On the Resilience of Text-to-Video Diffusion Models to Hardware Faults
A single hardware fault was enough to visibly change some generated videos, including their semantics.
The paper studies random computational and memory faults across three text-to-video diffusion models. It reports up to a 3.7% performance drop from one fault, with semantic correctness hit harder than perceptual quality. Memory faults were more damaging than computational faults, and bfloat16 was more vulnerable than other formats tested. The authors say 7-28% of faults produced visible artifacts, including added objects. ArXiv · AI/CL/LG's note
The paper studies random computational and memory faults across three text-to-video diffusion models. It reports up to a 3.7% performance drop from one fault, with semantic correctness hit harder than perceptual quality. Memory faults were more damaging than computational faults, and bfloat16 was more vulnerable than other formats tested. The authors say 7-28% of faults produced visible artifacts, including added objects. ArXiv · AI/CL/LG's note
score 4