🔍 Upscaler Advanced

Want to upscale a 1024px image to something print-sized on a 6GB card? The trick, as old as tiled upscaling itself, is to not do it all at once: cut the target into overlapping tiles, process each one, stitch them back with feathered seams. The hard part has always been the boring arithmetic - how many tiles, how much overlap, what shape, whether the latent matches the pixel grid. FiL Upscaler Advanced is a tile-grid calculator that does all of it, and if you connect an upscale model, it'll actually upscale the pixels too.

It's the anchor of the FiL Design ImageMind pack's upscale pipeline: plan the grid here, process the tiles, recombine with the pack's Tile Assembly node.

How it works

Give it an image (and optionally a latent to resize and tile 1:1 alongside it), set an upscale_factor, and it computes a tile-aligned target size, then lays out the grid. The geometry is real, not cosmetic: target dimensions are aligned to a 64px grid, the grid step is tile − overlap, overlap is clamped at half a tile so the tile count can't explode, and edge tiles shift inward to stay full-size instead of being zero-padded - so no black strips along the edges. If you feed a latent, it's resized with bislerp, which is the correct choice there - lanczos is an RGB-specific interpolation and doesn't belong in latent space.

The inputs that matter:

• upscale_factor - 2.0 doubles the size. This is the master dial.
• tile_size / tile_overlap - the manual knobs. Bigger tiles = fewer seams but more VRAM per pass; more overlap = cleaner blends but more tiles. auto_overlap derives overlap from tile size (~12.5%) so it stays proportional when you change tile_size.
• auto_mode + auto_profile - the lazy path: pick Low VRAM, Balanced, High VRAM, Max Quality, or Ultra Quality, and it chooses tile size, overlap, and even a denoise hint for your downstream sampler.
• upscale_model - optional. Connect an ESRGAN-style model from your upscale_models folder and the image output is actually upscaled through it. Leave it disconnected and the node becomes calculator-only - image passes through unchanged.

That last one is the key insight: the node is useful in both modes. With no model it's a planner that emits a full numeric plan for downstream nodes; with a model it's a real upscaler.

The outputs you'll actually wire

21 outputs sounds intimidating; you use about four. tiles is a batch of the actual cropped pieces - feed those (or latent_tiles) into per-tile processing, which is where the tiled-upscale magic happens. layout is the per-tile rectangle map, and it's the one you must feed into FiL Tile Assembly to recombine - that's what tells the assembler where the feathered seams go, so always use the same layout that produced the tiles. The numeric outputs (upscale_by, denoise, tile_width, tile_height, overlap, tile_count) exist so you can feed them into external tiled-sampler stacks - Ultimate SD Upscale-style downstream nodes read exactly this kind of plan. And warnings tells you when the grid left a thin edge tile, which is why auto_fix_thin_edges exists.

If you ever want to do this whole thing with a tile ControlNet keeping each tile faithful - the classic "tile diffusion" workflow - this planner is the grid half; your ControlNet does the faithful half.

Installing it

Ships in FiL Design ImageMind (MIT). ComfyUI Manager: search FiL_Design_ImageMind, or:

cd ComfyUI/custom_nodes
git clone https://github.com/FiL-Design-Ai/FiL_LLM FiL_Design_ImageMind
pip install -r FiL_Design_ImageMind/requirements.txt

Restart; it's under 🎨 FiL Design/Image. ComfyUI 0.3.60+, Python 3.10+, light dependencies. It doesn't ship upscale models - it reads your existing models/upscale_models folder - so grab a 4x ESRGAN (4x-UltraSharp or a Remacri-class model are the usual picks) if you want actual pixel upscaling.

Gotchas

The number-one confusion: "the tile count looks wrong." It's usually right - overlap grows the count by design, since step = tile − overlap. If you want a fixed count, pin manual_tile_cols and manual_tile_rows and the grid holds exactly that, growing the tile instead. Second: at least one of image or latent is required, or the node raises - source dimensions come from whichever you connected. And remember: no upscale_model, no actual upscaling. The most common "this node does nothing" report is really "I forgot to connect the model."