NovelAI 一键分块放大

The pack's tile upscale exists in two flavors: the staged pipeline (planner → refine → merge → seam fix, four nodes you can inspect individually) and this one node that does all of it. NAITiledRefineAllInOne is the "just make it bigger and detailed" button. Fewer moving parts, but you trade away the ability to inspect the plan and the call count before committing.

What it is

It bundles everything: it plans the tiles from your source_image and upscaled_image, refines each tile through NovelAI img2img, merges them back, and optionally runs the seam fix - all in one execution. Every input from the staged nodes is here, so the node is big. The ones that matter:

• source_image / upscaled_image - the original and its local upscale. Same aspect ratio required, same 2% tolerance as the planner.
• tile_width / tile_height - 0 = auto (source dimensions). The intended default.
• overlap_x / overlap_y (128) and blend_width (64) - the geometry that decides how clean the seams are.
• strength (0.28), steps, guidance, sampler, noise_schedule - the per-tile sampling settings. Keep strength low.
• enable_seam_fix (off by default) plus seam_strength (0.16), seam_noise, seam_band_width - the optional cleanup pass.
• use_cache / cache_max_gb - the resumable cache that makes interrupted runs cheap to continue.
• seed - the base for the stable per-tile derived seeds.

Outputs are image (the final merged result) and stats, which reports the base plan, base API calls, cache hits, and the seam-fix call counts. That stats string is your invoice breakdown.

When to reach for it

The one-node form wins when you know the recipe works and you just want the result: you've dialed in the settings once, and now every big image is source → upscale → this node → save. It also wins on canvas real estate - four chained nodes turn into one.

You lose the ability to check planned_calls before the run. In the staged pipeline you can look at the planner, see "this will be 24 API calls," and back out. Here, you find out in the stats string afterward. If you're new to the pack or the images are large, run the staged version once to learn the geometry, then switch to this node once you trust it.

The chroma-speckle detail

The pack applies its chroma-speckle suppression to every refined tile (it's always on in this node), which handles the colored-static artifacts low-strength img2img likes to produce. It's one of the quiet quality wins that makes tile refining on NovelAI look clean instead of noisy.

Install and setup

Part of ComfyUI-NovelAI-GENYTOOLS. Manager search, or:

cd ComfyUI/custom_nodes
git clone https://github.com/XTOGENY/ComfyUI-NovelAI-GENYTOOLS.git
pip install -r ComfyUI-NovelAI-GENYTOOLS/requirements.txt

Restart, set your NovelAI Persistent API Token under NovelAI 2.0: Persistent API Token.

Gotchas

• No call-count preview. The one cost-control feature of the staged pipeline is gone here. Set realistic expectations before you run, and use the cache so a bad run is cheap to redo.
• Keep strength at or near 0.28. This node is happy to burn Anlas on tiles that invent their own content if you push it.
• enable_seam_fix defaulting to off is the pack being honest: only enable it when the merged result shows seams, because it's an extra paid pass over the whole boundary set.
• The aspect-ratio check still applies - mismatch the source and upscale and the run fails before any tile is paid for. That's the seatbelt working.