ComfyUI nodes load INT8-quantized SAM3/3.1 into VRAM with Tensor Core speed
Zestyclose_Bake3680 · reddit · 2026-08-30
The open-source ComfyUI-HSWQ-Loader-and-Tools adds an HSWQ SAM3 Loader (ConvRot INT8) and HSWQ SAM3 Detect node, keeping ConvRot/TensorWise INT8-quantized Segment Anything 3/3.1 weights resident in VRAM at 8-bit precision to cut memory usage.
- Execution uses comfykitchen's fast int8linear GEMM kernel with online activation rotation for ConvRot layers.
- Unaligned layers (K not a multiple of 4, e.g. boxRPBembedx) automatically fall back to float, while heavy backbone/transformer blocks run accelerated INT8 Tensor Core math.
- Detect supports text conditioning, bboxes, point prompts as JSON, thresholds, and refinement iterations, outputting masks and bboxes.
- Standard FP16 checkpoints load unchanged with identical results; example workflows included in the repo.
More from Multimodal
- The math: 5 rerolls at 480p plus a 4K upscale costs $5 vs $9 at 720p — HeyAmit_ · 2026-10-15
- Dreamina workflow: generate at 480p, upscale to 4K, cut video costs 44% — HeyAmit_ · 2026-10-15
- One-prompt GPT-6 Astra writes two Palestrina-style motets in LilyPond — deanwball · 2026-10-15
- DiffSynth-Studio Adds OPD Workflow: Few-Step Diffusion Training in ~2 GPU-Hours — bdsqlsz · 2026-10-15
- AI film and TV hit the market: Prime Video picks up AI spin-off, pilots screen at Mipcom — lmoroney · 2026-10-15
- Ramayana film taps DNEG's massive VFX and AI-assisted production for global release — CurieuxExplorer · 2026-10-15