CCSR

CCSR (Conditioning with Contrastive supervision for image Super-Resolution) is the 2023-era diffusion upscaler that everybody respected and almost nobody ran, because it was a VRAM monster. Where Real-ESRGAN is a pixel upscaler that adds resolution, CCSR is a generative restorer: it reconstructs plausible high-frequency detail - texture, sharp edges, recovered detail that was never really in the source. That's a different job, and it's the one people usually actually mean when they say "make this image look better." The historical problem was that local CCSR wanted a serious card (12GB+ and slow), which is exactly the kind of thing that ends up serving a niche. This node puts it in the cloud: feed it an image, get a restored, upscaled IMAGE back.

The key mental model, straight from the upscaling essay: this is a detail tool, not a pixels tool. Use it on soft or damaged images where you want the model to invent plausible detail. Use Real-ESRGAN (also in this pack) when the source is already sharp and you just want it bigger. And know that, like every generative restorer, it will rewrite fine features - don't point it at a face you need to recognize and expect identity to survive.

How it works

Another upscale-classified node in the pack: it sends the image to Runware, the API runs CCSR's multi-scale diffusion restoration, and the result comes back as a native IMAGE tensor. What makes this node different from the Real-ESRGAN node is the controls - CCSR exposes diffusion-style settings because it's fundamentally a diffusion process:

• image (required) - the input tensor.
• upscaleFactor - (default) or 2. Note there's no 4× here; CCSR's default restoration scale is around 2×, and pushing a diffusion restorer past its comfortable factor is how you get artifacts.
• settings.CFGScale - an off-by-default gate with CFGScale_value. Higher = output sticks closer to the prompt; for pure restoration you usually want it low. Leave the gate off and let the model default unless you know why you're changing it.
• settings.controlNetWeight - −4 to 4, default 1. This is how hard the ControlNet-style conditioning (the "keep the source structure" constraint) is applied. Lower it to let the model stray further toward its reconstruction; negative values steer away from the guide. At 0 it's pure generation from the prompt.
• settings.positivePrompt / settings.negativePrompt - yes, an upscaler with prompts. CCSR conditions on a text description of what the result should look like. Empty positive prompt means "restore this faithfully"; a prompt like "highly detailed, sharp, clean" can push texture quality. The negative prompt is ignored when guidance is disabled (CFG ≤ 1), per the tooltip.
• settings.steps - off-by-default gate; steps_value 20–60 (default 20). More steps = slower but more thorough restoration.
• settings.seed - gate + seed_value for reproducible runs.
• outputFormat / outputQuality - JPG/PNG/WEBP and compression quality.

The output is image → Preview or Save.

Installing

cd ComfyUI/custom_nodes
git clone https://github.com/Runware/ComfyUI-Runware
pip install -r ComfyUI-Runware/requirements.txt

Or ComfyUI Manager → search Runware → install → restart. API key via Settings, RUNWARE_API_KEY, or runware auth login. No local weights - the VRAM tax is gone, which is the entire appeal of the cloud version.

Where people get burned

Two classic mistakes. First, treating it like a pixel upscaler and being disappointed that it "changed" the image - it's supposed to; it's reconstructing detail, and that's the point. Second, running it on already-clean images, where restoration adds unwanted texture and detail noise (the exact complaint the KB logs against SUPIR for the same category). If the source is clean and sharp, Real-ESRGAN is the right node. CCSR is for the soft, damaged, compressed stuff - and now you can actually afford to run it.