ChitraMaya

ChitraMaya: The Agent-Native Studio on AMD MI300X ChitraMaya is a revolutionary, agent-native visual production console designed to tackle all three tracks of the AMD Developer Hackathon simultaneously. We set out to build more than just another prompt gallery—we built a completely autonomous, API-first control plane where an AI agent acts as the director, and an AMD MI300X GPU acts as the powerhouse rendering engine. The core process begins with Fine-Tuning on AMD. Users can upload a small dataset of reference images directly through our sleek Studio UI. Under the hood, the system dispatches a LoRA training job directly to the AMD MI300X utilizing ROCm. This process trains a highly accurate, custom character identity model in roughly 90 minutes—without relying on CUDA or external cloud services. Once the custom identity is registered, the true magic of our Agentic Workflow begins. Instead of manually tweaking complex nodes in ComfyUI, users interact with our AI agent using natural language (e.g., "Create a cinematic video of my custom character walking through a neon-lit cyberpunk city"). The agent autonomously plans a multi-shot project, selects the correct tools, and submits the optimized workflows to our FastAPI backend. For the Vision & Multimodal generation, ChitraMaya leverages the massive 192GB VRAM of the MI300X to run state-of-the-art models flawlessly. We use FLUX.2 for high-fidelity base images and the massive 14B parameter Wan 2.2 model for stunning text-to-video and image-to-video generation. By unifying custom fine-tuning, heavy multimodal video rendering, and autonomous agent orchestration into a single dashboard, ChitraMaya proves the unparalleled capability of the AMD hardware ecosystem for the next generation of AI film production.