Bria Increase Resolution

Every upscaler argument in this community is really three arguments. There's "more pixels" - ESRGAN, Lanczos, free and instant but it can't invent anything. There's "more detail" - SeedVR2, SUPIR, the heavy generative stuff that rewrites faces when you're not watching. And then there's "I want it bigger and I don't want to think about VRAM or tiling ever again." That third category is where Bria Increase Resolution lives.

It does exactly what the name says: you feed it an image, tell it 2x or 4x, and Bria's hosted upscaler returns a bigger image that keeps your content intact. No checkpoint download, no GPU usage, no Ultimate SD Upscale tile-juggling. It's one of the Bria partner nodes that shipped in ComfyUI core in August 2026 - your image goes up to Bria's API through your Comfy account and comes back as a result. You pay per call out of prepaid Comfy credits; there are no model files on your machine to place anywhere.

The inputs, and the two that bite

image is your source. desired_increase is a simple dropdown - "2" or "4". That's the happy path.

The node enforces a hard ceiling: the output must fit within 8192 pixels on each side. If your input is big enough that 4x would blow past it, you get an error - unless you flip auto_downscale on. That boolean is the safety net: when the output would exceed the limit, it automatically drops the multiplier (4x becomes 2x) and, if even that isn't enough, downscales the input image first. Leave it off and the node just refuses, which is honestly the right default for a paid call - better to fail before spending credits than to get a surprise.

The other trap is hidden in the pre-processing. Bria first enlarges the short side of your image to at least 224 pixels before upscaling, so a genuinely elongated image can end up un-upscalable at any multiplier - the error message tells you to crop it to a squarer shape first. It's the one failure mode that's about your source rather than your settings. moderation is the same dynamic combo as the rest of the Bria family: it expands into visual-input and visual-output moderation toggles, all off by default.

What comes out

A single IMAGE output. That's it - no mask, no prompt, no progress. Save it, composite it, feed it to whatever comes next in the graph. Because the node polls Bria's status endpoint and downloads the finished result, it behaves like a slightly slow, fully deterministic-feeling image operation: connect it in the middle of a workflow and it just slots in.

When to reach for it

Honestly? When your local options are annoying you. Bria's offering is a hosted, content-preserving upscale - closer to the "more pixels, but smart about it" end of the spectrum than to SeedVR2's detail injection, and without any of the setup. It shines on a laptop with no GPU, or for a handful of hero images where the per-call cost is nothing compared to babysitting a tiled diffusion pass. Where it's the wrong tool: batch jobs, where per-call pricing adds up fast, or anything where you actually want creative detail added rather than the original content kept. The classic gotchas apply too - you need to be logged into a Comfy account with credits on it, and your machine needs to reach comfy.org. If the whole Bria family is cluttering your node search and you never use it, --disable-api-nodes gets rid of them.