VRAM Pager brings compressed ComfyUI paging

This looks like the diffusion/video equivalent of the LLM community’s compressed offload tricks: not a new model, but a practical memory-path optimization that attacks PCIe bottlenecks directly.

• –The key idea is bandwidth-efficient paging, not just better swapping, so it should matter most when models are too large for VRAM but still need to stay in native precision.
• –The repo’s own benchmark claims are strong for a niche pain point: 3.4x faster INT8 transfers, 1.8x faster lossless FP16 transfers, and a Wan 2.2 14B run that finished where the standard low-VRAM path crashed.
• –It is most relevant for ComfyUI users running unquantized FP16/FP32/BF16 safetensors on 16GB-class GPUs; the project explicitly says GGUF users probably won’t gain much.
• –If the numbers hold up across more models, this could become a standard escape hatch for image/video workflows that currently rely on blunt `--lowvram` tradeoffs.
• –The biggest question is breadth: how many diffusion/video architectures and custom nodes can tolerate this paging layer without edge-case regressions?