Instagram story viewer> @hou2touch> Posts

Hou2Touch

@hou2touch

2.2K

followers

4524

following

• Generative artist.
• Ex-Head of Technology at @inter.nyc

POSTS STORIES REELS TAGGED

A skill that trains a LoRA with you — for Claude Code and Codex. Alex Korneev (@forkni) built it for everyone who comes to Sunday’s masterclass.

It isn’t a tutorial, it’s an assistant that sits in your terminal. You say what you’re training; it picks the phase and walks it with you: dataset, captions, training, checkpoint choice, validation. Six phases, each one a script you can run by double-clicking.

Why that helps. You finish a run and you have a folder full of checkpoints — which one do you ship? The usual answer is to render a few and eyeball them: GPU time, and still a guess. The skill screens the whole run in under a minute, on CPU, without rendering a single image.

And it knows where that screen lies. Numbers can’t see colour — one retrain came back clean on every weight-only check and still rendered a magenta cast. So it never stops at the numbers: real renders across several seeds decide.

It also refuses to call a LoRA done too early. A still-image tool runs about 25 denoising steps; a real-time component runs one to four. A LoRA that was only ever checked at 25 has not been checked — so the skill won’t sign off until you’ve judged it in the component, at the step count and resolution you actually run.

And it doesn’t answer from memory. Numbers come out of its reference files, and where one is still provisional, the file says so. The slides are screenshots of those files.

What you need: Windows, an NVIDIA GPU, Claude Code (or Codex), and a checkout of the OneTrainer fork it ships in. The masterclass walks the setup.

How you get it — just come to the masterclass. It opens after the session to everyone who registered, along with the replay and the written companion.

Free. Sunday 13 September · 12pm ET / 6pm CEST / 9am PT · 90 minutes. Registration — link in bio.

Visuals by Alex Korneev.
A collaboration between @creativecodeart and @hou2touch. by @hou2touch

65

2

a month ago

A stable image out of noise — Alex Korneev’s (@forkni) feedback loop for StreamDiffusionTD, in five steps.

1. The input isn’t a camera. It’s a looping noise generator — the Noise_looper node — feeding StreamDiffusionTD.

2. The output goes back in — as a mask multiplied onto the next noise frame. The brighter the mask, the more of the last frame survives into the next. That is what holds the image together instead of letting it reinvent itself every frame.

3. Depth ControlNet, faked. You don’t need a real depth map: the model is trained on black-and-white images, so all it cares about is a white foreground on a black background. Alex takes a loop from the web, cuts out the subject, thresholds it and softens it slightly — real depth maps are soft too.

4. A second feedback kills the «boiling» — the frame-to-frame shimmer raw diffusion output has. Push the smoothing too far and you get motion blur, so it’s a dial, not a switch.

5. NVIDIA super-resolution sharpens the result, and then it goes back into the loop. Every cycle starts from a better frame than the last.

That’s half the setup. On Sunday he shows the other half — how to train your own style for it — and you can ask him anything in the Q&A.

Free masterclass, September 13 · 12pm ET / 6pm CET / 9am PT · 90 minutes. Link in bio.

Visuals and setup by Alex Korneev.
A collaboration between @creativecodeart and @hou2touch. by @hou2touch

78

2

a month ago

Train a diffusion model on your own images until it paints in your style — then load that style into TouchDesigner and repaint a live webcam with it. Alex Korneev (@forkni), co-author of the StreamDiffusionTD component, does the whole thing live in 90 minutes.

12pm New York · 6pm Berlin · 9am LA
Free. Registration closes an hour before we start. Link in bio.

The replay goes to everyone who registers, so sign up even if the time doesn’t work for you. by @hou2touch

20

2

a month ago

Train an AI on your own images until it paints in your style — then run that style live on a webcam inside TouchDesigner.

Free masterclass with Alex Korneev (@forkni), co-author of the StreamDiffusionTD component. He trained six versions of one style: one by every rule, five each breaking exactly one. You’ll see what each mistake does to the render — including the model that passed every test and then started printing pink.

Sunday, September 13 · 12pm ET / 6pm CET · 90 minutes · free
Replay and the full training kit go to everyone who registers. Link in bio. by @hou2touch

53

18

a month ago

Running real-time AI in TouchDesigner? That CPU round-trip is costing you.

Every time a GPU texture leaves for CPU memory and comes back, you’re paying in latency, bus bandwidth, and dropped frames. CUDA-Link eliminates it.

⚡️ Zero-copy GPU transfer: TD↔️Python, TD↔️TD, Python↔️Python
⚡️ ~3.4× faster end-to-end at 1080p vs CPU SharedMemory
⚡️ Zero-copy PyTorch tensor reads in <5 µs — already on GPU, ready for your model
⚡️ Full pipeline: capture TD output → run AI → inject result back into TD
⚡️ Drops straight into diffusers, ControlNet, any PyTorch pipeline
⚡️ Tested across all TouchDesigner pixel formats — no surprises at runtime
⚡️ One-command install — ships as a wheel with a 5-mode installer that auto-discovers TD and Python

v1.9.0 just shipped!

Works with TouchDesigner 2022+, Python 3.9+, CUDA 12.x on Windows.

🔗 Link in bio

#TouchDesigner #CUDA #RealTimeAI #OpenSource #MachineLearning taken in Brooklyn, New York by @hou2touch

28

6

4 months ago

Real-time FLUX.2-Klein model running locally on a single RTX5090 GPU. Conditioned with prompt and RGB camera input. Custom KV cache bank implementation for temporal consistency + a ton of optimizations 😵‍💫😈
•
#flux2 #realtimerendering #nvidia #touchdesigner #blackforestlabs taken in Brooklyn, New York by @hou2touch

70

12

5 months ago

Real-time diffusion with FLUX.2-Klein4B model + streamV2V in Touchdesigner 😳
Very impressed with the results!
Running smoothly on Nvidia RTX 5090 GPU locally.
Swipe to view the input ⏩️
•
#flux2 #torch #nvidia #blackforestgermany #touchdesigner taken in Brooklyn, New York by @hou2touch

41

8

6 months ago

For the last couple months I’m diving into FLUX.2 inference acceleration and here’s latency results so far:
LEFT SIDE: real-time input
RIGHT SIDE: real-time diffused output
•
Technology stack:
• NVIDIA 5090 GPU
• FLUX.2-Klein 4b model
• TensorRT acceleration
• FP8 selective quantization
• Pre-inference prompt calibration
• CUDA-Link for minimizing latency overhead
• 2 steps, 320x512px resolution
• 13 FPS
•
#blackforestgermany #flux2 #nvidia #touchdesigner taken in New York, New York by @hou2touch

24

10

6 months ago

Times like this remind you how much work depends on the people around you.
What begins as a structured set can slowly open into a long improvisation. And every now and then the stage and the visuals briefly meet in the same moment — something you could never fully plan.
Nomad Path Live Performance by @firouzfarmanfarmaian and @interactive.items
Made with @touchdesigner and #streamdiffusion by @dotsimulate
#generativeart #audiovisual #realtime #newmedia taken in MAD | Museum of Arts and Design by @hou2touch

117

8

7 months ago

Why does sharing GPU memory between two processes still require copying data through the CPU?
At 60fps, that round-trip costs milliseconds you can’t afford.
I built CUDA-Link to fix that — an open-source Python library that shares GPU textures and tensors directly between processes. Zero-copy, no CPU detour. Started because I needed live TouchDesigner frames in AI models without the latency hit. The GPU → CPU → GPU path was eating my frame budget, so I wrapped CUDA’s IPC API in pure Python and cut it out entirely.

What it does:
→ Zero-copy GPU sharing between any two processes
→ Ring buffer — producer and consumer never block each other
→ PyTorch, CuPy, NumPy output modes
→ Pure Python — pip install and go

60fps in production. Easy.

Open-source: github.com/forkni/cuda-link (link in bio)
•
#touchdesigner #python #cuda #realtimeai taken in New York, New York by @hou2touch

61

8

7 months ago

Live depth feed from Azure Kinect camera → ControlNet + StreamV2V (temporal consistency) → diffused output at 30 FPS with SDXL-Turbo at 2 Timesteps on a single RTX 5090.
Most StreamDiffusion implementations use TensorRT (incompatible with StreamV2V) — torch.compile gives both speed and consistency. Currently working on further pipeline acceleration 🚀🚀🚀
#streamdiffusion #torch #realtimerendering #generativeart #ai taken in New York, New York by @hou2touch

432

54

8 months ago

Real-time StreamDiffusion from Kinect Azure input (RGB + Depth)
ControlNet Depth + StreamV2V
•
#kinect #streamdiffusion #touchdesigner #nvi̇di̇a #realtimerendering taken in Brooklyn, New York by @hou2touch

56

12

8 months ago