BEGIN:VCALENDAR VERSION:2.0 PRODID:-//UC Irvine//CML Seminars//EN CALSCALE:GREGORIAN METHOD:PUBLISH X-WR-CALNAME:What Generative Visual Models Understand (and Don't) About the Physical World X-WR-TIMEZONE:America/Los_Angeles BEGIN:VTIMEZONE TZID:America/Los_Angeles BEGIN:DAYLIGHT TZOFFSETFROM:-0800 TZOFFSETTO:-0700 TZNAME:PDT DTSTART:20070311T020000 RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=2SU END:DAYLIGHT BEGIN:STANDARD TZOFFSETFROM:-0700 TZOFFSETTO:-0800 TZNAME:PST DTSTART:20071104T020000 RRULE:FREQ=YEARLY;BYMONTH=11;BYDAY=1SU END:STANDARD END:VTIMEZONE BEGIN:VEVENT UID:2025-03-17-anand-bhattad@cml.ics.uci.edu DTSTAMP:20250317T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20250317T110000 DTEND;TZID=America/Los_Angeles:20250317T120000 SUMMARY:[CML Seminar] Anand Bhattad: What Generative Visual Models Understa nd (and Don't) About the Physical World LOCATION:Donald Bren Hall 4011 DESCRIPTION:Anand Bhattad\, Research Assistant Professor\, Toyota Technolog ical Institute at Chicago\n\nTitle: What Generative Visual Models Understa nd (and Don't) About the Physical World\n\nAbstract: Generative visual mod els like Stable Diffusion and Sora generate photorealistic images and vide os that are nearly indistinguishable from real ones to a naive observer. H owever\, their grasp of the physical world remains an open question: Do th ey understand 3D geometry\, light\, and object interactions\, or are they mere pixel parrots of their training data? Through systematic probing\, I will demonstrate that these models surprisingly learn fundamental scene pr operties — intrinsic images such as surface normals\, depth\, albedo\, a nd shading — without explicit supervision\, which enables applications l ike image relighting. But I will also show that this knowledge is insuffic ient. Careful analysis reveals unexpected failures: inconsistent shadows\, multiple vanishing points\, and scenes that defy basic physics. All these findings suggest these models excel at local texture synthesis but strugg le with global reasoning: a crucial gap between imitation and true underst anding. I will then conclude by outlining a path toward generative world m odels that emulate global and counterfactual reasoning\, causality\, and p hysics.\n\nhttps://cml.ics.uci.edu/seminars/2025-03-17-anand-bhattad X-ALT-DESC;FMTTYPE=text/html:Anand Bhattad\, Research As sistant Professor\, Toyota Technological Institute at Chicago

Ti tle: What Generative Visual Models Understand (and Don't) About the Ph ysical World

Abstract: Generative visual models like Stable Diffusion and Sora generate photorealistic images and videos that are near ly indistinguishable from real ones to a naive observer. However\, their g rasp of the physical world remains an open question: Do they understand 3D geometry\, light\, and object interactions\, or are they mere pixel parro ts of their training data? Through systematic probing\, I will demonstrate that these models surprisingly learn fundamental scene properties — int rinsic images such as surface normals\, depth\, albedo\, and shading — w ithout explicit supervision\, which enables applications like image religh ting. But I will also show that this knowledge is insufficient. Careful an alysis reveals unexpected failures: inconsistent shadows\, multiple vanish ing points\, and scenes that defy basic physics. All these findings sugges t these models excel at local texture synthesis but struggle with global r easoning: a crucial gap between imitation and true understanding. I will t hen conclude by outlining a path toward generative world models that emula te global and counterfactual reasoning\, causality\, and physics.

< a href="https://cml.ics.uci.edu/seminars/2025-03-17-anand-bhattad">https:/ /cml.ics.uci.edu/seminars/2025-03-17-anand-bhattad URL:https://cml.ics.uci.edu/seminars/2025-03-17-anand-bhattad END:VEVENT END:VCALENDAR