Sign In
homemodelsimagesvideos3D Modelsarticlescomicsbountieschallengesupdatesshop

H3 Infinite Video Continuation and automatic Stitching with Fl2Va Model - v1.3 Showcase

HR

HRGTMRGT

Loading Images

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 multiple shorter FL2VA clips together.

The Nodepack was specifically built around the MiniMax H3 FL2VA / First Frame Last Frame Model.

The main reason is pretty simple: In my testing, FL2VA gives me noticeably better visual quality than Ref2VA, while Ref2VA normally gives you more flexibility and control over longer sequences.

The idea behind these Nodes is therefore to get much of the control of a Ref2VA workflow while keeping the higher FL2VA quality.

Instead of generating one very long H3 clip, you generate several shorter clips and continue the previous video + audio latent directly into the next generation.

This has a few advantages:

  • Motion and audio can continue naturally between Clips.

  • You can use a new Last Frame for every Clip as a visual target and quality reset.

  • Shorter H3 generations can be considerably faster than generating the same total duration as one very long Clip.

  • If one part of the video doesn't turn out well, you only need to regenerate that individual Clip.

  • Saved Clips can later be stitched without decoding the complete video into memory at once, so memory usage stays roughly tied to one Clip instead of increasing with total video length.

I've attached a longer example video generated using this approach.

What's new in v1.3+

The Workflows are much more flexible now.

The Start Workflow supports:

T2VA: Prompt only
I2VA: First Frame only
L2VA: Last Frame only
FL2VA: First + Last Frame

You can also add multiple Qwen Reference Images and use them for things like Character Identity, Clothing or other visual details.

The Nodes include a Picture Map that tells you exactly which connected image corresponds to Picture 1, Picture 2, etc., so you can reference them directly in your prompt.

For the Continue Workflow, the previous AV Latent provides the starting motion / scene context while the optional Last Frame is used as the new visual endpoint.

Included Workflows

1. Start Video — Herrgotts_H3_Infinite_v1.3.1_01_Start.json

Use this Workflow to create your first Clip.

You'll need your normal H3 Models / VAEs, a Prompt, Resolution and Clip Length.

The following inputs are optional:

  • First Frame

  • Last Frame

  • Qwen Reference Images

For the classic Infinite Continuation setup I'd still recommend using First + Last Frame, since the Last Frame also serves as a quality reset and gives H3 a clear target for the end of the Clip.

The Picture Map automatically shows how connected images can be addressed in your Prompt.

For example:

First Frame = Picture 1
Last Frame = Picture 2
Qwen Reference 1 = Picture 3
Qwen Reference 2 = Picture 4

After generation, the Workflow automatically saves the complete Video + Audio Latent that can be used to continue the sequence.

2. Continue Video into infinity — Herrgotts_H3_Infinite_v1.3.1_02_Continue.json

Use this Workflow for Clip 2 and every Clip after that.

Load the saved Latent from the previous Clip, enter the Prompt for your next segment and optionally provide:

  • a new Last Frame

  • one or more Qwen Reference Images

The important part is that the previous Clip is not simply rendered and fed back as an image.

Its Video + Audio Latent is used directly as temporal context for the next H3 generation.

The new Last Frame acts as the endpoint / quality reset for the new Clip.

In v1.3.1 the Last Frame is intentionally not a Qwen Picture during Continue. This keeps the beginning of the new Clip anchored to the previous Latent instead of letting the future End Frame influence the starting composition too strongly.

So in Continue:

Previous Latent = not a Picture
Last Frame = not a Picture
Qwen Reference 1 = Picture 1
Qwen Reference 2 = Picture 2

The Picture Map inside the Node always shows the actual mapping.

Just make sure that the correct Latent Clip Index is loaded and saved according to your Output Folder.

Example:

Clip 1 → Save Latent 1
Clip 2 → Load Latent 1 / Save Latent 2
Clip 3 → Load Latent 2 / Save Latent 3

Using an existing Save Index will overwrite that Latent, which can also be useful if you want to regenerate a specific Clip.

3. 3-Clip Showcase / Auto Stitch — Herrgotts_H3_Infinite_v1.3.1_03_3Clip_Showcase_AutoStitch.json

This is probably the easiest Workflow if you just want to understand how the whole system works.

It runs:

Start Video → Continue Video → Continue Video → Automatic Stitching

in one queue and creates one final video.

The graph is intentionally structured so another continuation block can easily be duplicated for Clip 4, Clip 5 and so on.

You'll mainly have to provide your Prompts, optional Keyframes / References and the initial settings.

For very long projects I'd still recommend generating Clips separately using Workflows 1 + 2.

4. Stitch Saved Chain — Herrgotts_H3_Infinite_v1.3.1_04_Stitch_Saved_Chain.json

Use this Workflow once you're happy with your individually generated Clips.

It combines saved Latents like:

clip_00001
clip_00002
clip_00003
clip_00004
...

into one final MP4.

The Stitcher decodes one saved AV Latent at a time and writes the result progressively instead of loading the entire decoded video into memory.

This means memory usage does not continuously increase with every additional Clip, so the total length of the final chain itself does not create the usual length-dependent OOM limitation.

Installation

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

Herrgotts-H3-Infinite-Continuation-Suite

in ComfyUI Manager.

The included Workflows use the normal MiniMax H3 FL2VA setup.

Main Features

  • Flexible T2VA / I2VA / L2VA / FL2VA generation.

  • Multiple Qwen Reference Images with automatic Picture Mapping.

  • Direct Video + Audio Latent Continuation without decode/re-encode handover.

  • Repeated Last Frame Keyframe Anchors for visual control and quality resets.

  • Auto Handover that detects the frozen FL2VA tail instead of simply using a fixed trim.

  • Phase-aligned Context to avoid invalid H3 latent cut positions.

  • No-Lock Fallback when no final freeze can be detected.

  • Safe Tail Bridge to reduce small visual seams between Clips.

  • Short Video Crossfade + separate Audio De-Click transition.

  • Full, Stitch Ready and Final Clip output modes.

  • Save / Load complete Video + Audio Latents with Stitch Metadata.

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

  • Workflows are intentionally kept relatively simple so you can rearrange, duplicate or modify individual sections yourself.

The main goal is basically:

FL2VA Quality + Ref2VA-like Control + Video/Audio Continuity + repeated Quality Resets + faster long-form generation.

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

Have fun Prompting. :)

Comments