ControlNet Preprocess Tile

Tile is the odd one out among ControlNet preprocessors, and it's the one with the most hidden power. The others detect something about your image - edges, pose, regions. Tile feeds your image back as its own condition at a lower resolution, so the model redraws it with more detail while staying structurally faithful. That's the mechanism behind the classic "tile ControlNet + tiled diffusion = very realistic upscaler" workflow, and it's also why tile is the go-to for style transfer: feed a reference at low control weight and the model reinterprets it in the prompt's style.

This node runs that preprocessor on Runware's cloud. Feed it an image, get the tile-conditioned map back as an IMAGE, ready for the Runware ControlNet builder. No annotator models, no local tiling machinery - the preprocessor half of a serious upscale chain, minus the install.

The inputs

Required: image (IMAGE) - the image you want to keep-faithful-while-detailing.

Optional: outputFormat (JPG default, PNG/WEBP) and the ttl group. Output is a native image (IMAGE).

Where the real controls live is downstream, not here: in the ControlNet builder's weight (the upscale workflow runs tile at low weight - 0.2–0.5 - so the model adds detail instead of copying the source verbatim) and in the model node's steps and resolution. The preprocess node just prepares the condition; the fidelity knob is the weight you send to the sampler.

How it works

Pack-standard: your image is encoded as a PNG data URI, tile processing runs on Runware's GPUs, and the conditioned map comes back. Then: condition → Runware ControlNet builder → model node. For an upscale-style pass, keep the tile weight low and let the model's own steps add the resolution. For style transfer, keep it low and write a strong style prompt - the two together are the whole trick.

Install and gotchas

Install the pack once - ComfyUI Manager, search Runware - or git clone https://github.com/Runware/ComfyUI-Runware into custom_nodes + pip install -r requirements.txt, restart, and set a Runware API key (Settings → Runware API key, RUNWARE_API_KEY, or runware auth login).

The trap most people hit is weight. Crank tile to 1.0 and you get a slightly sharper copy of the source - the model never dares invent. Drop it to the 0.2–0.4 range and suddenly the "upscale" is actually adding detail. And since this is a paid cloud chain, remember you're paying for a generation per pass: do your weight-fiddling on one result before you batch. Tile done right is the closest thing in this pack to a free resolution upgrade - done wrong, it's a very expensive copy command.