TT Quick Image Upscaler

Sometimes you don't want a grading suite and film grain bolted onto your upscaler. Sometimes you just want to make the image bigger. TT Quick Image Upscaler is the TenserTensor node stripped to exactly that: an image in, a model picked, a bigger image out. Five inputs, one output, no decoration. If you're building a hires-fix pass or upscaling a finished render, this is the one you'd actually reach for.

The mechanism is the same engine the pack's bigger postproduction nodes use, just exposed plainly. It loads an upscale model with spandrel - the same loader ComfyUI core uses, so every ESRGAN/SwinIR-family .safetensors or .pt in models/upscale_models/ works - and runs it tiled. The tile (default 512) and overlap (default 64) inputs control how the image is chopped up and re-stitched. That tiling is what lets a 4x ESRGAN model upscale a 2K render on a 6–8GB card: it never holds the whole image in VRAM, only the tile plus its overlap margin. If it hits an out-of-memory anyway, it quietly retries with smaller tiles instead of crashing - a nice touch you don't get from every upscaler node.

The only inputs:

• image - what you're upscaling.
• upscale_model_name - which model. No model selected means no upscale.
• tile and overlap - VRAM vs. seam-quality tradeoff. Smaller tiles fit tighter VRAM but need more overlap to avoid visible seams; 512/64 is a safe starting point.
• device - default (GPU) or cpu. Pinning to CPU frees VRAM for a sampler running in the same graph.

Output is a single IMAGE, ready to feed a save node or chain into another pass.

Install is the pack standard:

cd ComfyUI/custom_nodes
git clone https://github.com/tenser-tensor/ComfyUI-TenserTensor

or search "TenserTensor" in ComfyUI Manager and restart.

Two honest expectations to set. First, this is a "more pixels" upscaler, not a "more detail" one - per the KB's upscaling taxonomy, an ESRGAN model cleans up and enlarges what's there but won't invent the fine detail a generative restorer like SeedVR2 would. For sharpening an already-good image, it's exactly right; for rescuing a soft one, it isn't the tool. Second, the dependency story is the pack's usual one: declared deps are gguf and kornia, but this node's code also uses spandrel, which ComfyUI core already bundles.

Common failure and fix: upscale_model_name is empty → "Failed to load model file" or the node just returns nothing. Drop a real model into ComfyUI/models/upscale_models/, restart, and pick it. 4x-UltraSharp, Remacri, and the RealESRGAN models are all fair game. And if tiles leave faint grid lines, bump overlap toward 128 before touching tile.

Last note, same as every node in this pack: the V1 class is marked deprecated while the author migrates to API V3, with TT_QuickImageUpscalerNode as the maintained successor. Fine to use today; new graphs should prefer the V3 version.