Tiled VAE Settings
Four tiling numbers as wires, so a subgraph can share them
- tile_size
- overlap
- temporal_size
- temporal_overlap
Tiled VAE Settings is a small but genuinely useful idea: it takes the four numbers every tiled-VAE node makes you set - tile size, overlap, and the two temporal versions for video - and exposes them as connectable integer outputs. Set them once in the widgets on this node, and any tiled encode or decode downstream can read them from a wire instead of you hunting for each value inside a nested node. The author's own description of it is disarmingly honest: "This one's mostly for me tbh."
Why tiled VAE at all? Because VAE encode/decode is one of the most VRAM-hungry steps in ComfyUI, and tiling splits the image into overlapping chunks so a 4K decode fits on a card that would otherwise run out of memory. The community's own numbers from the upscaling side: tiling is far slower but used less than a third of the VRAM of a single-pass decode. For a 6GB card, tiled VAE is often the difference between "it runs" and an out-of-memory error. The temporal_size and temporal_overlap knobs are the video-VAE equivalent - how many frames per temporal tile, and how many frames those tiles overlap.
How it works
There's no cleverness here, and that's correct. The node is a pure value source: four integer widgets on the front, four integer outputs on the back, and run simply returns the widget values. The tiling itself still happens in whatever tiled-VAE node you wire these into - this node just stops those parameters from being buried. The settings each carry a tooltip straight from the author, and the defaults are sane: tile_size 512, overlap 64, temporal_size 64, temporal_overlap 8.
Inputs and outputs
tile_size(INT, default 512, range 64–4096) - pixels per tile.overlap(INT, default 64, range 0–4096) - overlap between adjacent tiles.temporal_size(INT, default 64, range 8–4096) - frames per temporal tile (video VAE).temporal_overlap(INT, default 8, range 4–4096) - overlap between temporal tiles, in frames.
Outputs: tile_size, overlap, temporal_size, temporal_overlap - the same four, as INT wires.
Install
Manager → search "Tojioo Passthrough" → Install → restart, or:
cd ComfyUI/custom_nodes
git clone https://github.com/Tojioo/tojioo_passthrough.git
No models, no pip deps. Category: Tojioo Passthrough → Other. Small single-maintainer pack, GPL-3.0, active on r/comfyui.
Common issues
Two practical notes. First, these outputs only do anything if the tiled-VAE node you're using accepts its parameters as inputs - most do (right-click a widget and "convert to input"), but it's worth checking before you wire four numbers to nothing. Second, keep overlap sensible relative to tile size: overlap exists to hide seams between tiles, but pushing it toward or past the tile size just wastes compute, since the same pixels get processed repeatedly. Start at the defaults (512/64) and only raise overlap if you're seeing seams at tile boundaries. It's a niche tool - but if you're building a reusable tiled decode subgraph, it's exactly the kind of thing that saves you from editing four widgets in four places every run.
Inputs (4)
| Name | Type | Default | Description |
|---|---|---|---|
| tile_size | INT | 51264–4096 | Size of each tile for VAE encoding/decoding |
| overlap | INT | 640–4096 | Overlap between adjacent tiles in pixels |
| temporal_size | INT | 648–4096 | Number of frames per temporal tile (video VAE) |
| temporal_overlap | INT | 84–4096 | Overlap between temporal tiles in frames |
Outputs (4)
| Name | Type | Description |
|---|---|---|
| tile_size | INT | — |
| overlap | INT | — |
| temporal_size | INT | — |
| temporal_overlap | INT | — |