Last week in Image & Video Generation

Last week saw a surge of innovation in the open-source multimodal AI space, with significant releases spanning image generation, audio synthesis, and robotics simulation. The standout announcement was BiTDance, a 14-billion parameter autoregressive model for image generation, representing a substantial open-source alternative in a field dominated by diffusion models. Alongside this, NVIDIA released DreamDojo, an open-source world model that allows robots to practice tasks in a simulated visual environment using only motor controls as input, eliminating the need for physical hardware during training. This release underscores the growing importance of simulation for scalable robotics development.

On the audio front, the AudioX research project introduced a unified model capable of generating audio from any input modality—text, video, image, or existing audio—demonstrating progress toward truly general-purpose multimodal systems. Other notable tools included an updated LoRA forensic copycat detector for identifying model copies and a new inpainting node for the LTX-2 video model that simplifies fixing specific regions in generated clips. These releases, from the 14B parameter BiTDance to the versatile AudioX, highlight the community's focus on building larger, more capable, and more accessible foundation models and tooling, pushing the boundaries of what's possible with open-source AI.