Comfyui-Mesh Splits Large AI Models Across Two GPUs Without NVLink

    
        By vramkickedin    
     | 
    
            May 27, 2026 at 12:19 pm        
    
     | 
    
        2 min read

Comfyui-Mesh is a new open-source ComfyUI node pack that splits diffusion model inference across two Nvidia GPUs. The system compresses the model’s internal activations using each GPU’s dedicated NVENC video encoder chip, which normally sits idle during AI work. This lets large models like FLUX.2 Dev or LTX 2.3 run on two consumer cards without needing expensive NVLink bridges.

Developer Shootthesound who recently released ComfyUI-Angelo, built the project to help people who don’t own a single powerful GPU. He created a custom codec that repurposes unused video encoding hardware, shrinking the data that travels between the cards by three to ten times. The goal is to make demanding generative models accessible on mixed or older hardware, saving users from costly upgrades.

Split models across any two Nvidia cards

Key Features

Split FLUX.2 and LTX 2.3 across two GPUs.
NVENC compression reduces wire data by 3–10×.
Works over gigabit Ethernet, VPN, and mobile.
LoRA patches auto-forward to the remote GPU.
Same-machine dual GPU without NVLink.
Live connection indicator and block sync.
Near-lossless quality mode for pixel-perfect output.

The tool is aimed at hobbyists with mismatched Nvidia GPUs and small studios that want to run large models without buying a single high‑end card. It helps privacy‑conscious professionals keep the model’s heavy half on a local server while editing on a weaker laptop. Anyone with spare GPUs can now combine their power to tackle models that previously required an expensive single‑workstation setup.

Developer notes and what's next

Currently, only FLUX.2 and LTX 2.3 models are supported, but Wan, FLUX.1, SD3.5, and others are planned based on community demand. The server handles only one client at a time, and the pipeline does not yet overlap codec work with compute. Neill also has a working proof‑of‑concept for splitting large language models and expects to release that codec separately soon.