Tile Test: Render

Everything above this node is cached. The layout is solved, the canvas is upscaled, the tile texts are written. Tile Test: Render is where the diffusion happens, and since it's the only node in the chain holding a seed, it's the only one you re-run for a different result.

That's the whole design. A 24-tile 8K refine is minutes of GPU time, and finding out tile 17 has a weird hand shouldn't cost you the other 23. Name tile 17, change the seed, re-run.

The engine underneath

Worth knowing, because it isn't what Ultimate SD Upscale does. The VL nodes keep the entire image in one shared canvas latent, diffuse every tile one step at a time, and between steps consolidate the tiles back into that canvas - the overlap bands cross-dissolved in latent space by the same directional feather the base node applies in pixel space. Each tile runs its own sampler, but all of them are held to one schedule, which is why the sampler list is restricted.

MultiDiffusion and Mixture of Diffusers also touch every tile on every step, but they average the overlapping predictions, and averaging disagreements gives you a soft blend. Here only the thin shared band gets predicted twice.

Inputs

The wiring is the chain you already built: image from Tile Test: Upscale, layout from Tile Test: Layout, plus model, clip and vae. clip carries the same constraint as the Captions node - a vision-language encoder from the Krea 2 family. There's no positive prompt input at all, because every tile's positive is built from the image.

Then seed, sampler_name, scheduler, steps, cfg, denoise. Respect denoise: at 0 it hands the image back unrefined, and the README's own samples sit around 0.35.

anchor_source is the setting with a real personality. source image (default) fills the context ring with the unmodified input, locking placement and style to what you started with - flaws included. live canvas uses neighbors' in-progress results instead, so the refine can repair flawed content; expect more invention and slightly brighter output. If you get drift you didn't ask for, flip this back first.

surface picks what fills each tile's positive: vision tokens, vision tokens and captions, or captions. The combined one is the default and the one to use - tokens keep the tile faithful and at the right scale, the text names what's there. Both caption surfaces need Tile Test: Captions connected.

canvas_tokens and crop_tokens (165 each) are how many vision rows a tile takes from the whole-image encode and from its own crop; zero turns a source off. Tile Test: Settings will hand you these.

Then the two control widgets. tiles is comma-separated tile numbers as the overlay labels them; empty means a full run over the entire canvas. with_neighbors (on) renders the named tile with its bordering tiles as lanes of one run, which is what a full run does. Turn it off and the tile renders alone - and its anchor ring then stays unrefined source, because no neighbor lane is refining it. Optional inputs are captions and negative.

Outputs

Two, both image lists. tiles is one image per named tile, the canvas cut at that tile's crop rect. blocks is the refined block rendered for each named tile - spanning its neighbours when with_neighbors is on, so it's the one to judge seams with. With an empty tiles input, both are the entire refined canvas.

The sampler restriction isn't arbitrary

The synchronized engine has to time every sampler's model evaluations, because all tiles share one schedule. Eleven are supported: euler, heun, dpm_2, dpmpp_2m, the four dpmpp_2m_sde variants, and the two exp_heun_2_x0 ones.

Anything else is rejected before sampling starts. dpm_fast, dpm_adaptive and uni_pc are unsupported by design - they own their own schedule or internal history, so no evals-per-step figure can time them. uni_pc is a common pick on distilled models, so this one bites. The upside: the rejection is caught at queue time, before the encoder and model load.

Installing

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

One dependency, logit-classifier>=0.4.0. Restart after. You also need the Krea 2 stack - the diffusion model, the Qwen3-VL text encoder (~8GB) and the Qwen-Image VAE.

Where it goes wrong

"renders one picture at a time." Batch size one, like every node in this pack.

A size mismatch. This node wants the exact canvas Tile Test: Upscale returned. A visually identical image resized by one pixel gets rejected.

Caption surfaces with nothing wired into captions. Setting surface to either caption option and leaving the input empty stops the run - the wiring mistake people make when swapping surfaces to compare.

And the one that looks like a crash: a syntax error in the settings file makes this node vanish from the menu, with the reason printed to the console. Fix the TOML, restart, it's back.