Fine-Tuning NVIDIA Cosmos Predict 2.5 with LoRA/DoRA for Robot Video Generation

NVIDIA's Cosmos Predict 2.5 is a large-scale world model capable of generating physically plausible videos, which can be fine-tuned for specific domains like robot manipulation. To avoid the high cost of full fine-tuning, parameter-efficient methods like LoRA and DoRA are used to inject small, trainable adapters into the frozen base model. This guide provides a walkthrough for fine-tuning the model using the `diffusers` and `accelerate` libraries for both single and multi-GPU setups. The process enables the generation of synthetic robot trajectories, offering a scalable alternative to expensive real-world data collection for robot learning tasks.