A detector reported 0.91 mAP and a GAN reported FID 0.24 — both numbers were lies told by the pipeline, not the model. The cross-project habit that caught them, and that killed a tempting hypothesis cheaply, in one place.

A reusable protocol for reporting FID that someone can actually reproduce: pin the feature extractor, the input pipeline, the sampling, and the controls — with a copy-paste reporting table and checklist, distilled from a real FID bug.

After ruling out discriminator-dominance, the next suspect for our GAN's FID-160 plateau was limited-data discriminator overfitting. DiffAugment — applied to both real and fake at the discriminator only — cut the subset FID from 163 to 118 and erased the late collapse. The ceiling was partly data-bound.

Our text-to-image GAN peaked at epoch 20 then degraded, with the textbook 'discriminator is winning' loss signature. Four targeted runs (EMA, TTUR, label smoothing, n_critic) ruled that hypothesis out: weakening D never robustly improved FID, which floored at ~160.

A structural map of MS-CLIP-GAN: CLIP text embeddings, conditioning augmentation, a 64→128→256 generator, stage-wise discriminators, and why that architecture made the discriminator-balance hypothesis worth testing.

A text-to-image GAN reported FID 0.2423 and it looked near-perfect. It wasn't: pytorch-ignite's default FID scores in a 1000-d logit space, not the standard 2048-d pool3. The same model's real FID was ~165, and the two scales don't convert.

Solution for SolveSQL VIP of Cities using CTEs and window ranking.

Position-aware ESM2 features, frozen embedding caching, and honest evaluation caveats for a supervised GOF/LOF protein variant classifier.

My safety-gear detector reported 0.91 mAP and a −93% false-positive cut. Then I audited my own evaluation and found three validity problems — two data leaks and a seed confound — that make both numbers softer than they sound. Here's the one I'd actually trust.

RTMDet trains great on one consumer GPU — once you survive the prebuilt-wheel lock, a pkg_resources crash, and a pip ResolutionImpossible (and debunk the SyncBN scare). The exact pins and fixes that worked.