for the typical static FL2VA tail.

• Video + Audio aware stitching.

• Full, Stitch Ready and Final Clip output modes.

• Save / Load complete AV Latents including stitch and handover metadata.

• Memory-bounded Saved Chain Stitcher for very long projects.

• Older v1.3 workflows remain compatible through the included Legacy Nodes.

Why generate shorter Clips instead of one very long H3 generation?

Apart from having much more control, shorter individual generations can also be significantly more practical than asking H3 to create one very long Clip in a single pass.

For example, instead of generating one 15-second Clip you could generate:

5s + 5s + 5s

If something goes wrong in the third section, you only have to regenerate that individual Clip instead of starting the entire sequence again.

Using regular Last Frames also gives H3 a new visual target every few seconds, which in my experience helps a lot with maintaining quality during longer sequences.

If you use the Nodepack or the Workflows, it'd be awesome if you could share your favorite creations!

Have fun Prompting. :)