Paper splits visual generation into autoregressive knowledge and diffusion rendering
A 3B text-to-image model trained in 6,600 H100 hours scored 86 on GenEval and 87 on DPG-Bench.
Stories per day · 30 days
4 stories, newest first · one card per story
Stories tagged #visual-generation
Rethinking Visual Provenance: Detection and Watermarking Across Direct Visual Generation and LLM-Driven Code Rendering
A conceptual agenda compares detection and watermarking for generated and code-rendered media.
OmniHarness: Harnessing Generalizable Visual Generation via Symbolic Policy Learning
OmniHarness learns symbolic policies for visual generation agents, reaching a 95.0% resolve rate on ComfyBench Creative tasks, 27.5 points above the strongest baseline.
Agentic Visual Generation: From Generative Models to Agentic Control
Researchers propose an L0-L4 control taxonomy for agentic visual generation, classifying controllers from fixed conditioning to experience-adaptive decision-making.