shootthesound

shootthesound / comfyui-mesh

Public

Split FLUX.2 across two GPUs (LAN or same-machine) — NVENC compresses activations live on the wire. Icarus (ComfyUI node) + Daedalus (back-half server).

48
4
89% credibility
Found May 17, 2026 at 52 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

ComfyUI-Mesh is a tool that lets you split large AI image generation models across two GPUs — either in the same computer or over a network. The front half of the model runs on your ComfyUI machine while the back half runs on a second machine. NVIDIA's normally-idle video encoding hardware compresses the data between them, so even a regular home network becomes fast enough for real-time generation. It currently supports FLUX.2 Dev and FLUX.2 Klein 9B models, with plans to add more architectures.

How It Works

1
💡 You discover the project

You learn that you can split a large AI image model across two GPUs, even over a home network, using idle video encoding hardware to compress the data between them.

2
🖥️ You install the client on your ComfyUI machine

You drop the Icarus node into your ComfyUI custom nodes folder and restart. It appears in your node menu ready to use.

3
🖧 You set up the server on another machine

You copy the server folder to a second computer (or second GPU in the same machine), point it at your FLUX model file, and launch it with one click.

4
🔗 You wire them together

You place the Icarus node between your model loader and sampler, enter the server's address, and everything connects automatically.

5
How are the machines connected?
🏠
Same machine, two GPUs

Both processes run on the same computer, sharing data over PCIe at blazing speed without any network overhead.

📡
Over your home network or VPN

Your main machine sends compressed data to the second machine across gigabit ethernet or even residential broadband.

6
✨ You generate an image

You queue a generation and watch it complete in seconds — the front half runs locally, the back half runs remotely, and the video encoder compresses everything in between.

🎉 Your image appears

You got a beautiful FLUX.2 image using half the VRAM, or generated it faster than your single GPU could manage alone — and you saved money by not buying a new card.

Sign up to see the full architecture

5 more

Sign Up Free

Star Growth

See how this repo grew from 52 to 48 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is comfyui-mesh?

This is a Python project that lets you run FLUX.2 image generation across two GPUs instead of one. The client-side ComfyUI node (Icarus) handles the front half of the model plus VAE, while a separate server process (Daedalus) runs the back half. Activations between them get compressed by NVIDIA's NVENC video encoder hardware, which sits idle during ML inference anyway. This means you can pair an RTX 5090 with a friend's GPU over VPN, or use two cards in the same machine without NVLink. The compression is essentially free since it uses dedicated silicon that wasn't doing anything else.

Why is it gaining traction?

The headline numbers are compelling: FLUX.2 Klein 9B at 1024x1024 generates in about 4.4 seconds split across two GPUs over plain gigabit ethernet, with only 0.5 seconds of wire overhead. The trick is treating ML activation tensors as video frames and letting NVENC compress them 3-10x before crossing the wire. LoRAs work transparently across the split, the block count is reconfigurable without restarting ComfyUI, and the node shows live connection status inline. Setup is straightforward: install the ComfyUI node, deploy the server on another machine, wire it into your workflow, and queue a generation.

Who should use this?

ComfyUI users running FLUX.2 Dev, FLUX.2 Klein 9B, or LTX 2.3 who want to leverage multiple GPUs without NVLink. Developers with a beefy desktop and a spare GPU elsewhere on the LAN. Anyone who wants to avoid buying a single massive GPU by combining two mid-range cards instead. Budget-conscious users who want more generation throughput without the NVLink tax.

Verdict

With only 48 stars, this is early-stage software and shows in the single-tenant server and untested same-machine dual-GPU path. That said, the documentation is thorough, the architecture is sound, and the NVENC trick is genuinely novel. The 0.8999999761581421% credibility score reflects real maturity concerns, but the project is actively maintained by a solo developer with clear priorities. If you have two modern NVIDIA GPUs and want efficient cross-machine FLUX or LTX inference, this is worth a serious look—just budget time for the initial setup and watch for updates as the community grows.

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.