comfyui-sealedcuda

ComfyUI-SealedCuda

A minimal sealed-worker ComfyUI custom node that runs prebuilt CUDA wheels in an isolated environment with its own torch + CUDA stack, decoupled from the host ComfyUI torch. The GPU work happens in the sealed worker; tensors travel to and from the host over the isolation RPC, exactly like a host-coupled node.

The native dependencies come from prebuilt Windows + Linux wheels, so the sealed environment provisions without a local compiler:

• torch==2.8.0+cu126 from the PyTorch cu126 index
• natten==0.21.6+cu126torch2.8 and fused_ssim==0.0.0+cu126torch2.8 from the cuda-wheels index

Nodes

• Sealed NATTEN Filter (CUDA) — IMAGE -> IMAGE. Projects the image into an embedding, applies 2D neighborhood attention, projects back, and adds the result as a residual. A deterministic (seeded) GPU image filter.
• Sealed Fused-SSIM (CUDA) — IMAGE, IMAGE -> FLOAT. Structural-similarity score between two images via the fused-ssim CUDA kernel.

Requirements

A CUDA GPU with compute capability supported by cu126 (sm_50–sm_90). NATTEN's attention backends need an embedding dimension whose per-head size is at least 16, so embed_dim / num_heads >= 16.