Text-to-Image: GW Direct vs GW Attention vs MaskGen-KL vs CAD-I

Two 10-caption qualitative probes. MaskGen-KL uses turkeyju/generator_maskgen_kl_xl with the TA-TiTok BL32 VAE tokenizer. CAD-I is kept as an optional ImageNet-trained SD-latent-grid comparison. Our columns use the frozen ImageNet-only GW checkpoint and RepTok decoder.

ImageNet Captions

COCO Captions

Manifest: manifest.json. Summary: summary.json.