ComfyUI-Mesh distributes FLUX 2 across networked GPUs—14 seconds per megapixel image
A new ComfyUI custom node splits diffusion models across machines over Ethernet, enabling fast generation without NVLink. A 5090 desktop plus 4090 laptop generates 1MP FLUX 2 Dev images in 14 seconds over gigabit Ethernet.
ComfyUI-Mesh is a custom node that distributes large diffusion models across multiple GPUs on separate machines using standard Ethernet or Wi-Fi, eliminating the need for NVLink. The node currently supports FLUX 2 Dev, Klein 9B, and LTX 2.3, with planned support for Wan, Qwen, and Chroma. A desktop RTX 5090 paired with a laptop RTX 4090 over gigabit Ethernet generates 1-megapixel FLUX 2 Dev images in 14 seconds; faster network links achieve 4.4 seconds per image.
The system works by encoding model layer outputs with a custom NVENC-based codec and streaming them between machines. Any Nvidia card with NVENC support can participate. The developer tested the setup over mobile tethering—70 percent of the model on a home desktop, 30 percent on a laptop in a café, connected via Tailscale VPN—and generated 1-megapixel images in under 8 seconds. Wi-Fi 6 performs well; multi-GPU setups within a single PC are also supported.
Network optimization
The node transmits active LoRA weights from client to server automatically, though performance improves when LoRAs are pre-loaded server-side and selected remotely. For FLUX 2 Dev with the 2.5GB turbo LoRA, the recommended workflow is to load the LoRA in the server app's LoRA field and place it to the right of the Icarus node in the ComfyUI graph, avoiding duplicate weight transfers over the network. This configuration enables 1MP FLUX 2 Dev generation in 14 seconds over 1GB Ethernet.
The underlying codec has its own repository and a parallel version for splitting 32B and 70B LLM models across two machines, with release planned for the following week. Multi-node architecture foundations are in place; future updates will add multi-LoRA handling, client-side remote LoRA selection, and smarter server-side LoRA management. Full setup instructions and source code are available on GitHub.