Trajectory Forcing: Structure-First Generation with Controllable Semantic Trajectories

Trajectory Forcing (TF) addresses the "black box" nature of modern generative models, where the path from noise to image is hidden and inaccessible. While diffusion and flow-based models produce high-quality results, they treat intermediate states as transient computational by-products. TF redefines the generative process as a sequence of semantically structured stages—progressing from global layout to object, part, and detail-level representations—that are explicitly decodable and editable.

The authors propose a trajectory-centric framework that organizes synthesis into a hierarchy. To make this operational, they use pretrained DINOv2 visual representations, which naturally capture semantic structure. By applying unsupervised clustering to these features, the model constructs a teacher hierarchy that defines how an image should be decomposed into parts and subparts.

Instead of iterative denoising, TF employs a hierarchy-conditioned one-step flow-matching model at each level. This allows the model to generate a coherent intermediate state at each stage of the hierarchy. Because each stage is decodable via a shared Representation Autoencoder (RAE), users can inspect, evaluate, and perform localized edits on the image structure before the model proceeds to the next, finer-grained level of detail.

By shifting the focus from the final image to the generative path, TF provides a new interface for human-AI collaboration. It transforms the generation process from a single, opaque mapping into a controllable, interactive workflow. This approach not only improves transparency but also allows for precise, localized interventions that are impossible in standard end-to-end generative models, all while maintaining competitive image quality and efficient inference speeds.