I let an AI build a Yolo26 node set for vvvv. It escalated a bit

[tgd]

Yolo26 for vvvv

This is a preview / test implementing the latest Yolo release by ultralytics (models exported to ONNX).

Link to this sectionOverview - Ultralytics YOLO26 | Ultralytics Docs

Preview

GitHub - tgd-git/VL.Yolo26: Yolo26 node set for vvvv gamma

There is a local nuget inside the dist folder, install from vvvv´s nuget command line:
nuget install VL.Yolo26 -Source "C:\dist;``https://api.nuget.org/v3/index.json``" -Version 0.4.2 -PreRelease
(modify C:\dist to your actual folder you put the nuget into)

There is no nuget.org package yet but i will continue to burn as many tokens as it takes to publish it soon with all features:) (3 of 5 implemented by now)

Note: Just noticed the read back pipeline nodes wont work in vvvv 7.0. Did not investigate this further. Suggest you just run it in 7.3
Note2: Smoothest pipeline is Texture (GPU only), Most demanding is the Segment Task (Mask generation)

What has been built:

I have explored 3 different input and pipeline approaches and kept them all around. The outcome for now is 3 different node families across 3 yolo tasks (Detect, Pose, Segmentation). The difference at first was about image input (IVideoSource VS Stride Texture readback), then came a third variant that is a GPU only implementation for almost the whole processing pipeline. Skia was out of scope for now and the category naming might be worth reconsidering.
As a bonus there is extensive documentation written along by the agent including all node and pin descriptions in the editor.
Find help patches in help browser.

Video vs Texture vs Readback

	Video	Texture	Texture Readback
Input	IVideoSource (CPU frames)	Stride Texture (GPU)	Stride Texture (GPU)
Pixels to CPU	Source already CPU	Never (zero-copy)	GPU→CPU readback, delayed
Preprocess	CPU ImagePreprocessor	GPU compute	CPU ImagePreprocessor
ORT EP	CPU or DirectML	DirectML only, DML1 + IoBinding	CPU or DirectML
Threading	Background worker	Mainloop (sync GPU wait)	Mainloop readback + worker
Mainloop blocking	No	Yes	No
Patch deps	VL.Video typical	VL.Stride required	VL.Stride required

All nodes

Node	Category	Input	Inference
Detect (Video)	ML.YOLO26	IVideoSource	CPU or DirectML; background thread
Detect (Texture)	ML.YOLO26.Stride	Stride Texture	Zero-copy GPU (D3D12 + DirectML IoBinding)
Detect (Texture Readback)	ML.YOLO26.Stride	Stride Texture	Non-blocking GPU→CPU readback, then CPU/DirectML worker
Pose (Video)	ML.YOLO26	IVideoSource	CPU or DirectML; background thread
Pose (Texture)	ML.YOLO26.Stride	Stride Texture	Zero-copy GPU
Pose (Texture Readback)	ML.YOLO26.Stride	Stride Texture	Readback + CPU/DirectML worker
Segment (Video)	ML.YOLO26	IVideoSource	CPU or DirectML; four toggled mask outputs
Segment (Texture)	ML.YOLO26.Stride	Stride Texture	Zero-copy GPU; proto on GPU, mask assembly in compute shaders
Segment (Texture Readback)	ML.YOLO26.Stride	Stride Texture	Readback + same CPU mask assembly as Segment (Video)

Credits to @tonfilm for the vvvv skills. Agent said they helped a lot:)

AI Agent Skills for vvvv gamma

Hey evvvveryone, I put together a first shot on skills that teach AI coding assistants how to write proper C# code for vvvv gamma. If you ask plain Claude, Cursor, Copilot, Codex, or similar tools to write code for vvvv, they often get it wrong on the first try. These skills should help with that: What is this? “Agent Skills” is an open standard that works across all major AI coding tools. They are like cheat sheets that your AI assistant loads automatically when it notices you’re working on…

Models:

The repository ships with nano models for now, for convenient testing and help patches.
From: @zwh20081 on Hugging Face (yolo26-onnx)
See docs/models.md for more info.

Aaaand fingers crossed the nodes work for you too.

[bjoern]

Did you know that there is a Yolo VL package already? It’s based on YoloDotNet which afaiu also supports Yolo26. Maybe extend the existing package instead of creating a new one from scratch?

[tgd]

Yes i know there is one. This here is also a proof of concept and learning project for me to explore AI coding. I was surprised how well it went after the first prompt and then could not let go :)
And actually it was joreg insisting to publish it :)

[joreg]

but i also said it would be smarter to extend the existing one…

[tgd]

Honestly, i dont remember that