Context-Anchored Tile Refine (VL)

Tiled upscaling has a sneaky failure mode with a global prompt: the prompt says "a cyberpunk city," and now every tile wants to hallucinate a skyscraper into a tile that's pure sky. Context-Anchored Tile Refine (VL) solves this by throwing the prompt away entirely. On this node, each tile's conditioning is built from the image itself, using the same Qwen3-VL vision encoder Krea 2 already runs on. It's the sibling of the base Tile Refine node, built for Krea 2, and it's the one you want when "describe what's in front of me" beats "restate the prompt."

How it works

The base node refines tiles one after another and blends the seams. The VL node takes a different route: every tile becomes a lane of one synchronized diffusion run over a single shared canvas latent. All tiles step together, and between steps the overlap bands are cross-dissolved in latent space. Seams aren't hidden after the fact - they're prevented, which is why this path needs none of the base node's boundary-cut or color-match machinery.

Then there's the conditioning trick. The whole image is encoded once into a grid of vision tokens that carry their region's tone, palette, and objects. Each tile conditions on its own slice of that grid - a RoIAlign-style idea in conditioning space. Because the tokens know roughly where they are, tiles stay aware of their surroundings without ever being told about a prompt object that isn't there. No global text prompt is used, so nothing phantom gets re-instantiated. The vlm_method input picks the flavor: vision tokens (fastest, one shared encode), captions (the VL model writes a short description of each tile's crop - better at repairing mushy areas, slower), or the default vision tokens and captions, which does both.

One input worth understanding is anchor_source. source image (default) shows the context ring the unmodified input, which locks placement and style - flaws included. live canvas shows the neighbors' in-progress result instead, so the model can repair flawed content. Expect more invention and slightly brighter output on live canvas; use it when the source has artifacts you want gone.

What wires in where

The required inputs are the base node's (image, guider, sampler, sigmas, vae, noise, the tile geometry) plus three VL-specific ones: clip (must be a vision-language encoder - Krea 2 family; a plain SD/SDXL CLIP gets a clear error), anchor_source, and vlm_method. The guider's positive prompt is ignored - that's by design - but its negative still applies. A mask works and keeps the global view: the whole image is still encoded once, so a masked region refines aware of what's around it. Feed an inverted mask on a second pass to do background and subject separately.

Where people get burned

• Samplers are restricted. The synchronized run must time every sampler's model evaluations, so only euler, heun, dpm_2, dpmpp_2m, dpmpp_2m_sde (all variants), exp_heun_2_x0, and exp_heun_2_x0_sde work. dpm_fast, dpm_adaptive, and uni_pc run their own schedule and are rejected before sampling starts. Not after a wasted run - before.
• ControlNet is silently ignored. Every tile's positive is vision conditioning, so there's nothing for a hint to attach to. The node logs a warning. If you need ControlNet during refine, that's what the base node is for.
• Caption wording lives in a file. vlm_method's caption options come from settings.toml in the node's folder. Copy it to settings.user.toml if you want to edit the tile prompts - the user copy survives updates, and editing preset wording applies on the next run.

Install is the same as the rest of the pack: search "Context-Anchored Tile Refine" in ComfyUI Manager or git clone https://github.com/Blakeem/ComfyUI-ContextAnchoredTileRefine into custom_nodes. No pip deps, no model downloads - though Krea 2 itself needs its Qwen3-VL text encoder and Qwen-Image VAE files, which is the setup step people actually forget. Tested only with Krea 2 and Qwen3-VL; Krea 2 Turbo and other Qwen3-VL models are untested, so treat them as YMMV.