ComfyUI-Mesh adds LTX 2.3 support for split-GPU video generation over Ethernet

ComfyUI-Mesh, a third-party ComfyUI extension that splits model inference across multiple GPUs or networked machines, now supports LTX 2.3 Dev and distilled models. The update includes Nvenc codec acceleration for 50-series GPUs and addresses VRAM management bugs that previously prevented ComfyUI from releasing memory blocks on client machines. Users with RTX 50-series cards should select the "Nvenc 5090" codec option for fastest performance, while 30- and 40-series users can try standard Nvenc or Raw modes depending on whether they prioritize speed or pixel-perfect output.

The extension is designed for practitioners who can't run large models on a single machine, not for accelerating workflows that already fit in local VRAM. LTX 2.3's 22-billion-parameter video model typically requires 48+ GB VRAM for full-resolution generation; ComfyUI-Mesh routes inference blocks over Ethernet to secondary machines, making the model accessible to users with consumer-grade hardware. The developer notes that large LTX LoRAs should be loaded in the server app rather than through the node to avoid network round-trips for weight transfers.

The VRAM fix resolves a memory leak where ComfyUI held onto model blocks after inference completed, a problem that compounded across multiple generations and eventually crashed client machines. Raw codec mode bypasses hardware encoding entirely, producing output identical to single-machine runs at the cost of higher network bandwidth — useful for final renders where compression artifacts matter.

The next logical step is multi-node support for FLUX.1 Fill and other inpainting models, which currently lack distributed inference options in the ComfyUI ecosystem. LTX 2.3's open weights and ComfyUI's modular architecture make it a natural testbed for split-GPU workflows; expect similar extensions for Wan and Hunyuan Video once those models see wider adoption.