🔍 AGSoft Upscale Image

Upscaling in ComfyUI is usually a two-node dance: a model loader plus an upscale node, and if you want a plain geometric resize you're reaching for yet another node. 🔍 AGSoft Upscale Image collapses that into one: it can load an upscale model and apply it in the same pass, or skip the model entirely and just do a clean resize. The "None" default means it's also a perfectly good plain resize node - it earns its keep either way.

The bigger lesson from the upscaling community applies here too: decide which job you're doing. If your source already has the detail and you just want more pixels, skip the model and use Lanczos. If you want the model to invent detail, that's where upscale_model earns its keep.

How it works

The node reads your upscale models from the standard ComfyUI/models/upscale_models/ folder - .pth, .pt, .ckpt, .safetensors all show up in the dropdown. Select one and it's loaded once and cached by filename (so re-runs don't reload it every queue). Leave it at None and the node becomes a pure PIL resize.

The pipeline order matters: model upscale first, then your size math on top. Three modes cover the size handling:

• rescale - multiplies the original dimensions by rescale_factor (default 2).
• resize_width / resize_height - set one side via resize_value (default 1024), the other scales to keep proportions.

There's a genuinely nice touch in supersample (on by default): when upscaling, it renders at 2× the target and then downsamples back down - the classic anti-aliasing trick that gives noticeably cleaner edges than a straight single-pass resize. And rounding_modulus (default 8) snaps the output to model-friendly multiples; set 64 for SDXL pipelines.

Inputs and outputs worth knowing

The set is small: image in, upscale_model, mode, rescale_factor/resize_value, resampling_method (Lanczos default), supersample, rounding_modulus, device (auto picks your GPU). Out comes image, plus width and height ints you can feed into downstream latent math or text labels. Batch inputs are handled frame by frame.

Installing it

Pack standard: ComfyUI Manager → search comfyui-AGSoft → install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/Art-xmaster/comfyui-AGSoft.git

No model downloads - the node ships empty and waits for you to add your own .safetensors files to models/upscale_models/. Drop a 4x-UltraSharp or an anime-specific Real-ESRGAN in there and it appears in the dropdown on the next refresh.

Common issues

• Dropdown shows only "None" - your upscale models aren't in models/upscale_models/, or they have a format the node doesn't scan (.onnx won't show). Move them, restart.
• Looks soft after model upscale - you may have gone model → then downscaled via resize_width to something small. The model ran, then the resize undid the detail. Pick one job.
• Odd dimensions down the line - your source was an awkward size and the round to rounding_modulus left leftovers; that's what the rounding_modulus field is for, and it's on by default.

The one thing to keep straight: this node adds pixels, and with a model it can add detail - but a generative upscaler on a face will rewrite the face. If you need identity preserved, treat the face as its own pass, not this node's problem.