Image Transformer - Resize to Max Pixels

Diffusion models have a hard native-resolution ceiling - push SDXL or Flux much past 1024 and you get duplicated anatomy, not detail. The community's answer is the same as it's always been: generate at native size, then upscale afterward. But that means anything you feed in - a reference photo for img2img, a ControlNet source, a face to swap - has to come down to a sane size first. That's the whole job of Image Transformer - Resize to Max Pixels: take whatever monster you dragged in and cap it at a pixel budget without touching the aspect ratio.

It's a one-node pack from teamalpha-ai, a tiny utility that does exactly one thing, and nothing else. No models, no API, no dependencies. It's basically a sharper, VAE-aware cousin of ComfyUI's built-in ImageScaleToTotalPixels.

How it works

The math is refreshingly simple. The node looks at your image's height × width, and if that product beats your max_pixels, it scales both dimensions down by sqrt(max_pixels / current_pixels). Aspect ratio survives, and because it rounds down, the result is guaranteed to sit at or under your cap - never over. If the image is already small enough, nothing happens and the original tensor passes straight through.

Resizing uses ComfyUI's own Lanczos resampler (comfy.utils.common_upscale), which is the right call for downscaling - it keeps more fine detail than bilinear and doesn't mush the image the way a lazy bicubic can. There's no cropping anywhere; it's a pure, ratio-preserving resize.

The inputs that matter

There are only three, and two of them are the story:

• max_pixels (INT) - your pixel budget. This is the one you'll actually tune: 1048576 for a ~1MP cap, 2097152 for 2MP, whatever your VRAM and model tolerate. Set it. The default is 1.
• vae_scale_factor (FLOAT) - snap the result to multiples of this number, so the output plays nice with the VAE. 8 is the standard for SD 1.5, SDXL, and Flux (the value is truncated to an int, so 8.0 and 8 are the same). This applies even when the image was already under the cap - which means you can also use the node purely as a "round these dimensions to multiples of 8" pass by setting max_pixels huge.
• image - any IMAGE tensor, batches fine.

The single IMAGE output wires straight into whatever's next - VAE encode, sampler, ControlNet, another transform.

Installing it

Via ComfyUI Manager, search ComfyUI-ImageTransformer (it's registry-published, so it should pop up). Or the manual route:

cd /path/to/ComfyUI/custom_nodes
git clone https://github.com/teamalpha-ai/comfyui-image-transformer.git

Then restart ComfyUI. That's the entire install - pyproject.toml declares zero Python dependencies and the README confirms it. No model downloads, nothing to cache, nothing that can break your environment. If you find it in a shared workflow and you're security-minded (custom nodes run arbitrary code, and a couple of bad ones slipped through the registry in the past), the whole source is ~60 lines and trivial to read before you run it.

Gotchas

• The default max_pixels is 1. Wire this node in without touching the field and a 1024×1024 image becomes a 1×1 gray smudge. This is the #1 way people get burned - it's a shrink-only node with a self-destruct default. Always set a real budget.
• It never upscales. If your image is under the cap, it passes through. Don't use this expecting a hires pass.
• It's a young pack - v0.0.1, one node, README says "more utilities planned." For most people the built-in ImageScaleToTotalPixels does fine; this one earns its keep when you want the cap and VAE-safe rounding in a single step, or as a cheap pre-downscale before a generative upscaler like SeedVR2 - the community's standard trick for soft, oversized sources is to shrink them first so the restorer works on a sharper-relative base.

Expect it in niche workflows more than in your daily graph. But when you need a one-node ceiling on image size, it does exactly what it says.