After temporal upscaling, reconnect earlier LTX keyframes by separating them before the latent is rescaled, processing the resulting plain video latent, then adding the separated keyframes to the new canvas with LTXV Generated Keyframes To Guides. Leave override_frame_indices empty to use the node’s automatic position scaling, or provide one checked target-frame index per keyframe when you need explicit placement. The reinserted guides use a zero noise mask, so they are frozen rather than denoised again.
Decide what must cross the stage boundary
LTXV Add Generated Keyframes creates keyframes that participate in the earlier sampling stage. They are not yet the same thing as either the separated keyframes latent or the later frozen image guides. For a handoff that must preserve their identity and placement, use LTXV Separate Generated Keyframes before spatially upscaling the latent or changing its frame-token layout.
Follow this order:
- Connect the earlier stage’s positive and negative conditioning and video latent to Separate Generated Keyframes. Its conditioning outputs have the generated-keyframe metadata removed; its video-latent output has those keyframes stripped out. Its separate
keyframesoutput retains the keyframe frames and their pixel-frame positions. - Send only the stripped video latent through the later temporal processing. Keep the separated keyframes output and both cleaned conditioning outputs for the handoff.
- On the longer canvas, connect those outputs to Generated Keyframes To Guides, along with the target video latent and the VAE. Confirm the conditioning comes from the separation node, not from a branch that still carries generated-keyframe metadata.
- Decide how positions should map. An empty
override_frame_indicesuses recorded positions, automatically scaled if the target has a different number of pixel frames. For explicit placement, supply one comma- or space-separated integer per keyframe. The indices must be unique and lie between 1 and the target latent’s number of pixel frames minus 1, inclusive. Check each index against that range and the intended placement before sampling. - Send the resulting conditioning and latent to the later sampling stage. Add the guides before merging video and audio latents with Concat AV Latent.
Use a handoff worksheet
The following is an explicitly hypothetical planning example, not a measured node result. Assume one generated keyframe was recorded at pixel-frame index 24, and the target canvas has twice as many pixel frames as the source canvas. Before sampling, compare the recorded location, the automatic mapping shown or reported by the workflow, and any explicit override you propose. Do not infer a particular scaled index from “twice as long”: the documentation says positions are scaled but does not specify a rounding rule.
| Handoff check | Hypothetical observation | Decision |
|---|---|---|
| Separated position | One keyframe; recorded pixel index 24 | Keep the keyframe latent and its position together; confirm the count matches |
| Target canvas | Twice the source pixel-frame length | With an empty override, inspect the automatic mapped index before sampling |
| Explicit remap | Proposed override is index 48 | Use it only if it is the intended destination, is unique among overrides, and lies between 1 and the target latent’s number of pixel frames minus 1, inclusive |
| Guide state | Reinserted keyframe is intended to remain unchanged | Confirm the node’s frozen-guide behavior; its noise mask is 0 |
| Input contract | Target input is a plain video latent; batch size is 1 | Resolve either mismatch before connecting the guide node |
For a practical comparison, record the target length and intended keyframe locations first, then inspect the workflow’s resulting indices before running the later stage. This is an editorial check, not a claim that every interface exposes the mapping in the same way.
Limits and next steps when it fails
Generated Keyframes To Guides supports batch size 1. Its target latent must be a plain 5D video latent, not a combined video-and-audio latent; add guides before Concat AV Latent. The separation node also requires the supplied latent to match the frame-token layout and recorded range associated with the keyframes. That is why separation belongs before spatial upscaling. If audio and video are still combined at separation time, use Separate AV Latent first.
If separation reports missing keyframe metadata, check that the earlier stage used Add Generated Keyframes and that its corresponding conditioning and latent are connected. If reinsertion rejects conditioning, use the positive and negative outputs from Separate Generated Keyframes. If counts do not agree, compare the number of separated keyframes with the recorded positions. If an index is rejected, check that overrides are unique, within the documented target range, and supplied one per keyframe. If an automatic position is not where the new edit requires it, replace the automatic mapping with deliberate overrides; do not assume temporal upscaling preserved the intended editorial timing.
Three quick questions
Should I leave the override empty after a temporal x2? Leave it empty when the node’s automatic scaling places every keyframe where the new canvas requires it. Inspect the resulting positions first; use explicit indices if they do not match the plan.
Can the later sampler change a reinserted keyframe? The node adds these as frozen image guides with a zero noise mask. Its strength also controls guide strength; the documentation describes 1.0 as a hard pin and lower values as more relaxed, so choose according to the intended constraint.
Can I separate after spatial upscaling? The separation node requires the keyframe token layout to match the supplied latent. The documentation warns that rescaling first breaks that match, so separate before spatial upscaling.
Sources and next step
- LTXVSeparateGeneratedKeyframes - ComfyUI Built-in Node Documentation
- LTXVGeneratedKeyframesToGuides - ComfyUI Built-in Node Documentation
This guide interprets public documentation and uses explicitly hypothetical acceptance examples, not application tests. Controls and availability can change; verify the cited manual against your installed version.
For work created in Imagild, carry the input requirements and acceptance checklist into the film workspace. The external software controls discussed above belong to their respective products.