Reduce memory at the tiled VAE decode stage

Try to recover a video decode that runs out of memory by reducing the spatial tile size at the existing VAEDecodeTiled node 109. Start with LTX-2 video with generated audio, version 0.1.6, and the reported failing node. The recipe already uses tiled decoding; this intervention makes its tiles smaller.

The intended deliverable is Sweet Tea Preview Video, exposed as output:75. A successful retry must both reach that output and produce imagery you can use. The retained sources establish the control and its stage of action, but provide no measured memory savings or guarantee that the retry will complete.

Node 109 receives samples from LTXVSeparateAVLatent node 110, slot 0, and vae from CheckpointLoaderSimple node 116, slot 2. It produces decoded IMAGE frames at slot 0. Those frames enter images on CreateVideo node 108, which also receives the separately decoded audio from node 144. Node 108's video then enters video on SweetTeaPreviewVideo node 75.

The saved spatial controls are tile_size = 512 and overlap = 64. tile_size determines the spatial tile dimensions requested for decoding; overlap determines the overlap between tiles. The implementation converts tile dimensions using the VAE compression factors and calls decode_tiled. Smaller tiles change the spatial work handled by each tile, which is the mechanism behind the existing memory-reduction adaptation. They can also change seams and decode work; the sources establish no numerical memory or speed prediction.

The implementation reduces overlap to one quarter of tile size when tile_size is less than four times overlap. At the proposed tile_size = 256 with overlap = 64, the values are exactly at that boundary, so this guard does not reduce overlap.

The saved video controls are temporal_size = 4096 and temporal_overlap = 8: the node schema describes these as frames decoded at a time and frames of temporal overlap for video VAEs. The published adaptation changes only the spatial tile_size; it does not establish a temporal tuning recommendation.

The existing tiled_decode rule, revision 2026-10-09.1, is bound to node 109 as tiled_decode:109. The recipe's published smaller_decode_tiles adaptation, revision 2026-10-09.1, uses binding decode_tiles:109 and applies when failure.stage is decode. Use that existing annotation for this selected revision.

Node 109 is marked hidden by default within the Text to Video (LTX 2.0) subgraph, scope root:92. Locate its internal controls if your editor exposes them; the source pack does not establish the exact UI navigation. For this retry, retain the saved overlap and temporal settings and the rest of the recipe configuration so the spatial tile change can be assessed on its own. Completion and output quality remain uncertain until the retry.

Keep the smaller setting if the retry completes and the imagery is acceptable. Completion demonstrates that this particular run reached the output; it does not establish a memory reduction measurement or a capacity recommendation.

The retained preparation evidence has no measured profiles, scoped GPU or RAM recommendation, or comparable timing. It also leaves dependency preparation incomplete. Neither the guide title nor the published adaptation establishes total-machine fit. These instructions are scoped to the selected recipe revision and the retained ComfyUI implementation.