SAM Loader (not shipped)

SAM Loader (not shipped) is the other half of a pair that doesn't work. Radiance 3.5.0 ships no SAM runtime, so this loader has nothing to load a model into, and its one input says as much: model_name is documented as "Not used: no SAM runtime ships with Radiance 3.5.0." It's hidden from the node menu, it still opens in saved graphs, and it's kept around purely for compatibility.

The pack's known-issues list is specific about the plan: either a real SAM2 backend or removal in 4.0. Until then, treat both SAM nodes as furniture.

What the schema still shows

Worth a look only because it tells you what the author was aiming at. model_name is a STRING. device is a cuda / cpu / mps choice, defaulting to cuda - the familiar three-way, and mps is the Apple Silicon path if you're on a Mac. offload_to_cpu is a boolean, the flag you'd use to keep the model off the GPU between runs so a big checkpoint isn't sitting in VRAM while the diffusion model works. dtype is float16 / bfloat16 / float32, defaulting to float16.

That's a normal, sensible loader interface. It just has no backend behind it, and the single output, sam_model, feeds a generator that also does nothing. There is no code path where either node produces an image.

Use a real SAM node instead

The KB's SAM panel is the honest map: SAM and SAM 2 are Apache-2.0 and in wide use, SAM 3 (November 2025) added promptable concept segmentation and is the tier most 2026 mask editors build on. ComfyUI's built-in SAM nodes and the mainstream segmentation packs cover point, box and text prompting - and if you want a foreground matte rather than a specific object, a background-removal model is the better tool entirely (background-removal.md).

On licensing, one thing to carry forward: SAM 3 isn't Apache - it ships under Meta's SAM Licence with its own acceptable-use terms - so if masks feed into something you distribute, check that before you commit to it.

Install, for the rest of the pack

Radiance is still the best thing in ComfyUI for EXR/DPX I/O, OCIO colour management, HDR review and Nuke handoff, and those nodes ship fine.

ComfyUI Manager → search Radiance → install → restart → refresh. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/fxtd-studios/radiance.git
cd radiance
python -m pip install -r requirements.txt

Windows portable users: run that pip line with python_embeded\python.exe, or the imports won't be visible to ComfyUI. This is the pack where "which Python did pip use" is genuinely the number one cause of a bad install, because OpenEXR, OpenImageIO and opencolorio are all compiled extensions rather than pure Python.

Bottom line

Two nodes, zero segmentation. If an old workflow of yours is stopping on this node, delete it and rebuild the mask with a maintained SAM implementation - the rest of the graph is unaffected. If you're here because you searched for "SAM loader" and hoped Radiance had one, it doesn't, and the display name is the pack telling you so before you waste the evening.