Guidance Strength as a Pseudo-Label Filter in Closed-Loop Diffusion Self-Training

Semi-supervised learning often lets a classifier label unlabeled data with its own predictions, called pseudo-labels, and then trains on them. A recent twist adds a generative model to this loop. A conditional diffusion model is trained on real images paired with the classifier's pseudo-labels. A diffusion model here is a generator that learns to turn random noise into samples by removing noise step by step, and "conditional" means it can be asked for a specific class. Its synthetic images are then fed back to the classifier. Run once, this works well: Dual Pseudo Training reaches 59.0% ImageNet top-1 with one label per class. Its authors stop after a single round because they worry that repeating the loop turns a "virtuous cycle" into a "vicious" one. Nobody knows which way the loop goes, or what decides it.

This note argues that one knob decides it: how strongly the sampler pushes toward the requested class (the guidance strength). We claim that the "noise-filtering" property earlier work credited to conditional GANs and the mode dropping that makes generative models unstable are the same operation. Both remove probability mass from regions where classes overlap. Stronger guidance purifies mislabeled classes, but it does so by deleting the ambiguous boundary samples the classifier most needs. The proposal is to make this trade-off explicit and predict the best guidance strength from an estimate of the teacher's confusion. We then test whether setting it this way, per class and per round, keeps a closed teacher–diffusion loop improving where a fixed setting drifts or collapses.

Core Claim

The starting point is a result from Label-Noise Robust Diffusion Models. That paper corrects diffusion training on static noisy datasets using a time-dependent auxiliary classifier. It proves that the distribution a model learns under a noisy label is a mixture of clean-class distributions. Write $\pi_{ij} = p(y=i \mid \hat y = j)$ for the probability that an image pseudo-labeled $j$ truly belongs to class $i$. Then

$$p(x \mid \hat y = j) = \sum_i \pi_{ij}\, p(x \mid y = i).$$

Guided sampling draws from a sharpened density $p(x\mid \hat y=j)^{1+w}/p(x)^{w}$, where $w \ge 0$ is the guidance strength and $p(x)$ is the unconditional image density. Suppose the class components barely overlap and true class $i$ has marginal frequency $\rho_i$. On the region of component $i$, both $p(x\mid\hat y=j)$ and $p(x)$ are proportional to $p(x\mid y=i)$. So the guided sampler for label $j$ produces class-$i$ images with weight

$$s_{ij}(w) \propto \pi_{ij}^{\,1+w}\,\rho_i^{-w}.$$

When the teacher's confusion is diagonal-dominant, meaning the correct class is the largest component for each label, the synthetic purity $s_{jj}(w)$ rises toward one as $w$ grows. This is the precise form of the "filtering" that Generative Pseudo-label Refinement observed informally for conditional GANs on digits.

The approximation fails exactly where components overlap. In those regions the sharpened density drops below the true density, so large $w$ removes near-boundary images. These are the samples Region Uncertainty Quantification identified as the "danger zone" where pseudo-labels are most often wrong and most informative. The next-round teacher's error therefore has two parts. One falls with $w$ (label impurity) and one rises with $w$ (missing boundary coverage). This predicts an interior optimum $w^*$ that increases with the noise level and differs across classes.

The second new element is using this optimum to control a closed loop. Iterating the mixture formula gives a recursion on the confusion matrix across rounds. Anchoring each round with a fraction of real labeled data plays the role of the real-data fraction $\alpha$ in the discounted-sum law of Quantifying Error Propagation and Model Collapse. That paper bounds drift for unconditional recursive diffusion training only. We extend the question to the conditional, label-coupled case it leaves open.

Research Program

The system alternates three steps.

  1. A teacher classifier labels the unlabeled pool.
  2. A conditional diffusion model is fine-tuned on real images with those labels, warm-started from the previous round.
  3. A new teacher is trained on the labeled set, real unlabeled images with its standard semi-supervised loss, and guided synthetic images.

The controlled variable is the per-class guidance strength $w_j^{(k)}$ at round $k$. We estimate $\pi$ from two sources: the teacher's confusion on the small labeled set, with cross-validation across rounds, and agreement between successive teachers. We then set $w_j$ to the smallest value whose predicted purity $s_{jj}(w)$ reaches a target, capped by a coverage floor. The coverage floor is measured as feature-space recall of the synthetic class against real images the teacher confidently assigns to that class. The labeled-set estimate of $\pi$ is noisy at one to four labels per class. We treat that noise as part of the question and report how often the estimate is off enough to hurt.

The analysis has three parts, run in order of cost.

First, a 10-dimensional Gaussian mixture of the kind used by Quantifying Error Propagation and Model Collapse. Here $\pi$, overlap, and $\rho$ are known, so the predicted $w^*$ and the round-to-round confusion recursion can be checked exactly.

Second, CIFAR-10 and CIFAR-100 with injected instance-dependent noise on the unlabeled pool, plus the genuine few-label setting. On the injected-noise runs the true confusion is known and the purity formula can be validated on real images.

Third, ImageNet at 64×64 for multi-round loops, and at 256×256 for two rounds only, under the DPT protocol. The 256×256 runs are limited to two rounds because of compute.

A side study on the digit benchmarks repeats the loop with a conditional GAN in place of the diffusion model. We track the generator's Jacobian condition number, the quantity Odena et al. showed causally precedes GAN mode dropping, alongside synthetic purity. If our claim is right, rising purity and rising condition number should move together, which would settle the tension between PLR's filtering claim and the known instability of GANs.

Baselines

The naive reference points are the teacher alone, with no generative step, and one-round DPT. For one-round DPT, Dual Pseudo Training reports 59.0% ImageNet top-1 at 1 label per class, 69.5% at 2 and 74.4% at 5, FID 3.08 at 256×256, and 4.68% CIFAR-10 error at 4 labels per class. We reuse its teacher and backbone where feasible. Any closed-loop method must beat these numbers at matched total diffusion training steps to justify the extra rounds.

The strongest direct competitor is closed-loop DPT at a fixed global guidance strength. It is the obvious way to iterate DPT. We sweep $w$ on a grid with the same number of tuning runs our controller consumes and report the best fixed value, not a default. Beating it means higher final-round accuracy, and no round-over-round accuracy decline, at matched compute.

Three further competitors each test a different explanation:

  • Confidence-filtered loop. The same loop, with synthetic images the teacher classifies with low confidence discarded. This is the filtering reviewers will propose first.
  • Anchor-only loop. The loop at fixed $w$, with real labeled data mixed in at several ratios. This isolates the Khelifa-style stabilization from our control.
  • TDSM-in-the-loop. Each round's diffusion training uses transition-aware weighted score matching from Label-Noise Robust Diffusion Models. That paper reports class-wise FID improving from 30.45 to 15.92 at 40% noise on CIFAR-10. It is the strongest existing method for training diffusion on noisy labels, but it was designed for static noise and costs about 1.5× training compute. We therefore compare it both at its native cost and at equal compute.

On digits, PLR's reported 97.3% for SVHN→MNIST and 63.4% for MNIST→SVHN are the reference numbers. For FixMatch-style pure semi-supervised training without generation, no reported result was found on Lacuna under the DPT protocol, so we will run it ourselves.

The key ablation separates the prediction from the adaptivity. It has three settings: per-class adaptive $w$; a single global $w$ set by our formula; and the best global $w$ found by the grid sweep.

Why not just…?

Why not just threshold pseudo-labels or synthetic samples by teacher confidence? Thresholding uses the teacher's own judgement to decide what counts as clean. Under structured shift noise the teacher is confidently wrong, as PLR shows (Morerio et al.), and softmax confidence is miscalibrated near the boundary (Wang et al.). Guidance instead purifies toward the majority of each pseudo-labeled group, which is a different signal. This alternative is a real threat, because it also removes boundary samples. The confidence-filtered loop is run at matched compute. If it matches our controller on the injected-noise CIFAR runs, where true purity is measurable, the generative mechanism adds nothing and we will say so.

Why not just mix real labeled data into every round? Anchoring bounds drift through the discounted-sum law (Khelifa et al.). However, with one to four labels per class the anchor is tiny, and anchoring slows the spread of old errors rather than removing them. The anchor-only loop measures how much stability comes from anchoring alone. Our controller is also evaluated with the same anchor fraction, so any gain beyond it is attributable to $w$.

Why not just plug in a transition-matrix correction such as TDSM or rGAN? Label-noise robust GANs assume noise that depends only on the class and a known transition matrix. TDSM handles noise that depends on the individual image, but it assumes the noise is fixed and requires an auxiliary classifier trained across all diffusion timesteps. In a loop, the confusion changes every round, so the classifier would need retraining each round. TDSM-in-the-loop is still run. If it wins at equal compute, the conclusion is that correcting the training objective beats controlling the sampler, which is itself a useful answer.

Why not just tune one global guidance scale? This is the most likely way the adaptive controller fails to matter. With balanced classes and uniform noise, per-class control may add little. The ablation settles it. If the best global $w$ ties with the adaptive version, the contribution shrinks to the purity/coverage analysis and the finding that $w^*$ must grow as noise grows, which is still what practitioners would need to close the loop safely.

Decisive Checks

The idea stands or falls on four results.

  1. The purity formula holds on real images. On injected-noise CIFAR-100, where true labels of synthetic samples can be recovered with an oracle classifier, measured synthetic purity should track $s_{jj}(w)$ as $w$ varies.
  2. The next-round optimum is interior. Next-round teacher accuracy plotted against $w$ should peak at an interior value, and that peak should move right as injected noise rises from 20% to 60%.
  3. Fixed low guidance drifts in the closed loop. On CIFAR-10 at 4 labels per class and ImageNet 64×64 at 1 label per class, the fixed low-$w$ loop should show round-over-round accuracy decline, confirming DPT's "vicious cycle" worry. The controlled loop should keep improving, or at least never fall below round one.
  4. Overlap loss explains the failures at high guidance. Where large $w$ fails, the missing synthetic mass should be concentrated in regions of high teacher uncertainty, measured as feature-space recall restricted to uncertain real images.

If the first check fails, the mechanism is wrong. If the third shows no drift at any $w$, closing the loop was never the problem, and the proposal reduces to a study of guidance as a pseudo-label filter.