WTF?! Simulation-Free Reinforcement Learning with Wasserstein-Tilted Flow Maps

Checkpoints for WTF?! (NeurIPS 2026, Spotlight).

Code · Project page · Paper

File	Benchmark	Steps	Used for
wtf_t2i_step1460.pt	Text-to-image (TiM-T2I, LoRA r16)	1,460	text-to-image table, all WTF rows
wtf_t2i_euclidean_step800.pt	Text-to-image	800	exact vs. Euclidean ablation
wtf_t2i_exact_step800.pt	Text-to-image	800	exact vs. Euclidean ablation
wtf_imagenet_seed{0,1,2}_step12500.pt	ImageNet-256 (DMF XL/2)	12,500	ImageNet table

The text-to-image files hold rank-16 LoRA weights for TiM-T2I: lora is the evaluated adapter and ema its exponential moving average. The ImageNet files hold the EMA weights of the full DMF XL/2 network (ema), the weights that are evaluated. Every file loads with torch.load(path, weights_only=True).

Usage

Set up the environment and the base models as described in the code repository, then:

huggingface-cli download wasserstein-tilted-flow-maps/WTF --include "*.pt" --local-dir checkpoints
python eval.py t2i/configs/wtf_t2i.yaml --ckpt checkpoints/wtf_t2i_step1460.pt --nfe 4 --gpus 1
python eval.py imagenet/configs/wtf_imagenet.yaml --ckpt checkpoints/wtf_imagenet_seed0_step12500.pt --nfe 1

REPRODUCE.md gives the commands and expected values for every reported number.

Base models

The text-to-image adapters apply to TiM-T2I (TiM, Apache-2.0). The ImageNet weights are fine-tuned from Decoupled MeanFlow XL/2 (MIT).

Citation

@article{mammadov2026wtf,
  title={WTF?! Simulation-Free Reinforcement Learning with Wasserstein-Tilted Flow Maps},
  author={Mammadov, Abbas and Huang, Jerry Y and Lin, Justin and Kaushik, Partha and Shah, Sheel and Nair, Kartik and Teh, Yee Whye and Boffi, Nicholas M},
  journal={arXiv preprint arXiv:2609.27033},
  year={2026}
}