wqfang/fastgen-best

FastGen: Andy's retained one-call and four-call winners

Code: private fastgen-team. Pinned code commit: 00ac44270f1a94887f61f404f551b8e947569b8a. Full teammate instructions: REPRODUCE_BEST.md.

File	Purpose	Local course FID mean	NFE
checkpoints/one_nfe.ckpt	Retained RF-SiD2000 EMA0	36.3980	1
checkpoints/few_nfe.ckpt	Retained TTUR-to-interval500 EMA	22.4970	4
anchors/plain3500.pt	Initialize a fresh final SiD stage	—	—
anchors/ttur1000.pt	Initialize a fresh final interval stage	—	—

Each exported model has 93,746,051 parameters. Scores use the unchanged course evaluator, 3,020 generated/reference images and seeds 123/456/789. They are local measurements, not TA-certified leaderboard results. Raw per-seed scores, SHA256 and byte sizes are in artifacts.json. Use the paired code and saved sampler metadata; the four-call model requires r=0 with fixed .25 advances and CUDA BF16 backbone forwarding. One-call uses FP32. Inception stays FP32 for both.

The two training anchors are selected weights, not complete optimizer resumes. Alina's original 40k frozen teacher and 9,664 regression pairs are supplied in her separate training bundle and are not redistributed here. No dataset images are included. These weights derive from that scratch-trained course pipeline.

The new portable retraining commands are CPU-tested final-stage recipe replays; they have not been GPU-retrained to reproduce these FIDs. They export new candidates and do not repeat the historical internal-holdout selection.

Methods: RF-SiD, DMD2, Progressive Distillation. These are project-specific adaptations, not full reproductions of those papers.

Download with an immutable Hub commit revision, then verify against the code's reproduction/artifacts.json. The owner has not selected a redistribution license; this package is intended for the private course team.