Load CCSR Model (TensorRT)

Every custom upscaler in ComfyUI has a loader node, but this one is doing something different than the usual "point me at the weights" dance: there is no full CCSR checkpoint involved at all. Load CCSR Model (TensorRT) is an engine-only loader. The heavy ControlNet+UNet denoise network ships as a prebuilt TensorRT engine, and the only other thing you need is a small aux file with the VAE and cond_encoder weights. The fp16 and even the ConvRot INT8 PyTorch versions of CCSR were dropped from this pack entirely, because on a TensorRT engine the same work runs roughly 1.4x faster and takes less VRAM. For a 2023 upscaler that's normally a sluggish proposition, that's the whole appeal.

It's part of a two-node pipeline in the category CCSR: this loader, then CCSR Upscale (TRT). The loader's job is to hand you a ready-to-run ccsr_model, and it has exactly one input to set: an engine dropdown. That dropdown auto-lists every .rtxplan file it finds in the pack's nodes/CCSR/trt_engines/ folder. If you haven't put anything there yet, the dropdown won't be empty - it'll literally read "(no engine in trt_engines)", which is the node's way of telling you the download step is still on you.

Under the hood the loader deserializes the engine through the TensorRT-RTX runtime (tensorrt_rtx), loads the aux weights into the CCSR stage-2 config in fp16 on your GPU, and wraps the model so that every apply_model call on a 64x64 latent (a 512 px tile) dispatches straight into the engine instead of PyTorch. There's a PyTorch fallback wired in for tiles that don't match that shape, but the upscale node keeps tiles pinned at 512, so in practice you're always on the engine.

The one output, ccsr_model, has the custom type CCSRMODEL - and this matters, because it is not a standard ComfyUI MODEL. You can't route it into a KSampler or any normal model input. Its only consumer is the CCSR Upscale (TRT) node, so wire it straight there. The upside of the custom type is that the loader also frees you from everything that makes CCSR fiddly in its original form: no checkpoint download, no text encoder, no prompt.

To install, search "NunchakuFluxLoraStacker" in ComfyUI Manager (the pack title is bigger than the repo name, but the search works) or clone it:

cd ComfyUI/custom_nodes
git clone https://github.com/ussoewwin/ComfyUI-NunchakuFluxLoraStacker.git

Then do the two-file dance:

cd ComfyUI/custom_nodes/ComfyUI-NunchakuFluxLoraStacker
mkdir -p nodes/CCSR/trt_engines
# download both files into nodes/CCSR/trt_engines/ from:
# https://huggingface.co/ussoewwin/CCSR-ConvRot-INT8-and-TensorRT-Engine

Those two files are ccsr_apply_f16io.rtxplan (the engine) and ccsr_trt_aux.safetensors (VAE + cond_encoder). Restart ComfyUI after dropping them in so the dropdown refreshes. If you installed through Manager, its install.py hook already handled the runtime stack - pinned tensorrt-rtx==1.6.1.120 plus a Windows triton wheel, all --no-deps so it doesn't trample your torch install. If you cloned by hand, run python install.py once in the pack folder.

The gotchas, in order of how often people hit them: this is NVIDIA RTX-only - an .rtxplan engine requires the TensorRT-RTX runtime, so don't expect it to run on AMD or an older non-RTX card. If the loader throws a "engine not found" or "aux weights not found" error, the files aren't where the loader looks (they must sit next to each other inside trt_engines/, not in your models folder). And because this pack is a grab-bag, installing it also pulls in nunchaku and the Florence-2 stack via requirements.txt even if you only wanted the upscaler - expect a heavier install than a two-node pack should be, and wait for it to finish before restarting.

Worth knowing before you build around it: the engine path is the only CCSR this pack ships, so this loader is not optional scaffolding - it's the gatekeeper. Get the two files in the right folder and the whole upscaling pipeline is two nodes and an image. Miss them and the dropdown tells you exactly what's wrong, in plain English.