Updated v1.6 Yogurt LTXDirector_Extender Workflow (Inspired by WAN SVI, notoriously missing for LTX)

What Changed

1. Guide attention accounting fix

The LTX Director guide node now appends the guide attention metadata that ComfyUI expects for each inserted guide latent. This resolves the failure:

ValueError: guide pre_filter_counts != keyframe grid mask length

2. Overlap-aware guide timing

Extension latents can start with preserved overlap/reference frames from the previous generation pass. The timeline guide positions are extension-relative, so the guide node now detects that front prefix and shifts guide insert positions forward into the newly generated region.

3. Overlap-aware PromptRelay timing

The PromptRelay timing logic now detects when the runtime latent is longer than the schedule-configured latent. When that happens, it treats the extra leading frames as overlap/reference context and shifts the local prompt schedule forward.

That shift is applied to:

• the normal video cross-attention path

• the scaled attention path used by LTX audio conditioning

This matters for extension passes where video and audio conditioning must align to the new section instead of the preserved overlap prefix.

4. Transition and seam fixes for the extender path

The extender workflow now includes a validated transition path for continuing rendered clips without the hard one-frame cut that earlier workflow variants could fall into.

The working v1.6 workflow uses direct embedded-subgraph inputs for overlap and seam controls instead of relying on the older virtual SetNode/GetNode bridge inside the extender subgraphs.

Specific node updates used by the working path:

• AlignedOverlapCutTransition in frame_transition_nodes.py

  • adds automatic seam scoring across the overlap window

  • exposes blend_frames to soften the splice on both sides of the seam

  • exposes seam_shift so the cut can be moved earlier or later inside the overlap region

  • infers a usable overlap window from the actual frame batches when the workflow resolves the overlap control incorrectly

  • keeps IS_CHANGED forcing recomputation so transition tuning is not hidden by prompt/subgraph caching

• AlignedAudioCutTransition in frame_transition_nodes.py

  • uses the image seam index to keep the audio splice aligned with the chosen visual seam

• LTXExtensionFrameLogic in frame_transition_nodes.py

  • remains the overlap/frame reference utility node used by the extender sampler subgraphs

  • its outputs are now consumed through the corrected v1.6 subgraph wiring instead of the broken overlap bus that previously collapsed to a one-frame cut

The v1.6 workflow is the release workflow to use when you want the smoother transition behavior that was validated during testing.

Use example_workflows/Yogurt_LTXDirector_extender_v1.6.app.json as the current tested template workflow. Files under example_workflows/ are exposed through ComfyUI's workflow template system.

Before loading the v1.6 workflow, update the custom node itself from this repo. The workflow depends on the current transition and patch behavior in __init__.py, frame_transition_nodes.py, and patches.py; older installed node files can load the workflow but still miss the seam, overlap, and audio-alignment fixes it expects.

Custom nodes now available via comfy registry:
via cli "comfy node install LTXDirector-Extender" or just search the manager for LTXDirector-Extender

Workflow file:

- example_workflows/Yogurt_LTXDirector_extender.json

**If the workflow opens as an app, just select graph -> exit app mode

The Yogurt LTXDirector_Extender workflow: generate the first clip with the standard LTX Director path, then continue with the extender path using prior video and audio context.

Primary nodes in this fork:

- LTX Director Extender

- LTX Director Guide Extender

- LTX Prompt Relay Encode + Temporal LoRA

The standalone LTX Prompt Relay Encode + Temporal LoRA node included in this fork. This is kept separate from the tested LTX Director extension path until the combined workflow is validated.

This node allows you to apply specific lora weights to segments within the prompt. This is useful when using a lora without a trigger word, and you don't want it influencing part of the generation. for example, under the lora segment weights you put 0.0,1.0,1.0,0.5 if you want that particular lora to have no effect during the first segment, 1.0 during the second and third segments, and 0.5 for the last segment.

What The Workflow Does

The intended flow is:

1. Generate the first clip with the standard LTX Director node.

2. Use the Yogurt LTXDirector_Extender flow to append new segments.

3. Feed the extender sampler subgraph with the previous clip's video and audio context so the next section continues from what was already generated.

This lets the user keep extending a sequence while preserving context from the earlier result instead of treating each segment as a fresh standalone generation.

Cut vs Blend In The Sampler Subgraph

Inside the sampler subgraph, there is a choice between cutting frames or blending them depending on the effect you want.

- Use cut when you want a harder transition.

- Use blend when you want a softer overlap transition.

If you swap this behavior, change the output connector from cut to blend on that node and reconnect it to the following node in the subgraph.

Duration Must Match Manually

The duration configured in LTX Director Extender must match the duration configured inside the sampler subgraph.

This is manual for now. A GetNode attempt was not reliable enough in this workflow, so both places should be set by hand until that is improved.

If those durations do not match, the extension path can drift away from the intended overlap and continuation timing.

Extending From A Previously Created Video

If you want to extend from a video that already exists:

1. Use a Load Video node to bring the previous clip into the workflow.

2. Disable the standard LTX Director generation path for the initial clip.

3. Connect the loaded video's frame output and audio output into the extender sampler subgraph.

That lets the extender path continue from an already rendered clip instead of requiring the first section to be regenerated inside the same workflow.

Practical Runtime Notes

Tested results so far:

- Up to about 150 seconds of generation on 16 GB VRAM

- Occasional 64 GB system RAM exhaustion on the final step for very long runs

- 10 to 15 second extension durations tend to work consistently

For stability, shorter extension chunks are the safer default. Long chained runs are possible, but they are more likely to hit RAM limits near the end.

UI Note
The extender workflow is intended to use the renamed fork-specific nodes:

- LTX Director Extender

- LTX Director Guide Extender

- LTX Prompt Relay Encode + Temporal LoRA

If a workflow still references the original LTXDirector or LTXDirectorGuide node types, it will bind to the upstream WhatDreamsCost nodes instead of the forked extender versions.

Workflow V 1.1 corrects an issue so that the nodes show up as needed in comfyui manager.

Download the nodes:

https://github.com/Yogurt1192/LTXDirector-Extender.git or search the manager for LTXDirector-Extender

## Install

1. Clone this repo into your ComfyUI custom_nodes folder.

2. Restart ComfyUI.

3. Re-open your LTX Director extension workflow and reconnect any nodes if your local workflow still points at an older custom-node install.

Example:

```bash

cd /path/to/ComfyUI/custom_nodes

git clone https://github.com/Yogurt1192/LTXDirector-Extender.git

```

The included workflow is designed for the extender node names in this fork. If you install both the upstream repo and this fork side by side, use the extender-named nodes when building or updating workflows.