Kuaishou has officially open-sourced Kolors, a state-of-the-art text-to-image diffusion model that leverages ChatGLM3 for advanced bilingual text understanding and employs a two-stage training strategy to achieve photographic image quality rivaling leading proprietary systems.
The article reviews recent automated quality assessment advances for AIGC image generation, including an aesthetic scoring framework with the APDD dataset and AANSPS network, a human‑preference benchmark (HPD v2 and HPS v2) that outperforms IS/FID, and the Pick‑Score model trained on user‑driven Pick‑a‑Pic data, all enabling faster, unbiased evaluation, cost savings, and more effective model iteration, with ongoing work in home‑improvement AI.
This article explores multimodal representation learning by introducing a Bilinear Residual Layer that automatically fuses image and text features, demonstrates its superiority over traditional concatenation and FiLM methods on text‑guided image editing and fashion synthesis tasks, and reports state‑of‑the‑art results on several benchmark datasets.
