Local multimodal models bottleneck on simple vision tasks

A developer attempting to filter 5,000 images of red cars using local Vision-Language Models on an 8GB GPU found inference taking up to three minutes per image. The community discussion highlights a growing trend of developers over-engineering simple computer vision pipelines with massive generative AI models.