DownloadAndLoad CCSRModel

If you're setting up ComfyUI-CCSR fresh, this is the loader you want. DownloadAndLoadCCSRModel gives you a two-option dropdown - real-world_ccsr-fp16.safetensors (the default) or real-world_ccsr-fp32.safetensors - picks one, downloads it for you, and hands back a ready-to-use ccsr_model for CCSR_Upscale. No hunting for a Google Drive link, no guessing which folder it belongs in, no "why is my dropdown empty" moment.

That's the whole point of the node: it's the modern replacement for the pack's older CCSR_Model_Select, which expects you to have already downloaded the original full checkpoint and dropped it into models/checkpoints. The download-and-load node removes both steps.

How it works

Under the hood it's a small snapshot_download from Kijai's HuggingFace repo (Kijai/ccsr-safetensors) using an allow_patterns filter, so it grabs only the one file you chose, not the whole repo. It lands in ComfyUI/models/CCSR, which is also the folder you can drop a safetensors into by hand to skip the download entirely. Then it builds the model from the pack's stage-2 config and loads the weights.

Two details worth knowing. First, the dtype is decided by the filename: pick the fp16 file and you get an fp16 model, pick fp32 and you get fp32. Second, unlike the old loader, it doesn't cache the model on the node instance - each run is a fresh, offload-friendly load, which is why keep_model_loaded on CCSR_Upscale defaults to off.

Which option to pick

The fp16 file, and it isn't close. CCSR's entire problem is memory - the README calls this pack "not very efficient" and the community reports first runs that hammer RAM and VRAM. fp16 halves the download and the VRAM footprint for output that's visually indistinguishable on an upscaler doing 45 steps of noise removal. Reach for fp32 only if you've actually proven a precision artifact and want to chase it, which for this tool basically nobody does.

Installing

Same shared pack install as the rest of ComfyUI-CCSR - ComfyUI Manager (search "ComfyUI-CCSR") or:

cd ComfyUI/custom_nodes
git clone https://github.com/kijai/ComfyUI-CCSR

Restart, let Manager pull the requirements (taming-transformers, omegaconf, einops, pytorch-lightning>=2.2.1, open-clip-torch>=2.23.0 - a heavier stack than most packs, so let Manager handle it).

Gotchas

First run needs an internet connection and a couple of GB of download, so give it a minute on a slow link - it looks frozen, it isn't. If the download keeps failing, grab the file from the HF repo and drop it into ComfyUI/models/CCSR yourself; the node will pick it up and skip straight to loading. And remember the output is a ccsr_model (CCSRMODEL), not an image - it wires into CCSR_Upscale's ccsr_model input, and that's where the upscaling actually happens.