For tiled video-VAE processing, first confirm that the connected VAE is a video VAE. temporal_size and temporal_overlap have no effect with a standard image VAE. For VAEDecodeTiled, ComfyUI documents that if temporal_size is less than twice temporal_overlap, the temporal overlap is halved. The supplied VAEEncodeTiled documentation confirms that these temporal controls apply to video VAEs, but does not document the same automatic-halving rule for encoding. A small chunk and an automatically reduced overlap are therefore different observations: the former is your selected frame-chunk size; the latter is a documented VAEDecodeTiled constraint applied to the requested overlap.
Concrete Checks and Steps
- Identify the node and VAE path. The relevant controls are
temporal_sizeandtemporal_overlaponVAEEncodeTiledorVAEDecodeTiled. Use the actual input type or model configuration as evidence; node color is only a clue. - Record the three temporal values separately: requested chunk size, requested overlap, and expected effective overlap. Do not replace the last value with the input value when the constraint is triggered.
- Calculate the comparison
temporal_size < 2 × temporal_overlap. ForVAEDecodeTiled, if it is false, the documented temporal constraint is not triggered; if it is true, expect the overlap to be halved according to that node's documentation. ForVAEEncodeTiled, use the inequality only as a recorded comparison unless the applicable implementation or documentation separately confirms an adjustment rule. - Check spatial controls independently.
tile_sizeandoverlapdescribe spatial tiling, while the temporal pair describes frame chunks. A large spatial tile does not prove that the temporal chunk is large enough.
The official VAEEncodeTiled documentation states that temporal controls apply only to video VAEs and have no effect on standard image VAEs. The VAEDecodeTiled documentation describes the same temporal constraint for decoding.
A New Acceptance Worksheet
The following is a hypothetical documentation-based worksheet, not a software test or a guarantee about an output sequence. It uses temporal_size=32 and temporal_overlap=20, so the comparison is 32 < 40 and the documented adjustment is expected to apply.
| Field | Hypothetical observation | Decision |
|---|---|---|
| VAE role | Video VAE is connected | Temporal fields are relevant |
| Temporal chunk input | 32 frames | Record as selected chunk size |
| Temporal overlap input | 20 frames | Record as requested overlap |
| Constraint check | 32 is less than 2 × 20 | Automatic-reduction condition is met |
| Expected effective overlap | 10 frames after halving | Use as the expected adjusted value |
| Spatial tile inputs | tile_size=512, overlap=64 | Keep in a separate spatial record |
This worksheet does not infer how many chunks the whole clip produces, how latent compression maps frames, or how boundary frames are blended. The documentation notes that VAE compression ratios are considered internally, but it does not provide enough information here to calculate those results manually.
Limits and Next Steps
If your observation does not match the worksheet, first verify that the workflow reaches a video VAE node and that the values were read from the same node instance. Then record the exact node, VAE type, input values, and any visible runtime message. Do not diagnose the issue from tile_size or spatial overlap alone. To cross-check your parameter record, an author-suggested comparison is to hold the VAE and spatial settings fixed and record two temporal-setting cases with the result of temporal_size < 2 × temporal_overlap. For VAEDecodeTiled, the documented rule predicts halving when that condition is true; for example, 32/20 gives an expected effective overlap of 10. When the condition is false, this documented rule is not triggered. This is a way to organize a check against the documentation, not an official ComfyUI detection feature or a software test. If the interface does not expose an adjusted value, label it as a rule-derived expectation, not a runtime observation.
Three Quick Questions
Does `temporal_size=32` automatically mean the chunk is too small?
No. It is simply the selected number of frames per temporal chunk. Whether it is suitable depends on the workflow and available resources; the cited rule only determines whether the overlap constraint is triggered.
Is `overlap=64` the same as `temporal_overlap=64`?
No. The first is spatial overlap between image tiles. The second is temporal overlap between frame chunks. Record them in separate dimensions.
What does `32/20` mean in the example?
It means a 32-frame temporal chunk and a requested 20-frame temporal overlap. Because 32 is below 40, the documented expected effective overlap is 10 frames.
Sources and next step
- VAEDecodeTiled - ComfyUI Built-in Node Documentation
- VAEEncodeTiled - 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.