TOPPAN and Keio test quantum+AI anime images, accuracy up to 78%

The proof-of-concept chose flow matching as its generation method. Flow matching is a way to generate images in fewer steps than diffusion models, which are now the mainstream approach for image generation AI. Diffusion models repeatedly remove noise in small increments, so they involve many computation steps and take time to generate an image.

In the hybrid model, Keio University implemented flow matching on a quantum circuit to generate core images such as character structure, while TOPPAN added background generation, colorization, and high-resolution upscaling using AI. Testing on a real 10qubit quantum computer produced a 16×16-pixel core image with specified hairstyle and facial expression, and after AI processing it was confirmed to output a 256×256-pixel color image practical enough for use.

In simulation testing, the smaller quantum circuit cut the number of operations to about one-sixth of the conventional level while raising discrimination accuracy from 16% to 78%. The 78.3% accuracy on 10,000 small training samples, above the 54.3% for conventional computer AI, suggested an advantage in avoiding local solutions. The stated remaining challenge is whether quantum hardware can show superiority over conventional computer AI in image generation accuracy and generation time.