EbbFlow: PINN-Inspired Progressive Curriculum for MeanFlow Training
Authors: Mengfan Yuan, Xunheng Kuang, Jianbin Chen, Lei He
Organizations: Beijing Institute of Technology (bit.edu.cn) · Tsinghua University (mail.tsinghua.edu.cn) · Northwestern Polytechnical University, Xi’an, China · Tsinghua University (tsinghua.edu.cn)
MeanFlow learns interval-average velocity for few-step generation, but its fixed temporal sampler mixes diagonal and interior supervision throughout training. We interpret temporal coordinates as collocation points: diagonal samples provide a stochastic boundary target, whereas finite intervals supply a differential-consistency target. The diagonal fraction controls how often the consistency component appears, while interval length enters its gradient through an explicit factor. Motivated by this distinction, EbbFlow uses Temporal Supervision Allocation (TSA) to schedule the diagonal fraction and interval geometry. It starts with diagonal supervision, then admits interior samples and expands their maximum interval; a left-anchored interval law preserves coverage across time. The per-sample loss, target, backbone, and inference procedure remain unchanged. In matched class-conditional ImageNet 256×256 runs without classifier-free guidance (DiT-B/2, one training seed per method, 400k updates), EbbFlow achieves 50k-sample FID of 39.65 at one network evaluation and 36.02 at two-evaluation ODE sampling, versus 44.82 and 39.86 for MeanFlow. Gradient probes show a mid-training difference in component geometry. Ablations support allocation scheduling and left-anchored intervals, although a fixed full-horizon variant performs better at one evaluation. These results suggest that temporal sampling is a useful training control whose benefits depend on the inference protocol.
Figures & tables
Figure 1: EbbFlow overview: (a) sampled temporal coordinates; (b) diagonal allocation ρk and horizon hmax(k) ; (c) mean component-gradient cosine at 100k and 200k. Figure 3 gives the full 100k–400k comparison.
Figure 2: Full-gradient diagnostics for MeanFlow and AlphaFlow at 100k, 200k, and 400k. Top: ∠(gMF,gTFM) in degrees; bottom: κ=∥gMF∥/∥gTFM∥ , unlike the component ratio q in Figure 3 .
H←hmax(k),n←⌊Bρk⌋,L←0
for i=1,…,B :
if i≤n : t∼pt,r←t
else :
a,b∼iidLN(−0.4,1)
h←Ha,r←(1−h)b,t←r+h
xt←(1−t)xi+tϵi
Table 3
Table 4
Figure 3: Component-gradient diagnostics at 100k–400k. Left: mean cos(gTFM,gTC) ; right: mean q=∥gTC∥/∥gTFM∥ . Each point averages eight interval bins.
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Paper notation
Implementation quantity
r
t_next or sigma_next
t
t or sigma
xt
x_t
vt
velocity_cfg
uθ(xt,r,t)
pred_mean_velocity
gθ
videos_dudt_mf
Appendix
Table 3: Correspondence between the paper notation and the AlphaFlow training implementation.
Figure 4: Sampler-based exposure before batch rounding: (a) interior fraction and mean interval; (b) their product Sk ; (c) cumulative exposure relative to MeanFlow’s 400k total. AlphaFlow ∗ denotes coordinate exposure, not gradient strength.
1-NFE
2-NFE ODE
Method
FID
FDD
FID
FDD
MeanFlow
44.82
843.8
39.86
805.7
AlphaFlow
41.89
787.8
37.67
772.2
EbbFlow
39.65
796.9
36.02
768.2
Appendix
Table 4: Complete matched ImageNet 256×256 results with DiT-B/2 at 400k updates. AlphaFlow averages two evaluations of the same checkpoint. Lower FID and FDD are better; two-NFE results use ODE sampling.
ρk
Law
Horizon
1-NFE
2-NFE ODE
2-NFE cons.
(a)
0.75
MM
full
44.82
39.86
45.51
(b)
0.82525
MM
full
44.13
38.77
44.78
(c)
1→0.65
MM
full
41.80
37.22
42.35
(d)
0.75
LA
quadratic
40.78
36.92
41.67
(e)
1→0.65
LA
full
38.85
36.25
39.37
(f)
1→0.65
LA
quadratic
39.65
36.02
39.97
Appendix
Table 5: Complete DiT-B/2 ablations at 400k updates, measured by 50k-sample FID. MM denotes min–max intervals and LA denotes left anchoring. Static ρ=0.82525 approximates the time-averaged allocation of 1→0.65 . Cons. denotes two-NFE consistency sampling.
Configuration
Value
Network architecture and data
Backbone
DiT-B/2
Backbone parameters (M)
approximately 131
Transformer depth / hidden dimension
12 / 768
Attention heads / latent patch size
12 / 2×2
MLP expansion ratio
4
Appendix
Table 6: Shared configurations for the matched ImageNet 256×256 experiments. Parameter count refers to the trainable DiT backbone and excludes the frozen autoencoder. The learning rate includes a 10k-update warm-up.
Configuration
MeanFlow
AlphaFlow
EbbFlow
Diagonal timestep law
LN
LN
LN
Interior coordinate law
MM (LN)
MM (LN)
LA (LN)
Diagonal fraction ρk
0.75
0.50
1→0.65
Allocation schedule
constant
constant
sigmoid
Allocation start / end (updates)
—
—
0 / 400k
Allocation steepness γρ
—
—
20
Appendix
Table 7: Temporal-sampler configurations for the matched baselines and EbbFlow. LN denotes LogitNormal(−0.4,1.02) ; MM denotes min–max ordering of two independent draws, and LA denotes Equation 10 . Dashes indicate an inactive schedule.
Figure 5: Uncurated EbbFlow samples: ImageNet class 33 (loggerhead), seeds 0 – 15 , 400k EMA. Top: one NFE; bottom: two-NFE ODE.
Figure 6: Uncurated EbbFlow samples: ImageNet class 88 (macaw), seeds 0 – 15 , 400k EMA. Top: one NFE; bottom: two-NFE ODE.
Flow-based generative modeling learns probability transports from simple noise to data for fast sampling, density estimation, and controllable generation.
Timestep sampling p(t) is a central design choice in Flow Matching models, yet common practice increasingly favors static middle-biased distributions (e.g., Logit-Normal). We show that this choice induces a speed--quality trade-off: middle-biased sampling accelerates early convergence but yields worse asymptotic fidelity than Uniform sampling. By analyzing per-timestep training losses, we identify a U-shaped difficulty profile with persistent errors near the boundary regimes, implying that under-sampling the endpoints leaves fine details unresolved. Guided by this insight, we propose Curriculum Sampling, a two-phase schedule that begins with middle-biased sampling for rapid structure learning and then switches to Uniform sampling for boundary refinement. On CIFAR-10, Curriculum Sampling improves the best FID from 3.85 (Uniform) to 3.22 while reaching peak performance at 100k rather than 150k training steps. Our results highlight that timestep sampling should be treated as an evolving curriculum rather than a fixed hyperparameter.
Flow-based image generative models exhibit stable training and produce high quality samples when using multi-step sampling procedures. One-step generative models can produce high quality image samples but can be difficult to optimize as they often exhibit unstable training dynamics. Meanflow models exhibit excellent few-step sampling performance and tantalizing one-step sampling performance. Notably, MeanFlow models that achieve this have required extremely large training budgets. We significantly decrease the amount of computation and data budget it takes to train Meanflow models by noting and exploiting a particular discretization of the Meanflow objective that yields a consistency property which we formulate into a ``Discrete Meanflow'' (DMF) Training Curriculum. Initialized with a pretrained Flow Model, DMF curriculum reaches one-step FID 3.36 on CIFAR-10 in only 2000 epochs. We anticipate that faster training curriculums of Meanflow models, specifically those fine-tuned from existing Flow Models, drives efficient training methods of future one-step examples.
Chia-Hong Hsu, Frank Wood
Department of Computer Science, University of British Columbia, Vancouver, BC, Canada
Flow-matching-style generative models learn time-dependent vector fields, but standard objectives supervise each timestep independently, leaving shared-path temporal structure unused. We introduce Temporal Pair Consistency (TPC), a simple training-time objective that couples velocity predictions at paired timesteps along the same probability path. TPC leaves the architecture, probability path, sampler, and inference-time computation unchanged. Under matched backbones, training budgets, solvers, and numbers of function evaluations, TPC improves sample quality across flow matching, OT-CFM, rectified flow, DiT-style backbones, and MeanFlow. On CIFAR-10, TPC improves FM from 6.35 to 3.19 FID and OT-CFM from 3.58 to 2.90 FID at the same NFE. Mechanism diagnostics show lower measured gradient variance, higher paired-gradient correlation, and reduced temporal roughness, while random and local pairing controls are substantially weaker. Higher-resolution ImageNet experiments further show consistent gains for DiT-XL/2 at 256 and 512 resolution and for MeanFlow at 256 resolution, all at matched inference cost.