Why the One‑Step “Drift Model” by He Kaiming’s Team Was Rejected at ICML (Scores 5‑4‑4‑3)

Earlier this year He Kaiming’s team introduced the drift model , which shifts the distribution evolution from the inference stage to the training stage so that after training only a single forward pass is needed for generation. This contrasts with diffusion and flow‑matching models that require many inference steps.

The paper reports impressive results on ImageNet 256×256: the latent‑space version (L/2) reaches an FID of 1.54 with 1 NFE, and the pixel‑space version (L/16) attains an FID of 1.61 with 1 NFE. Both generators are trained from scratch and are not distilled from multi‑step models.

ICML 2026 ultimately rejected the submission. The four reviewers gave scores of 5, 4, 4, 3. While they praised the one‑step generation quality and the clarity of the writing, they raised two major concerns.

1. Dependence on a feature encoder : the drift field relies on similarity between samples, which is weak in raw high‑dimensional image space. The authors therefore compute the drift loss in a pretrained self‑supervised feature space (latent‑MAE) during training only; the encoder is not used at inference. Ablation experiments show that the choice of encoder dramatically affects FID, and without any encoder the ImageNet version fails to run.

2. Contribution of the drift field versus classifier‑free guidance (CFG) : reviewers asked how much of the quality gain comes from the drift field itself and how much from CFG. The authors added a no‑CFG ablation, finding that the model still trains but scores drop noticeably, especially when the feature encoder is weak, indicating that CFG provides a clear boost while the drift objective remains effective.

Additional components such as feature normalization, drift normalization, multi‑temperature scaling, CFG conditioning, and a sample queue also appear in the implementation. Reviewers requested further ablations to isolate each component’s impact.

The paper positions the drift field among existing theoretical tools—DMD, MMD, GANs, score estimation, Sinkhorn distance, and gradient flow—arguing that many implicit generative models can be interpreted as inducing a drift field. Reviewers asked for a clearer exposition of these relationships; the authors responded that they make the drift vector V explicit to enable design, normalization, and cross‑feature aggregation.

On the theoretical side, the authors prove that when the drift field is zero the model achieves distribution matching, but the converse does not hold in general except under specific constructions, a point highlighted by the reviewers.

Further controversies include the high variance of Monte‑Carlo estimates of the drift field in high‑dimensional uniform sampling, the failure of the method in raw pixel space, and the sensitivity to kernel and bandwidth choices. The authors note that attempts with different kernels and bandwidths still could not make the raw‑pixel version run.

Despite the rejection, the authors argue that the work remains valuable: high‑quality generation does not necessarily require complex multi‑step inference; training can shoulder the heavy distribution evolution. The drift model thus offers a concrete, well‑documented step toward one‑step generative synthesis.