Will this approach work with image generation models?