🔍 CR Apply Multi Upscale

Big upscales often go better in stages - a 4x model, then a resample, then another pass - rather than one giant jump. This node runs a whole UPSCALE_STACK in sequence against an image, so a multi-step upscale is one node instead of a chain of them.

It's the applying half of Comfyroll Studio's upscale system (Suzie1 / RockOfFire), paired with CR Multi Upscale Stack where you define the steps. Worth setting expectations up front: this is pixel upscaling - resampling and model-based enlargement - not a generative restorer. It makes the image bigger and cleaner; it doesn't hallucinate new detail the way SUPIR or SeedVR2 do. Know which job you have before you reach for it.

How it works

You build an UPSCALE_STACK (a list of upscale steps - which model or method, what factor) in the stack node, and this one applies each step to the incoming image in order. Between and after model steps it can resample with your chosen method, and it snaps output dimensions to a rounding modulus so downstream nodes don't choke on odd sizes.

The inputs that matter

• image - the picture to enlarge.
• upscale_stack - the UPSCALE_STACK from CR Multi Upscale Stack. This is where the actual steps and factors live.
• resampling_method - lanczos, nearest, bilinear, or bicubic. lanczos is the sharp default and the one you'll usually want; nearest is for hard-edged pixel art.
• supersample (true / false) - render larger internally then scale down for a cleaner result. On costs more compute for smoother output; off is faster.
• rounding_modulus (default 8) - forces output dimensions to a clean multiple. Leave it at 8 unless a downstream node needs a specific alignment; it prevents off-by-a-pixel size mismatches.

Output is the upscaled IMAGE (plus the show_help link).

How to install it

• ComfyUI Manager: search Comfyroll Studio, install, restart.
• Manual: cd ComfyUI/custom_nodes && git clone https://github.com/Suzie1/ComfyUI_Comfyroll_CustomNodes.git, then restart.

Upscale models (ESRGAN-family .pth files like 4x-UltraSharp or Remacri) aren't bundled - drop them in ComfyUI/models/upscale_models for the stack node to find.

Common issues

If the output looks bigger but not better - soft, or plasticky - remember what this is: pixel upscaling can't invent detail that isn't there. On an already-sharp source that's exactly right and gaining favor over generative upscalers. On a soft or tiny source, no amount of Lanczos will rescue it; that's a job for a generative restorer, or the trick of downscaling to a sharper-relative base first and rebuilding from there.

Watch out for stacking too much magnification - a 4x model followed by another 4x is 16x, which balloons memory and can OOM on a modest card. Build the stack to reach your target size, not past it. And if faces are in frame, know that repeated resampling softens them; give a recognizable face its own dedicated pass rather than trusting a blind multi-step chain. If the node itself is missing, that's the Comfyroll pack failing to load, a pack-level problem rather than this node.