Many recognition systems must do two things at once: name the object in front of them, and notice when it belongs to no class they were trained on. One strong way to notice novelty is to train a generative model on the features of familiar data and flag inputs it cannot reconstruct. The trouble is that these generators are slow. A diffusion model, which learns to turn random noise back into data by removing noise a little at a time, needs many sequential network calls per input. Recent one-step generators remove that cost. Real training data, however, is long-tailed: a few classes have thousands of examples and many have a handful. We suspect the speed-up is not free. The rare known classes are the ones a generator fits worst, and when it has to jump from noise to data in one step, those are the classes it is most likely to reconstruct poorly. A one-step novelty detector would then start calling rare known objects "unknown." Current evaluation would miss this, because open-set ranking metrics are averaged over samples, and head classes supply most of the samples.
This note proposes to measure that hidden cost and then remove it. We will score novelty with generative models run at a range of step counts, from many-step diffusion down to one step, on the features of a long-tailed recognizer. Each detector is judged with a class-balanced version of a coupled open-set metric, so that a rare class counts as much as a frequent one. We then test whether a specific repair lets a one-step detector match the many-step detector on rare classes: making the generator conditional on class and weighting its training loss so every class counts equally. The repair only counts if it beats a simpler fix, which is to rescale each class's reconstruction error after the fact.
Core Claim
We make two claims, each falsifiable. First, when an unconditional latent generator used as a novelty scorer is cut from many steps to one, rare known classes lose far more coupled open-set performance than frequent ones, and sample-averaged OpenAUC hides most of that loss. Second, a class-conditional one-step flow trained with a class-balanced loss recovers most of the lost tail performance at one network call, and does better than post-hoc per-class rescaling of an unconditional scorer. If per-class rescaling does as well, the second claim fails, and the contribution becomes the measurement plus the finding that the fix is cheap. We treat that as an acceptable outcome, not one to explain away.
The three directions play different roles. Iterative inference latency is the variable we control, set by the number of network calls per input. Class imbalance is the regime we test in. Balanced open-set evaluation is the measuring instrument; without it, the effect we are looking for cannot be seen.
What Prior Work Establishes and Leaves Open
Latent reconstruction is the motivating mechanism. Diffusion-based Layer-wise Semantic Reconstruction for Unsupervised Out-of-Distribution Detection (DLSR) trains a diffusion model on pooled multi-layer features of a frozen backbone and flags inputs whose features it fails to rebuild. It reports 99.7% average AUROC on CIFAR-10 and about 999 images per second, against 11.4 for pixel-space DDPM. DOODLE does something similar for LiDAR segmentation, and states plainly that detection quality falls as the number of noise steps is reduced; the authors settle on five steps as a compromise. Both papers train on balanced in-distribution data, and neither has a way to tell a rare known class from an unknown one. That missing piece is the gap we target.
The tools for one-step scoring come from two papers. Mean Flows for One-step Generative Modeling learns the average velocity over a time interval rather than the instantaneous one. This allows generation in a single network call, reaching FID 3.43 on ImageNet 256×256, but it has been tested only as a sampler on balanced data. Classification Diffusion Models obtain exact likelihoods in one forward pass by relating the denoiser to a classifier over noise levels. They were evaluated only on balanced OOD benchmarks. We reuse both, but in a role neither paper studied: novelty scoring under long tails.
Evaluation both constrains and motivates the work. OpenAUC couples the two tasks by counting a pair of one known and one unknown sample as a success only when the known sample is classified correctly and the unknown sample receives the higher novelty score. OpenworldAUC extends this idea to prompt tuning. DeCoOp shows that the harmonic-mean score can rise without any improvement in real mixed-deployment accuracy. We adopt coupled pairwise ranking but change how known samples are weighted, because with a 100:1 head-to-tail ratio the pairs involving head samples dominate the average.
On the long-tail side, Large-Scale Long-Tailed Recognition in an Open World (OLTR) and LUNA handle open long-tailed recognition with discriminative memory and local-density methods. Both evaluate with thresholded F-measure, and neither involves generative scoring or latency. OLTR also concedes that borrowing features from head classes causes confusions between tail classes and related head classes. That warns us against transferring information from head to tail inside the generator, the approach taken by Long-tailed Diffusion Models with Oriented Calibration, so we give each class equal weight instead. Latent-based Diffusion Model for Long-tailed Recognition shows that class-conditional latent diffusion can represent tail features well enough to generate useful synthetic features for them, reaching 44.8% on ImageNet-LT. That makes it plausible that a conditional latent generator can fit tail classes, although LDMLR itself is closed-world and does not use the generator at test time.
Research Program
The measurement is a class-balanced coupled ranking. Let $\mathcal{K}$ be the set of known classes, $X_c$ the test samples of class $c$, $X_u$ the unknown samples, $s(\cdot)$ the novelty score, and $\hat y(\cdot)$ the predicted label. We define
$$\text{bOpenAUC} = \frac{1}{|\mathcal{K}|}\sum_{c\in\mathcal{K}} \frac{1}{|X_c||X_u|}\sum_{x\in X_c}\sum_{x'\in X_u} \mathbb{1}[\hat y(x)=c]\,\mathbb{1}[s(x')>s(x)],$$
which is OpenAUC computed within each known class and then averaged with equal class weights. We report it separately for head, medium and tail groups under the usual frequency cut-offs. The reweighting itself is simple, and we do not present it as the main contribution. It is the instrument that makes the effect of step count on tail classes visible.
The shared setup is as follows. A backbone is trained on the long-tailed known classes with logit-adjusted cross-entropy and then frozen. Following DLSR, we pool its multi-layer features into a single vector per input. The classifier on top of this backbone supplies $\hat y$ for every detector, so the detectors differ only in how they compute $s$.
We then compare a family of feature-space scorers along a latency axis. The first is a flow-matching or DDIM reconstruction scorer at 50, 10, 5 and 2 steps, which is DLSR-style and the reference for quality. The second is an unconditional MeanFlow scorer at one step: the feature is noised to level $t$, mapped back in one average-velocity jump, and scored by cosine discrepancy. The third is a single-pass CDM likelihood scorer. The fourth is our variant, a MeanFlow model conditioned on class and trained with the loss of each class weighted by the inverse of its frequency. Its score is the conditional reconstruction discrepancy under the predicted class, minimized over the top-$k$ predicted classes.
The analysis rests on the curve of bOpenAUC on tail classes against the number of network calls, compared with the same curve for head classes. The first claim predicts that the tail curve falls more steeply as step count drops. We also ask why. For each class we measure the one-step velocity regression error and plot it against training frequency, to check whether the degradation follows data sparsity or instead follows how semantically close the class is to others. If the pattern follows semantic closeness, the fix should target that, not frequency.
Baselines
The do-nothing reference is the frozen logit-adjusted backbone with maximum softmax probability as the novelty score. It costs one forward pass, needs no generator, and sets the floor that every detector must clear. For context we also report thresholded F-measure. On ImageNet-LT open, OLTR reports an F-measure of 0.474, OpenMax 0.368, and the plain model 0.295 (OLTR). On the naturally long-tailed MS-LT, LUNA reports an F-measure of 0.657 (LUNA).
The strongest discriminative competitor is LUNA's weighted center loss together with its local-density novelty score. It was built for exactly this setting of open-set recognition under long tails, and it runs in one pass. We re-run it on our shared backbone data with its reported hyperparameter search. On ImageNet-LT we found no reported result on Lacuna for OpenAUC or bOpenAUC for LUNA, OLTR, or any other method, so we will compute these numbers ourselves for every method.
The strongest generative competitor is multi-step DLSR at its default setting. It is the best available method for latent reconstruction, and its reported 99.7% CIFAR-10 AUROC and 999 images per second (DLSR) come from balanced data. We expect those figures to fall under long tails, but we will not assume they do. On balanced open-set data, OpenAUC reports 75.9% on TinyImageNet (OpenAUC), which provides an anchor for our balanced control split.
To keep the comparison fair, all generators use the same feature extractor and the same training data, the same parameter count, and the same number of GPU-hours. We budget the roughly 16% extra training compute that MeanFlow spends on Jacobian-vector products. Every method gets the same number of hyperparameter trials, selected on a validation split made of held-out known classes, never on the test unknowns. Latency is measured as wall-clock images per second on one GPU at a fixed batch size.
Two ablations isolate our ingredients. One removes the class conditioning, the other removes the balanced loss weighting, and each is run at one step. A third baseline is post-hoc per-class calibration: the unconditional one-step score is z-normalized by the mean and standard deviation of reconstruction error for the predicted class, estimated on training features. This is the most important competitor, discussed below. Our method counts as a success if, at one network call, it matches the 50-step DLSR on tail bOpenAUC within about one point, while being at least ten times faster. It must also beat both the calibration baseline and LUNA on tail bOpenAUC with confidence intervals over three seeds that do not overlap.
Why not just…?
Why not just use a reweighted discriminative classifier with a confidence score? It is the cheapest option, and it could be competitive. However, OLTR's numbers show that confidence-based rejection lags badly under long tails, with OpenMax at 0.368 against 0.474 (OLTR). The explanation is that tail classes have low logits, which makes their confidence look like novelty. We include this baseline with Seesaw-style reweighting (Wang et al.). If it matches the generative scorers on tail bOpenAUC, we will conclude that generative scoring is unnecessary in this regime and report that result.
Why not just rescale each class's reconstruction error after training? This is the real threat. If one-step errors are uniformly inflated on tail classes, a per-class z-score removes the problem with no retraining. Our mechanism predicts something different. We expect that a one-step unconditional model trained mostly on head samples produces tail reconstructions that are pulled toward nearby head classes. That would make errors on tail samples overlap with errors on unknowns in their distribution, not merely shift them upward, and rescaling cannot separate overlapping distributions. The calibration baseline tests this directly. Tail bOpenAUC for calibration versus conditioning at one step decides the question, and if calibration ties, we will say so.
Why not just keep a few diffusion steps? DOODLE settles on five steps as a compromise (DOODLE), and five steps in feature space are already fast. Our step-count sweep shows exactly what each step buys on the tail. If five-step DLSR equals one-step DLSR on tail classes, then the latency trade-off does not matter for this problem, which is itself worth knowing. If they differ, the curve shows how much of the gap our method closes.
Why not just use local-density scoring like LUNA? LUNA's k-nearest-neighbor density is fast and was designed for tails. But it needs search against a stored training set, whose cost grows with that set, and its behavior under coupled pairwise ranking has not been measured. It is run as a full baseline, including its wall-clock cost on ImageNet-LT.
Proposed Experiments
The main benchmarks are ImageNet-LT with ImageNet-2010 open classes, Places-LT with Places-Extra69, and the naturally long-tailed MS-LT marine species set from LUNA. MS-LT contains closely related species, which makes it the most demanding test of whether the method can tell a rare known class from a similar unknown one. CIFAR-100-LT at an imbalance factor of 100 is the known set for fast iteration, with the OOD sets used by DLSR as unknowns. A balanced TinyImageNet 20/180 split is the control, where no tail effect should appear.
Three checks decide the outcome. The first is whether the gap between head and tail in the slope of bOpenAUC against step count is significantly nonzero on ImageNet-LT and MS-LT and close to zero on the balanced control. If so, the latency cost falls on the tail. The second is whether the ranking of methods under standard OpenAUC disagrees with the ranking under tail bOpenAUC, measured by Kendall's tau across all methods and step counts. If so, standard evaluation hides that cost. The third is whether the conditional, balanced one-step model beats per-class calibration and LUNA on tail bOpenAUC at one network call without losing more than a point on head classes. If the first check fails, the premise is wrong and the project reduces to a negative result on generative scoring under long tails. If the third fails because calibration is enough, the recommendation is the cheap fix, and the evidence for it comes from this study.
References
- Diffusion-based Layer-wise Semantic Reconstruction for Unsupervised Out-of-Distribution Detection
- DOODLE: Diffusion-based Out-of-Distribution Learning for Open-set LiDAR Semantic Segmentation
- Mean Flows for One-step Generative Modeling
- Classification Diffusion Models: Revitalizing Density Ratio Estimation
- OpenAUC: Towards AUC-Oriented Open-Set Recognition
- OpenworldAUC: Towards Unified Evaluation and Optimization for Open-world Prompt Tuning
- DeCoOp: Robust Prompt Tuning with Out-of-Distribution Detection
- Large-Scale Long-Tailed Recognition in an Open World
- LUNA: Localizing Unfamiliarity Near Acquaintance for Open-Set Long-Tailed Recognition
- Long-tailed Diffusion Models with Oriented Calibration
- Latent-based Diffusion Model for Long-tailed Recognition
- Seesaw Loss for Long-Tailed Instance Segmentation