Physics-Grounded Hallucination Repair and Physics-Driven Consistency Distillation

Rethinking Video Consistency with 'Full Consistency Distillation and Faithful Preference Optimization'

Full Consistency Distillation creates extremely consistent and dynamic motions, while Faithful Preference Optimization ensures the result actually aligns with the user's prompt without drifting.

Fixing Physics Hallucinations with 'Pressure De-Hallucinator'

A major step forward for video forensics, the Pressure De-Hallucinator predicts real-world physics consequences and edits out AI-induced hallucinations (like clipping) using RL.

Unified Spatio-Temporal Editing for Video Diffusion

This method brings object-level editing to video generators by unifying spatial and temporal masks, allowing for precise local modifications without affecting the rest of the scene.

Updating RF-INV for High-Fidelity Text-Based Image Editing

Taming RF-INV expands instruction-based editing capabilities, focusing on inversion robustness. Crucial for anyone building advanced control layers for image gen models.

Controlling Text in Your Designs with TextCtrl

TextCtrl demonstrates highly precise spatial control over design layouts, simplifying the process of locking characters or objects to specific canvas coordinates.