Flux.2-Klein pipeline for real-time webcam stream processing in 30 FPS
Open-source FluxRT runs Flux.2-Klein at 50 FPS on a single RTX5090...
TensorForger's new FluxRT pipeline brings real-time video processing to consumer hardware. Built on the Flux.2-Klein-4B model, it streams webcam footage at 30 FPS with latency as low as 0.2 seconds—all on a single RTX5090 GPU. The key innovation is a custom spatial-aware KV-cache that only recomputes image tokens where motion or change occurs, drastically cutting compute per frame. To smooth output, it integrates RIFE frame interpolation, supporting multipliers of 2x, 4x, and 8x (4x is recommended). Benchmark results show 50 FPS in mostly static scenes and around 20 FPS when the full input changes rapidly.
FluxRT is completely free and open-source, with a GitHub repo housing a Gradio demo, minimal cv2 examples, and even a paint-style app that updates the canvas in real time. This enables accessible, low-latency video-to-image streaming for creative tools, interactive AI art, or live camera filters without needing a server farm. The repository includes full benchmark data and setup instructions, inviting the community to experiment with their own hardware configurations.
- Custom spatial-aware KV-cache reduces token recomputation to only moving regions, cutting latency to ~0.2s
- RIFE frame interpolation boosts FPS by up to 4x, achieving 50 FPS in static scenes on a single RTX5090
- Open-source pipeline includes Gradio demo, cv2 examples, and a real-time paint app for creative use
Why It Matters
Makes real-time AI video processing practical on a single consumer GPU, enabling interactive applications without cloud dependency.