مولدات GigaGAN المتدرجة

GigaGAN, introduced by Adobe and researchers in 2023, addressed instability encountered when naively scaling earlier GAN architectures to large text-to-image datasets. The paper widens the generator and discriminator, uses sample-adaptive convolution filters, and incorporates cross-attention to text embeddings. The generator emits an image pyramid across multiple scales. The paper reports a 512px output in about 0.13 seconds and a 4096px output in about 3.7 seconds for the tested framework; the high-resolution pipeline uses the base generator with a separate GAN super-resolution model. Results apply to those reported configurations and do not establish that GigaGAN is faster than every diffusion model. The separate upsampler can also be applied to low-resolution inputs from other text-to-image systems. In one high-capacity framework configuration reported in the paper, the base generator has 652.5 million parameters and the separate upsampler has 359.1 million, for 1.0 billion total parameters across the two networks.