Came back to Iris after a while because responses were way too slow on free CPU hosting. Today I moved the model to Hugging Face ZeroGPU, so it actually runs on a GPU now. I rewrote the backend route to call the Space through the Gradio API, with retries and fallback messages in case something fail