Tile Test: Captions

There's a failure in tiled upscaling that no blend can fix: a tile that should be empty sky grows a car, because your global prompt says "red car" and the model doing that tile was never told it's looking at the sky.

This pack solves it by throwing the global prompt away: each tile's positive conditioning is built from the image itself - vision tokens of the tile's crop, its slice of a whole-image encode, and a short bit of text naming what's actually there. Tile Test: Captions is that text pass, on its own.

Why you'd split it out

Captioning is the slowest part of a VL upscale - every tile needs a vision encode plus a text pass. When something comes out wrong you want to know whether the tile text was bad or the sampling was, and in the all-in-one node you can't tell. Here you get the answer as text before rendering anything: every stage of the tags pass prints as Markdown.

It also has no seed, which matters more than it looks. ComfyUI caches it, so re-rolling the seed downstream leaves the tile texts alone.

Two kinds, and you pick one

The mode is decided by which instruction socket you connect. Connect tile_caption_instruction or tile_tags_instruction - never both, and leaving both off leaves the node idle.

Tags (the shipped tags preset) is the interesting one. Three stages: the VL model lists what it sees in the tile; each candidate is scored true-or-false against the tile and only near-certain ones are kept; then six strips - three rows, three columns - decide each tag's position, and a tag no strip can hold is dropped. Scoring runs through logit-classifier, the pack's one Python dependency.

Caption asks one question about each tile's crop and uses the answer as that tile's text. A preset can also write a style caption from the whole image once and place it on top of every tile's text.

The inputs a beginner actually touches

• image and layout - the canvas from Tile Test: Upscale, and the CATR_LAYOUT the tiles are captioned from.
• clip - a vision-language text encoder with a text generator, from the Krea 2 family. SD/SDXL CLIP, T5 and plain Qwen3 are rejected with a clear error.
• tiles - comma-separated tile numbers as the Layout overlay labels them. Empty captions everything. This is the widget the chain is built around.
• with_neighbors (on) - also caption the bordering tiles of each named tile. Must match Tile Test: Render's setting, or Render comes up short.
• position_terms (on) - the strip scoring. Off, tags are written without a term like top-left and nothing is dropped for failing a strip.

Then the tuning numbers, each wireable from Tile Test: Settings: caption_megapixels, the two token budgets tile_caption_max_tokens and global_style_max_tokens (which also cover the model's hidden reasoning turn, hence the generous 768 default), and the three tag thresholds, defaulting to near-certainty. Optional strings in: prompt, global_style_instruction, prompt_tags_instruction, tile_tags_verification_statement.

The outputs

captions is the CATR_CAPTIONS object - wire it to Tile Test: Render. tile_texts is Markdown listing the style caption and then each tile's text, exactly what Render conditions on. tiles passes the named tile list through for Render's tiles input; empty in, empty out.

Then the four inspection outputs - prompt_tags, tags_listed, tags_verified, tags_final - which are the reason to use this node instead of the all-in-one. They're Markdown for Preview as Text's Markdown mode, so Nodes 2.0 needs to be on.

About that prompt input

The diffusion model never reads your prompt. On the tags preset the VL model reads it once per image to list the physical things it names, and each tile checks that list against its own content - a checklist, not conditioning. In a caption preset, {PROMPT} in the instructions is filled from it, and a preset whose instruction carries {PROMPT} errors if prompt is empty; the tags preset just skips its prompt question.

Install

ComfyUI Manager → search "Context-Anchored Tile Refine". Or:

cd ComfyUI/custom_nodes
git clone https://github.com/Blakeem/ComfyUI-ContextAnchoredTileRefine
cd ComfyUI-ContextAnchoredTileRefine
pip install -r requirements.txt
cp settings.toml settings.user.toml

logit-classifier>=0.4.0 is the only dependency. You'll want the Krea 2 stack too - the model, the Qwen3-VL encoder (~8GB) and the Qwen-Image VAE - because without a vision-language encoder there's nothing to run on.

Where it goes wrong

Wiring both instruction sockets. Exactly one of the two must be on - an easy mistake, since both sit in the same optional block.

A tag list that looks empty. Usually the thresholds doing their job: near-certainty cuts hard, and a scene the VL model is unsure about produces short tags. Loosen tile_tags_verification_threshold carefully.

A broken settings file. If this node is missing from the menu entirely, the settings file has a syntax error; ComfyUI prints the filename and the problem to the console, and the node stays gone until you fix it.