Flux.2-Klein pipeline for real-time webcam stream processing in 30 FPS
I have built a pipeline based on the Flux.2-Klein-4B model that allows processing of a video stream with low latency (about 0.2 seconds) on a single RTX5090 GPU.
It is free and open-source, you can try it locally:
https://github.com/tensorforger/FluxRT
Under the hood, it uses a custom spatial-aware KV-cache, so it only recomputes a small number of image tokens per frame, specifically where something is moving or changing.
It also uses frame interpolation with the RIFE model, which can multiply FPS by a factor of 2, 4, 8, etc. I have found that 4 is the most appropriate for my setup.
Depending on scene dynamics, the output stream achieves up to 50 FPS in mostly static scenes and around 20 FPS when the entire input image is changing rapidly. Benchmark results are in the repo.
There is also a Gradio demo, several minimal cv2 examples, and a simple paint-style app with real-time canvas updates.
EDIT: Thanks a lot for support! Added int8 quantization mode, so it would now run smoothly on RTX 4090 too with 20 GB VRAM in peak.