LLM Steers Compounding Video Hallucinations to the Beat: A Controlled Drift Experiment

ART-ficial-Ignorance · reddit · 2026-09-04

The author chained Gemma 12b, LTX 2.3, an audio-reactive LoRA and Wan2GP into a controlled hallucination experiment: the LLM holds the full video structure and music energy map but writes each 6.9s clip prompt only after 'seeing' the previous clip's final frame, steering accumulated artifacts back toward the planned arc. Using a stationary camera on one morphing object made errors compound faster than the previous infinite-hallway run, yielding a black ceramic form that develops its own visual grammar before destabilizing into a network-like surface. Next step: let the LLM control LoRA strength and other settings.

Original post →

More from Multimodal

Multimodal channel →