CLM-020 Claim

Proposition

LDM 在无条件图像合成上达到新的 SOTA（CelebA-HQ 256²：500 DDIM 步 FID 5.11，优于此前 likelihood-based 与 GAN），并在 FFHQ / LSUN-Churches / LSUN-Bedrooms 上与最强 GAN 和扩散模型竞争（Bedrooms 上接近 ADM，参数量减半、训练资源少 4 倍）；且在 Precision 与 Recall 上系统性优于 GAN——印证其 likelihood（mode-covering）目标相对对抗训练的分布覆盖优势。

Evidence refs

• EVI-020

Notes

主表（ms_tables.tex, FID/Precision/Recall）：CelebA-HQ — VQGAN+T 10.2、PGGAN 8.0、LSGM 7.22、UDM 7.16；LDM-4 5.11（prec 0.72/recall 0.49）。FFHQ — LDM-4 4.98（prec 0.73/recall 0.50）。LSUN-Churches — StyleGAN2 3.86；LDM-8 4.02（prec 0.64/recall 0.52）。LSUN-Bedrooms — StyleGAN2 2.35、ADM 1.90、ProjectedGAN 1.52；LDM-4 2.95（prec 0.66/recall 0.48）。§4.2 原文："we report a new state-of-the-art FID of 5.11, outperforming previous likelihood-based models as well as GANs"、"LDMs consistently improve upon GAN-based methods in Precision and Recall, thus confirming the advantages of their mode-covering likelihood-based training objective over adversarial approaches"。注意 SOTA 措辞限于 CelebA-HQ；Bedrooms 的 ADM 1.90 优于 LDM 2.95，"close to ADM despite half its parameters and 4× less resources"。