HSWQ SAM3 Loader (ConvRot INT8)

Two names, one class

If you landed here because a workflow JSON, a Reddit comment or your own node search said HSWQLoadConvRotINT8SAM3, the short answer is: it's the same node as HSWQSAM3Loader, wearing its old name. In the source it's literally one line -

HSWQLoadConvRotINT8SAM3 = HSWQSAM3Loader

and the pack's __init__.py registers both keys in NODE_CLASS_MAPPINGS, with the same display name, "HSWQ SAM3 Loader (ConvRot INT8)", and the same category, loaders.

Why anyone does this

ComfyUI saves workflows by class name, not by the label you see on the canvas. So an alias is a compatibility hatch: workflows saved when the node only answered to HSWQLoadConvRotINT8SAM3 still resolve today, and workflows saved against the shorter HSWQSAM3Loader name resolve too. Nothing in the graph has to be re-wired, and nothing about behaviour changes - same sam3_name dropdown, same single MODEL output, same INT8-or-fp16 detection inside.

The naming itself is a tic of this pack, not a mistake. Its other ConvRot INT8 loaders follow the HSWQLoadConvRotINT8… pattern (the ControlNet loader ships the same way, as HSWQControlNetLoader with a HSWQLoadConvRotINT8ControlNet alias), and the SAM3 one was written in that style before getting the shorter, more searchable name HSWQSAM3Loader. Both were kept.

The version-history trap worth knowing

There is one situation where neither name works, and it's a pack-version problem rather than a your-install problem.

The SAM3 nodes were removed in v3.4.8 - the author's finding then was that the pack's startup INT8 patch already lets stock loaders (CheckpointLoaderSimple, the default Comfy SAM3.1 node) load ConvRot INT8 SAM3 checkpoints on their own, so the dedicated nodes weren't pulling their weight. They came back in v3.5.9, restored faithfully from the v3.4.7 git history, with both class names and the png/sam3.png illustration.

So: if a workflow throws a red "missing node" box for HSWQLoadConvRotINT8SAM3, check which build you're on before you start hunting for a repo. Anything from v3.4.8 up to v3.5.8 doesn't have it - at all, under any name. Update the pack via ComfyUI Manager and the node reappears.

There's a milder, more common symptom too: because both keys carry an identical display name, the node search dialog can list "HSWQ SAM3 Loader (ConvRot INT8)" twice when you filter by sam3. That's not a duplicate install, it's the alias. Pick either; the graphs are interchangeable.

Using it

Since it's one class, this is the same node documented under the HSWQSAM3Loader name - the full walkthrough of the INT8 load path, the ConvRot unaligned-GEMM fallback and the CLIP key remap lives on that page. The two-minute version:

Input: sam3_name, a dropdown populated from .safetensors files found in diffusion_models, sams, detection and checkpoints. File in a different folder, no dropdown entry.
Output: model (MODEL). Feed it to HSWQ SAM3 Detect, or to stock SAM3 nodes - a ConvRot / TensorWise INT8 SAM3 checkpoint loads with its Linear layers kept 8-bit in VRAM (MixedPrecisionOps + int8_tensorwise), while a plain fp16 SAM3 file loads normally with no quant handling attached.
Search terms that find it: sam3, sam3 loader, load sam3, sam 3, segment anything, hswq sam3, convrot sam3, int8 sam3. Typing "convrot sam3" is how most people arrive.

Install is the pack install - ComfyUI Manager, search ComfyUI-HSWQ-Loader-and-Tools, or:

cd ComfyUI/custom_nodes
git clone https://github.com/ussoewwin/ComfyUI-HSWQ-Loader-and-Tools
# restart ComfyUI

Troubleshooting, alias edition

Node missing in a workflow you downloaded. Check the pack version, not the spelling. Pre-v3.5.9 builds genuinely don't ship it.
Both names visible in search. Expected. Identical nodes.
You edited a workflow by hand and typed the wrong one. Both are valid; there's no wrong one to type as long as the pack is current.
You don't want the pack at all. Reasonable - the startup patch means stock CheckpointLoaderSimple handles these checkpoints for many users, which is the author's own conclusion from the v3.4.8 removal.