Hey Product Hunt 👋 Built StudioMI300 solo for the AMD x lablab hackathon (May 2026). Motivation was simple: Sora and Runway produce beautiful clips but leave you with zero leverage - closed weights, vendor cloud, your output under their ToS. I wanted to know: can the open-source stack (FLUX.2, Wan2.2, ACE-Step, Kokoro) be stitched into an autonomous end-to-end pipeline that produces a complete cinematic artifact from one prompt? Turns out yes - if you have a single AMD MI300X with 192GB HBM3, which lets four very different architectures share one card sequentially. What's different vs other AI video tools: 🎬 End-to-end autonomous - single sentence, no manual scene/shot writing 🤖 Director Agent emits inspectable JSON plan (edit shots, re-render selectively) 🔍 Vision Critic auto-retries bad clips with 10-label failure taxonomy 🎵 Music + 9-language narration in the same pipeline 🎨 Character identity via FLUX.2 reference editing (no LoRA training step) 📦 100% Apache 2.0 / MIT - outputs commercially usable Built in 7 days. Every line of optimization (FBCache 2× + torch.compile 1.2×) and every incident (AITER FP8 segfault, MagCache calibration miss) is documented publicly. Try the live demo on HF Space (link above) or fork the code on GitHub: https://github.com/bladedevoff/s... Would love feedback, especially from anyone who's tried to compose open-weights video stacks before. Happy to dig into architecture choices, optimization details, or the vision critic's failure taxonomy in comments.

StudioMI300

Open-source text-to-cinematic-reel on a single AMD MI300X

Open-source text-to-cinematic-reel on a single AMD MI300X