For a long LTXV generation, first choose a context_length that can contain the action you need to remain coherent; then consider context_overlap for behavior at window boundaries. Length and overlap are both expressed in real frames, but they are converted to latent frames differently. Neither setting guarantees a coherence improvement. Compare changes with the same seed and other sampling conditions held fixed.
Choose a window for the action, then assess its boundary
The ComfyUI LTXVContextWindows documentation says context_length must follow 8*n+1, with a default of 145 real frames. It converts length to latent frames as ((context_length - 1) // 8) + 1. Thus, 145 real frames map to 19 latent frames. context_overlap is also entered in real frames, but converts by integer division by 8, with a minimum of zero. For example, 40 real frames map to 5 latent frames. These conversions describe inputs, not a guarantee of a particular visible result.
- Find the longest action span whose continuity matters, such as a turn followed by a reach. Record its start and end in real-frame positions; do not assume that overlap can make a window contain an action that exceeds its length.
- Select a valid
context_lengthlarge enough to cover that span, allowing room for surrounding motion if practical. Use only values matching8*n+1; the documented step is 8. - Keep the documented default overlap, or another chosen baseline, and generate a candidate. Note where the action falls relative to the apparent window boundaries and inspect those moments.
- If the action fits within a window but discontinuity appears near a boundary, make a second candidate changing overlap only. If the action spans beyond the available window, compare a longer valid length first. Keep seed and all other sampling inputs fixed for each pair.
The node also exposes schedule, stride, fuse method, and other settings. This guide isolates length and overlap; if you change those settings too, the comparison no longer identifies which change mattered.
Proposed acceptance worksheet
The following is a hypothetical acceptance example, not a test result or expected node output. The 300-frame duration and window counts are planning entries; the supplied documentation does not define a window-count formula or guarantee exact boundary placement. Record observed counts from your own workflow if available, rather than deriving them from an undocumented rule.
| Candidate | Total frames | Context length, real / latent | Overlap, real / latent | Observed windows | Action position and boundary observation | Decision |
|---|---|---|---|---|---|---|
| A: baseline | 300 | 145 / 19 | 40 / 5 | 3, hypothetical observation | A 90-frame turn crosses a visible boundary; the turn seems less consistent there | Check whether the action span fits a window before blaming the join |
| B: longer window | 300 | 201 / 26 | 40 / 5 | 2, hypothetical observation | Same turn appears within one window; boundary looks similar | Keep only if the action improves without a new unwanted change |
| C: more overlap | 300 | 145 / 19 | 64 / 8 | 3, hypothetical observation | Turn still crosses a boundary; the join appears less abrupt in this hypothetical review | Compare with A; retain only if the boundary improves and the action remains acceptable |
Review the same short action in all candidates. Record the actual node inputs, seed, observed window count, action’s position relative to boundaries, and a concise visual judgment. Use frame positions, not rounded seconds, when marking the action.
Limits and what to try next
A larger overlap can alter how windows meet, but it does not extend the documented context length or establish that an action spanning windows will become coherent. Conversely, a longer context is not a promise of improvement. ComfyUI describes the node as dividing generation into overlapping windows to help manage memory and improve temporal coherence, but the documentation does not promise a specific quality outcome or provide a formula for window count. Also, the node’s stated input ranges and conversions are not a substitute for checking the actual accepted values in your workflow.
If neither controlled change helps, confirm the actual node and its input types in the workflow; a node’s color is only a clue, so rely on the displayed type labels. Then compare the same action under unchanged settings and note whether the defect is inside the action or localized at a boundary. Consider other exposed context settings only as a separate experiment, changing one at a time. Do not describe a visual detail absent from the generated frames as recovered by overlap.
Three quick questions
Should I raise overlap first? Only when the action fits the selected window and the issue is localized at its boundary. Otherwise, assess length first.
Can I enter any real-frame length? The documented pattern is `8*n+1`, and the node lists a step of 8. Choose a value satisfying that rule and verify the actual input is accepted.
Does a higher overlap always reduce visible seams? No such guarantee is in the documentation. Treat it as a controlled candidate and judge the resulting boundary.
Sources and next step
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.