SamplerCustom (Sampler Params, Tiled)

You can fit a lot more pixels in a generation than your GPU can hold in memory at once. Tiled sampling is the trick that squares that circle: instead of running diffusion over the whole latent at once, it splits the latent into a grid of overlapping tiles, samples each, and blends them back together. The trade is speed for VRAM - the numbers from the upscaling world hold here too: tiling is much slower but uses a fraction of the memory.

SamplerCustom (Sampler Params, Tiled) is this pack's version of that idea. It's the tiled sibling of the pack's SamplerCustom (Sampler Params), and it's worth being precise about what it does and doesn't do: it wraps the model with a tiling mechanism, then runs the official KSamplerSelect / BasicScheduler / SamplerCustom nodes underneath. No custom sampling math - same samplers, same schedulers, same results you'd get untiled, just processed in spatial chunks.

When it's actually the right tool

The author's own guidance is the honest version, so steal it: the tiled version is useful not just when you're under VRAM pressure, but also when you need output resolutions beyond a model's practical supported size. Verified models are SDXL (with ControlNet Tile) and Anima; the README is upfront that other models may work but compatibility isn't guaranteed. So this isn't a general-purpose "make everything bigger" node - it's a targeted one for those two lineages, and you should treat anything else as a gamble.

For tiled upscaling specifically, the usual warning applies: without a tile-ControlNet condition to anchor each tile to the source, tiles can drift and produce patchwork seams. That's why the author's verified list includes "with ControlNet Tile."

The inputs that matter

The heavy lifting inputs are the same as any sampler: model, positive, negative, latent_image, and sampler_params - the pack's bundle that carries sampler, scheduler, steps, denoise, seed, and CFG in one wire. Then the tiling knobs:

• tile_columns / tile_rows (default 2 / 2) - the grid split.
• tile_overlap (default 128) - overlap between tiles, in pixel-space units. Too little overlap and you'll see tile seams; the code converts this to latent units using the model's spatial downscale ratio.
• mini_unit (default 32) - tile window sizes round up to a multiple of this.

Outputs are output and denoised_output, the same pair stock SamplerCustom gives you.

The honest downsides

Tiling is slow. You're doing redundant work in the overlap regions and loading the model wrapper per tile, so a tiled generation at high resolution can take several times longer than a single pass would have. The payoff is being able to run at all on a small card. And because the wrapper is model-specific (it reads the model's latent downscale ratio), don't be surprised if an exotic model throws an error about the wrapper or produces seams - that's the "compatibility not guaranteed" clause showing up.

Installing

Part of kinorax/comfyui-info-prompt-toolkit:

cd ComfyUI/custom_nodes
git clone https://github.com/kinorax/comfyui-info-prompt-toolkit.git
cd comfyui-info-prompt-toolkit
pip install -r requirements.txt

or install "Info-Prompt-Toolkit" via ComfyUI Manager, then restart. No model files. You'll also want the pack's Sampler Params node (or Sampler Params Context) upstream to build the sampler_params bundle this node consumes - they're in the same pack and install together.