If VAEDecodeTiled receives a nested batch of latents, ComfyUI's node documentation says it decodes only the first item in that batch. That points to the input structure, not automatically to images being dropped by a later save node. Confirm the LATENT structure and count the IMAGE output before tracing the downstream workflow.
Check the latent structure and count the output
The node accepts a LATENT input, but that label alone does not tell you whether the value is a single latent, a regular batch, or a nested batch. Node color can be a clue; use the workflow's actual type labels and the producing node's output information where available. Do not infer nesting merely because you see several images later in the graph.
Use this ordered check when fewer images appear than expected:
- Identify the connection feeding
VAEDecodeTiledat itssamplesinput. Record the producing node and the displayed data type. Check the upstream node's description or output details for whether it emits a nested batch, rather than relying on color or a preview. - Write down what you expect to decode: one item, each item in a regular batch, or the contents of a nested batch. Keep this expectation separate from the number of previews shown elsewhere in the workflow.
- Inspect the
IMAGEoutput directly afterVAEDecodeTiled, beforeSave Imageor another downstream node. Record the number of images visible or reported there, if the interface exposes a count. - Compare actual output with the intended count. If the input is identified as a nested batch and only its first item appears at the tiled node's output, that matches the documented limitation. If the output already contains the expected images, look at downstream selection, filtering, or saving instead.
The node's `VAEDecodeTiled` documentation describes tiled decoding for latent input and explicitly notes the nested-batch behavior. That statement is specific to this tiled node; do not assume it describes every VAE decoder.
Record a before-save acceptance check
The table is a proposed diagnostic worksheet, not a guaranteed output specification. Its example observations are hypothetical. They do not establish an image-count rule for regular batches, whose behavior is not specified in the cited node documentation.
| Check | Hypothetical observation | Decision |
|---|---|---|
| Input type | Upstream details identify samples as a nested batch of 3 latents | The documented condition applies; test the tiled node's output before blaming saving |
| Intended count | Creator expects 3 decoded items | Record this as the workflow requirement, not a node guarantee |
| Output before saving | VAEDecodeTiled exposes 1 IMAGE item | The shortfall exists at decoding; investigate the nested input organization |
| Output before saving | VAEDecodeTiled exposes 3 IMAGE items, but the saved set has 1 | The decoder produced the expected count; inspect downstream nodes and save workflow |
For your own run, replace these hypothetical values with the type label, intended count, and observed output count. If no count is displayed, compare the image items available at the decoder output using the controls your interface provides; do not treat a single preview as proof that only one item exists.
Choose a next step without changing unrelated settings
When the input is nested and the decoder produces only one item, decide whether all items need decoding. If so, revise the batch organization upstream so the input to VAEDecodeTiled is in a form that represents the intended items, then check the output again. The cited documentation does not specify which upstream node or transformation will suit your workflow, or promise that a particular rearrangement is valid. Confirm the resulting LATENT type and structure before relying on it.
If the input is not nested, or the tiled node's output already has the intended count, this limitation does not explain the discrepancy you see. Trace the IMAGE connection after decoding and check whether another node selects, filters, or routes items before saving. Do not change tile_size or overlap to address an item-count mismatch: the documented nested-batch note concerns which item is decoded, not tile boundaries. The page describes those spatial settings separately and also covers temporal chunking for video VAEs; neither description establishes that adjusting them will decode additional nested items. See the node's documented inputs and outputs for that scope.
Common questions
Does this prove `Save Image` lost frames?
No. Compare the decoder's IMAGE output with the expected count first. If the shortfall is already present there and the input is nested, the documented tiled-decoder behavior is a better match.
Does the note apply to every VAE decoder?
The cited statement names VAEDecodeTiled. It does not establish the same behavior for other decoder nodes; check their own documentation before drawing that conclusion.
What if the type label is unclear?
Inspect the node that creates samples and its available output details. Until you can establish whether the input is nested, keep the cause open and compare the output immediately after decoding with the downstream result.
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 photo editor. The external software controls discussed above belong to their respective products.