High Resolution Tile Split & VAE Encode

Tiled diffusion is how you do detail-focused high-res work on a GPU that can't see the whole canvas: you slice the image into overlapping chunks, sample each one, and stitch them back. This node is the front half of that pipeline - it takes one pre-upscaled image, splits it into overlapping tiles, and VAE-encodes every tile sequentially so the sampler can work on them one at a time. Its partner, UC_HighResolutionTileAccumulator, does the stitching.

The reason this whole approach exists is VRAM. A 4K upscale is not a thing a 12 GB card does in one pass, but four 1024px tiles it can handle. The catch - and it's the classic tiled-diffusion failure mode - is seams: tiles diverge and you get patchwork. That's what the overlap and mask machinery here is for.

How it works

You give it an image and a VAE. It builds a tile grid in one of two modes:

• tile_size mode uses tile_width/tile_height (both default 1024) to size each tile.
• grid mode uses rows/columns instead. Controls for the mode you're not using are ignored.

Every neighboring pair of tiles shares overlap pixels (default 128). The overlap region is where the magic happens: each tile gets a soft mask - mask_profile defaults to cosine - so the shared zone is weighted rather than hard-cut, and feather_width/mask_strength control how gradual and how protected that transition is. Tile latents come out with Core-compatible noise masks, which is how the sampler knows to denoise the overlap gently.

If you're doing guided tiled upscaling, there's a bonus: optional Differential Diffusion via model + differential_diffusion_mode (off/core/advanced), and an optional depth_map whose depth_influence steers structure preservation across the feathered zones. Skip all of it and the node still works fine - those are depth-of-field extras, not requirements.

The inputs that matter

The ones a beginner actually sets: image, vae, tile_width/tile_height (or rows/columns), and overlap. Bump overlap for bigger transitions (slower, fewer seams); drop it for speed. mask_profile of cosine is a safe default.

Outputs

• tile images - the exact padded tensors sent to VAE encode, for spatially matched visual conditioning.
• tile latents - matching VAE latents with the noise masks.
• tile layout - the coordinate/overlap metadata that the accumulator needs to put it back together.
• model - the Differential Diffusion model for the sampler guider, or your unchanged model when diff-diff is off.

Installing it

ComfyUI Manager → search "ComfyUI-UtilsCollection", or:

cd ComfyUI/custom_nodes
git clone https://github.com/silveroxides/ComfyUI-UtilsCollection

Restart. Requirements are just opencv-python and typing-extensions - no model downloads. Newer-API pack, so keep ComfyUI updated.

Where people get burned: forgetting that the sampler must run in list-mapped mode so each tile gets its own encode/decode, and dialing overlap to zero for speed, which guarantees seams. Keep some overlap. And unlike old tiled-upscale stacks, there's no ControlNet Tile preprocessor wired in by default - this is a straight split/sample/accumulate pipeline, so the model's own prompt carries the detail work.