Dickson_TTP_Preprocessor_Simple

Tile ControlNet doesn't want your raw image. It wants a flattened version of it - the source detail smeared away enough that the model invents fresh detail instead of just re-rendering the same pixels. That preprocessing step is what the TTPlanet preprocessors exist for, and this one, Simple, is the fastest and least fiddly of the three in this pack. If you're doing tile-based upscaling on SDXL and you just want the conditioning image to look right with the least knob-twiddling, this is the one I'd reach for.

How it works

It's a short pipeline with no surprises: take the image, downscale it by scale_factor with area interpolation, upscale it back to the original size, then apply a Gaussian blur. The down-then-up trick is what flattens the detail - the model sees the composition and colors but not the grain. The blur then knocks down any remaining sharpness. Output is a single IMAGE, image_output, ready to feed into a ControlNet Apply node alongside the TTPlanet tile model.

One deliberate difference from its sibling: Simple blurs after resizing, so the blur operates at full resolution. The GF version flips the order and adds an edge-preserving guided filter on top. Simple is the one without that extra filter - fewer controls, faster, and fine for most upscale workflows.

Inputs that matter

• image - the source you're upscaling or re-detailing.
• scale_factor (1–8, default 2) - how far down it squashes the image before bringing it back. Higher = more detail erased = the model gets freer with new texture. Default 2 is a sensible starting point.
• blur_strength (1–20, default 1) - how hard the final Gaussian blur hits. Nudge it up when you're seeing the model copy source noise.

Output: image_output → ControlNet Apply (with the TTPlanet tile model loaded), then a VAE Encode so the regenerated detail actually lands in your sampling.

Install and the model you can't skip

Install from ComfyUI Manager (search dicksondickson) or:

cd ComfyUI/custom_nodes
git clone https://github.com/dicksondickson/ComfyUI-Dickson-Nodes

Then restart. The pack's requirements.txt is just Pillow and opencv-python-headless, so no heavy dependencies - but the README's most important line is about the model. These preprocessors are tuned for TTPlanet's SDXL tile ControlNet, which does not auto-download. Grab TTPlanet/TTPLanet_SDXL_Controlnet_Tile_Realistic from Hugging Face and drop the files into ComfyUI/models/controlnet. Forgetting that step is the #1 way beginners build a graph that produces nothing.

Two honest caveats: this is SDXL-era tooling - the preprocessor and the paired model are for SDXL base models, not Flux or the newer DiT models. And "Simple" is relative: the pack's own cufoff variant is broken in the shipped code, so this is effectively your baseline option. Keep the default scale_factor at 2, tune the blur only if the output looks noisy, and it'll behave.