Daniel Cohen-Or

LLM interrogations reveal diffusion model's internal tokens capture scene semantics surprisingly early, even without explicit prompts, and correlate with generation quality.

Joint audio-video diffusion models hallucinate not from diffuse attention noise, but because bidirectional audio-video cross-attention systematically overrides text conditioning in favor of learned canonical priors.

Achieving a 23-52x speedup in 3D urban modeling, SiZeUp leverages ordinal depth consistency to transform aerial imagery into accurate proxies without the pitfalls of traditional depth estimation.

Fine-grained identity tuning enables precise facial edits in text-to-image models without additional training, preserving identity consistency across diverse outputs.