An optimized, full-featured AI Web Studio and execution engine built on ComfyUI, custom-tailored for high-performance AI generation on AMD ROCm 7.14, RDNA 3.5 APUs (Radeon 890M / Ryzen AI 9 HX 370), FLUX.1 GGUF models, and MiniMax Music 0.1 / ACE Step audio generation.
- 3-Column Studio Layout: Clean, responsive tabbed creation suite for Image Generation, Audio Production, and Prompt Enhancement.
- Expandable Multi-Line Textareas: Dynamic input controls for prompt, style, vocal specifications, arrangement timelines, and song lyrics.
- Real-Time Audio Progress & FLAC Export: Live WebSocket progress indicator (
0%to100%) for MiniMax audio sampling with built-in.flactrack downloader.
- Real-Time Lyrics Search: Type any song title or prompt into the concept input and the backend automatically queries the open LRCLIB database to retrieve verbatim lyrics.
- Auto MiniMax Tagging: Formats retrieved lyrics into MiniMax music section tags (
[Intro],[Verse 1],[Chorus],[Verse 2],[Outro]).
- FLUX.1-dev GGUF Integration: Full support for quantized GGUF image generation models (
flux1-dev-Q4_0.gguf). - Automatic LoRA Trigger Word Detection: Scans safetensors header metadata and displays clickable trigger badges (e.g.
class1cpa1ntfor Classic Painting FLUX).
Tailored for AMD Ryzen AI 9 HX 370 APU / Radeon 890M (gfx1150) with 51.5 GB LPDDR5X Unified Memory:
- PyTorch 2.13 SDPA AOTriton Fused Cross-Attention: Direct kernel dispatch for matrix math (
TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1&MIOPEN_FIND_ENFORCE=3). - PyTorch Segment Expansion: Eliminates memory fragmentation with
PYTORCH_HIP_ALLOC_CONF&PYTORCH_CUDA_ALLOC_CONF(expandable_segments:True,garbage_collection_threshold:0.8). - HSA Hardware SDMA Direct Transfers: Enables
libhsa-runtime64.so.1system direct memory access (HSA_ENABLE_SDMA=1).
- Least Recently Used (LRU) Model Eviction: Active models (FLUX.1 GGUF, Klein, Juggernaut XL, MiniMax Audio) remain 100% resident in VRAM during consecutive runs, guaranteeing instant zero-reload generations.
- Context-Aware Dynamic Swapping: Automatically evicts inactive model weights when switching creation contexts (e.g., Image Studio ↔ Music Studio or switching checkpoint families), preventing PyTorch allocation pool fragmentation and OOM errors.
- Segment Expansion: PyTorch ROCm allocation pools dynamically expand to accommodate high-resolution VAE decodes seamlessly.
- Centralized Location:
HF_HOME="/home/jason/models/HF-Hub"andHF_HUB_CACHE="/home/jason/models/HF-Hub/models". - Universal Model Scanning: All downloaded checkpoints, LoRAs, text encoders, and GGUFs are automatically stored and scanned from the central
/home/jason/models/directory structure. - Segment Expansion: PyTorch ROCm allocation pools dynamically expand to accommodate high-resolution VAE decodes seamlessly.
./run_comfyui.shRuns the accelerated backend on http://127.0.0.1:8188.
./run_studio.shServes the custom Web Studio on http://localhost:8090.
├── run_comfyui.sh # Accelerated ComfyUI launcher script with ROCm 7.14 flags
├── run_studio.sh # Web Studio server launcher script
├── web_studio/
│ ├── studio_server.py # FastAPI server, LRCLIB lyrics lookup, model discovery
│ └── static/
│ ├── index.html # Responsive 3-column Web Studio interface
│ ├── app.js # Client event handlers, WebSocket progress & audio player wiring
│ └── style.css # Studio dark theme stylesheet
├── comfy/ # Core ComfyUI tensor engine & MiniMax/ACE audio patches
├── comfy_extras/ # ACE Step 1.5 music nodes
└── custom_nodes/ # GGUF model loader & custom extensions
This project inherits the core ComfyUI GPL-3.0 license.