Maestro AMD (github.com/yoyokits/Maestro-AMD) is a Pinokio wrapper that gets the Maestro creative studio running on AMD GPUs via ROCm — 100% local AI video, image, and music generation, no cloud, no subscription. It bundles Director Mode (prompt → finished short film/music video), Studio Mode (manual control over MiniMax H3, LTX‑2.5, Hunyuan Video, Flux 2, Krea 2, TTS/music/SFX models), and Edit Mode (retakes, character swap, repaint, outpaint). Runs on RDNA 2/3/4 cards (RX 6600–9000 series) and Ryzen AI APUs, on Windows or Linux.
⚠️ Before using Maestro AMD: change settings or risk a system hang
This is the one step you cannot skip. MiniMax H3 ships defaulting to NVFP4 text encoding — an NVIDIA-only format. On AMD hardware there is no NVFP4 support, so Maestro silently falls back to running the 32-billion-parameter text encoder on your CPU, which can exhaust system RAM and freeze the whole machine, not just the app.
Step-by-step fix (one-time)
- Open Maestro, select a MiniMax H3 model in Studio mode, and click the Advanced Settings icon (bottom-left, sliders icon) on the generation panel.
- Find H3 Text Encoder. It will show "NVFP4 AWQ (Recommended)" — ignore that label, it's NVIDIA-centric and wrong for AMD.
- Change it to GGUF Q4_K_M. If you're RAM-constrained, use GGUF Q2_K instead.
- In Settings → Performance, confirm your GPU is detected correctly (e.g. "AMD Radeon RX 7900 XTX") and leave Auto-tune on — it picks a safe performance profile for your VRAM/RAM.
NVIDIA-only settings to avoid entirely on AMD
- NVFP4 (text/model encoding) — CPU-offloads and can hang the system.
- SAGE Attention — CUDA-only attention kernel, no ROCm build exists.
- Triton — NVIDIA/CUDA compiler backend; not usable on AMD.
- CUDA anything — there is no CUDA on AMD; Maestro falls back to standard PyTorch attention on ROCm instead.
If a setting's description mentions any of the above, or says "requires an RTX GPU," treat it as NVIDIA-only and pick the GGUF/CPU/ROCm-compatible alternative instead.
Everyday use
Once the H3 Text Encoder is switched to GGUF, generation runs normally: pick Director or Studio mode, describe what you want, adjust resolution/duration/LoRAs as needed, and generate. Output quality matches NVIDIA; AMD is somewhat slower due to attention-kernel differences, and VRAM (24 GB+ is comfortable) matters more than raw GPU speed for what you can run.