mean_flow_base_4(pretrained: bool = False, overrides: object = {})The ablation backbone — Base width, patch 4.
Model Size
Parameters
pretrainedbool= FalseAccepted for signature parity with the rest of the zoo. No
MeanFlow checkpoint ships — the paper releases none, and a
True here would have nothing to load.**overridesobject= {}Optional
MeanFlowConfig field overrides.Returns
MeanFlowModelThe average-velocity network.
Notes
Reference: Geng, Deng, Bai, Kolter, and He, "Mean Flows for One-step Generative Modeling", arXiv:2505.13447, 2025, Table 4.
This is the configuration every ablation in Table 1 runs on, at 80
epochs rather than the 240 the reported models get — its FID of 61.06
is a comparison point between design choices, not a headline number.
Patch 4 over a 32-wide latent leaves 64 tokens against B/2's 256,
which is what makes a six-way study affordable.
Examples
>>> from lucid.models import mean_flow_base_4
>>> config = mean_flow_base_4().config
>>> config.patch_size, config.num_patches
(4, 64)