MeanFlowAdvantage

Code: HaCTang/MeanFlowAdvantage.

Promoter/     # FANTOM5 DNA
  pretrain/   # Riemannian MeanFlow snapshot
  onpolicy/   # teacher-free on-policy RL
  RL/         # reward-graded distillation
AnyFlow/      # SD3.5 / AnyFlow fine-tuning snapshots

Promoter — DNA

Four FANTOM5 checkpoints: one MFA and one NFT for reward-graded distillation (RL/), and one MFA and one NFT for teacher-free on-policy RL.

Path	Method	Test Sei MSE	6-mer
Promoter/RL/mfa/seed123	Reward-graded MFA (step 90)	0.0467	0.953
Promoter/RL/nft/seed123	Reward-graded MeanFlowNFT (step 90)	0.0570	0.955
Promoter/onpolicy/mfa/seed123	On-policy MFA (step 300)	0.0437	0.952
Promoter/onpolicy/nft/seed123	On-policy MeanFlowNFT (step 250)	0.0445	0.927
Promoter/pretrain/ckpts/0120	Pretrained RMF	0.0714	0.928

Each fine-tune folder is an Orbax StandardCheckpointer payload. Architecture config is in Promoter/pretrain/.hydra/config.yaml.

AnyFlow — images

SD3.5-Medium LoRA EMAs, copied from Jietson/MeanFlowAdvantage-SD35M, plus one existing training snapshot. Details and checksums are in AnyFlow/README.md.

Path	Checkpoint
AnyFlow/pretrain/generator_ema.pt	AnyFlow pretraining adapter
AnyFlow/onpolicy/generator_ema.pt	AnyFlow on-policy distillation adapter
AnyFlow/mfa/generator_ema.pt	MeanFlowAdvantage EMA (reported 4-NFE model)
AnyFlow/Anyflow-MeanFlowAdvantage-v3/checkpoint-500/	Step-500 training state