ComfyUI Node

TBG ETUR Tile Overrides

TBG ETUR Tile Overrides

By Ltamann·Created about a year ago·Updated 19 days ago· 156
TBG ETUR Tile Overrides
  • TBG_Pipe
  • Model_1
  • Model_2
  • Model_3
  • CnetPipe_1
  • CnetPipe_2
  • CnetPipe_3
  • TBG_Pipe
Rebuild_only_modified_tilestrue
requeue0

You ran a 20-tile upscale, and one tile - of course it's the face - came out wrong. The traditional answer is "re-run the whole job and hope." TBG ETUR's answer is the Tile Overrides node: it sits in the middle of the graph, between the Tiler and the Refiner, and lets you edit the sampler settings for individual tiles. The class is literally named TBG_TilePrompter_v1 - it's a descendant of the old McBoaty per-tile prompter - and the node description calls it "a note to set per tile sampler settings."

It's an output node that passes the TBG_Pipe straight through, so it doesn't change what's in the pipeline; it attaches per-tile overrides to it. The overrides live in a hidden JSON blob (tile_edits_json) that persists in your workflow, and you edit them through the pack's web UI - click a tile, then set that tile's prompt, denoise, seed, and ControlNet strength.

How it works

The node carries a list of per-tile edits keyed by tile id: prompts, denoises, seeds, cnet_strength, plus deeper overrides like cfg_overrides, model_overrides, cnetpipe_overrides, color_match_overrides, and ignore_general_prompts. The Refiner reads those when it samples each tile, so one tile can run at denoise 0.8 with a completely different prompt and a different model while its neighbors run at 0.3 on the default model.

The clever part is Rebuild_only_modified_tiles (on by default): it fingerprints your inputs and reprocesses only the tiles whose settings actually changed. The tooltip claims typical speed gains up to 90% on re-runs, which matches the pack's general caching story from the changelog. Change one tile, and you're not paying for the other nineteen again.

Inputs that matter

Required: TBG_Pipe and the Rebuild_only_modified_tiles toggle. That's it for the everyday path - everything else is optional firepower:

  • Model_1 / Model_2 / Model_3 - swap in a different model for specific tiles (say, a detail model for eyes while the base model does the rest). All three must share the same latent space/VAE.
  • CnetPipe_1 / CnetPipe_2 / CnetPipe_3 - per-tile ControlNet pipes, so you can apply structure/depth conditioning only where it's needed.
  • requeue - a manual/automatic re-run counter, useful while you're iterating on a single tile.

Output: TBG_Pipe, straight to the Refiner. The node is marked OUTPUT_NODE, so it's a legitimate endpoint in the graph - you can even stop your workflow here and inspect the pipe.

Install

Part of the whole pack:

cd ComfyUI/custom_nodes
git clone https://github.com/Ltamann/ComfyUI-TBG-ETUR
cd ComfyUI-TBG-ETUR && pip install -r requirements.txt

or ComfyUI Manager → "TBG Enhanced Upscaler," then restart. It needs a current ComfyUI (v3 API) and pulls in the pack's heavy requirements (opencv-contrib-python, kornia, timm, triton, etc.).

Common issues

  • Overrides not "sticking" - the edits persist through the node's hidden JSON in the workflow. If you reload the workflow in a fresh browser tab and your per-tile edits are gone, you've hit a persistence hiccup - the changelog explicitly added workflow persistence for these inputs, so a recent update usually fixes it.
  • "Rebuild only modified tiles" looks wrong - it's marked experimental in the tooltip. If a partially-rebuilt image looks inconsistent, toggle it off and do a full re-run to confirm the per-tile edit is the problem, not the caching.
  • Mixing models - Model_1/2/3 must be in the same latent space (same VAE family). Sticking a Qwen model in with FLUX tiles will produce garbage, not overrides. The source comments spell this out: "same Latent space (same VAE)."
CategoryTBG/ETUR Tiled Upscaler and Refiner

Inputs (9)

NameTypeDefaultDescription
TBG_PipeTBG_Pipe
Rebuild_only_modified_tilesBOOLEANtrueExperimental: Detects user input changes and reprocesses only affected tiles instead of the full image. Typical speed gains up to 90%
Model_1optMODEL
Model_2optMODEL
Model_3optMODEL
CnetPipe_1optControlnet_Pipe
CnetPipe_2optControlnet_Pipe
CnetPipe_3optControlnet_Pipe
requeueoptINT00–99999999999

Outputs (1)

NameTypeDescription
TBG_PipeTBG_Pipe