Load LDSR Model

There's no button to press here. LDSRModelLoader does exactly one job: it looks at the files sitting in ComfyUI/models/upscale_models, picks one, and hands it out as an UPSCALE_MODEL socket for the LDSR Upscale node to run. The interesting part isn't the node - it's what it loads, and that's where most people get stuck before they ever upscale a single image.

What you're actually loading

That last.ckpt in the dropdown isn't some random model. It's the original Latent Diffusion Super Resolution (LDSR) checkpoint - a 4x diffusion upscaler from the High-Resolution Image Synthesis with Latent Diffusion Models paper (Rombach et al., 2022), made by the CompVis group at LMU Munich. Yes, the same group whose latent-diffusion architecture became Stable Diffusion, and whose founders later went on to Black Forest Labs. LDSR is the family's 2022-era upscaler: a full diffusion model, not a one-pass ESRGAN-style CNN, which is why the file is a hefty multi-gigabyte checkpoint and why it needs a diffusion sampler at runtime instead of a fast filter.

The checkpoint itself lives on a Heidelberg University server (HeiBox), not on Hugging Face or CivitAI. The dropdown lists whatever .ckpt/.safetensors files are in models/upscale_models, and last.ckpt is the default because that's literally the name of the file the download drops on you.

Loader vs. all-in-one

This is the "proper" half of a two-node split: load the model once, wire the UPSCALE_MODEL output into LDSR Upscale. The LDSR Upscale (all-in-one) node skips the loader entirely and does load-plus-upscale in one step.

The one real difference: the loader parks the model on the CPU after loading (the source calls model['model'].cpu() right after instantiation), so a loaded LDSR model doesn't hold your VRAM hostage between runs. The all-in-one loads straight onto the GPU every time you execute it.

Installing it - the part that trips people up

You can grab the pack through ComfyUI Manager by searching "LDSR" or "ComfyUI-Flowty-LDSR", or do it by hand:

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

Then the model, which is the step everyone forgets:

# from the pack README - this is the original LDSR checkpoint
# save the download as: ComfyUI/models/upscale_models/last.ckpt

Download it from the HeiBox link in the README, drop it in upscale_models, and restart ComfyUI.

Watch the requirements: the pack pins pytorch-lightning and torchmetrics==0.11.4, on top of torch/torchvision/einops/omegaconf. That's a chunky dependency list for what is, mechanically, one loader node - and it's the source of the most common failure below.

Troubleshooting

• "No module named 'pytorch_lightning'" when the node runs. This is the classic. The pack's requirements.txt got installed into your system Python instead of the environment ComfyUI actually uses - a near-certainty on the Windows portable build. Install the requirements into ComfyUI's own python_embedded (or your venv), not wherever pip points by default.
• The dropdown is empty or the node errors on load. The file isn't in upscale_models, or you renamed it and the default last.ckpt can't resolve. It's an enum over the folder, so just select the file that's really there - once it exists, the node can't miss it.
• The torchmetrics==0.11.4 pin is old. If another pack in your install wants a newer torchmetrics, you can get dependency whiplash. Usually harmless, occasionally the cause of a bizarre import error elsewhere in ComfyUI.

Honest verdict: the loader is a formality - there are no settings to tune, and it will never be the reason your upscale looks bad. The actual decision you're making by installing this pack is "do I want to run a 2022 diffusion upscaler at all." More on that call in the all-in-one article.