Search papers, labs, and topics across Lattice.
Affiliation:
8
0
9
7
Iterative refinement in data space, not flow space, unlocks fast, high-fidelity generative sampling without manual timestep tuning.
Diffusion models no longer need to simultaneously plan and paint; inserting discrete "visual thought" tokens between VLMs and DiTs decouples high-level semantic reasoning from pixel synthesis for tighter alignment and direct intermediate control.
Converged diffusion loss in visual generation improves linearly with structured language, leading to a new training paradigm that outperforms both open-weight and closed-weight models.
Reinforcement learning boosts multimodal performance, raising task scores and creating unexpected synergies between image generation and editing.
Ditch the VAE bottleneck: Representation Forcing lets you train unified multimodal models to generate high-quality images directly from pixels, rivaling VAE-based approaches without the architectural constraint.
Training a single model across text, images, video, 3D geometry, and hidden representations unlocks "Context Unrolling," where the model reasons across modalities to improve reasoning fidelity.
Adversarial training can drastically improve the sample quality of existing flow-matching models, achieving FID improvements of over 4.5 points on ImageNet 256px.
Open-source VQ-VA models just got a massive boost: a new dataset and benchmark close the gap with proprietary systems on visual question-visual answering.
