🔍 Upscaler Simple

Upscaler Simple is the trimmed-down sibling of FiL_Design_ImageMind's Upscaler Advanced: identical tiling panel, identical math, but only the five outputs most graphs actually use. If you've ever wanted the Ultimate SD Upscale-style tiled workflow without bolting a whole other pack onto ComfyUI, this is the on-ramp - it plans a tile grid over your image, optionally upscales through a real ESRGAN-style model, and hands you the pieces.

The key thing to understand is what it is not. This is not a "make the image bigger, magic, done" node. It's a tile-grid planner. It splits your image into overlapping tiles so you can do the heavy lifting (a diffusion pass, a per-tile upscaler) on each tile within your VRAM budget, then recombine them. The whole point of tiling is arbitrary output sizes on a card that would never hold the full image at once - the same reason Tiled Diffusion and Ultimate SD Upscale exist.

The inputs that matter

• upscale_factor (2.0) - how much bigger. Range 0.1–8.0.
• tile_size (1024) and tile_overlap (64) - the grid geometry. More overlap means fewer seams and more VRAM. auto_overlap derives overlap from tile size (~12.5%, min 32px) so it stays proportional when you change tile_size.
• auto_mode + auto_profile - full auto picks tile size and overlap from a VRAM profile (Low VRAM → Ultra Quality). This is the "I just want it to work" path.
• manual_tile_cols / manual_tile_rows - force an exact grid; 0 means derive it from tile size.
• image, latent, upscale_model - at least one of image/latent required. upscale_model is an ESRGAN-style model loader and only applies when an image is connected.

The geometry deserves one note because it's where people get confused. Overlap isn't padding - the grid step is step = tile − overlap, so overlap genuinely increases the tile count. If the count looks wrong, that's why. Clamping keeps overlap at half a tile so the count can't explode, and non_square_tiles allows rectangular tiles within a 1.5:1 aspect.

The five outputs, and how they fit together

• image - the image upscaled through the model (if one is connected) and resized to the tile-aligned target.
• tiles - the actual cropped tile batch. Edge tiles shift inward to stay full-size, so no black padding strips.
• latent / latent_tiles - the latent resized with bislerp (not lanczos - that's RGB-only interpolation and doesn't belong in latent space) and one latent crop per image tile.
• layout - the per-tile positions. This is the one to keep track of. It's what FiL Tile Assembly consumes to stitch everything back together.

The intended pipeline: Upscaler Simple → process tiles (or latent_tiles) individually - upscale each with a diffusion pass, run ControlNet Tile, whatever - then feed the processed tiles plus the same layout into FiL Tile Assembly. Keep the layout matched to the tiles; mixing a layout from one grid with tiles from another is how you get seams in the wrong places.

Two honest notes. Without an upscale_model connected, image is just a passthrough - the node plans and crops but doesn't add pixels. And in latent-only mode the image/tiles outputs are empty placeholders, which trips people who wire them into a preview and see nothing.

The community's standing advice applies here too: decide whether you want more pixels (a plain ESRGAN pass - fast, can't hallucinate) or more detail (generative, tile-based, needs ControlNet Tile to stop tiles diverging). This node supports both paths; the difference is what you put between the tiles and the assembler.

Install is pack-standard: ComfyUI Manager search FiL_Design_ImageMind, or clone + pip install -r requirements.txt + restart. No model downloads ship with the pack - an upscale model comes from your own models/upscale_models.