Neon: Negative Extrapolation From Self-Training Improves Image Generation¶

Conference: ICLR 2026 Oral
arXiv: 2510.03597
Code: github.com/VITA-Group/Neon
Area: Image Generation / Self-Training
Keywords: self-training, model collapse, weight merging, negative extrapolation, FID

TL;DR¶

Neon is proposed as a post-processing method requiring <1% additional training computation. It involves fine-tuning the model on its own synthetic data to induce degradation, followed by negative extrapolation away from the degraded weights. The study proves that mode-seeking samplers cause anti-alignment between synthetic and real data gradients, making negative extrapolation equivalent to optimizing toward the real data distribution. It improves xAR-L to a SOTA FID of 1.02 on ImageNet 256×256.

Background & Motivation¶

Background: The scaling of generative models is constrained by the scarcity of high-quality training data. Self-training using synthetic data is an intuitive solution but leads to Model Autophagy Disorder (MAD / Model Collapse), characterized by rapid degradation in sample quality and diversity.

Limitations of Prior Work: (a) Self-improvement methods like SIMS require 2× inference NFE, large synthetic datasets (100K), and significant extra training computation (20%); (b) DDO requires multiple iterations (16 rounds × 50K samples); (c) Existing approaches lack a unified theoretical explanation for why self-training degrades and how to exploit it.

Key Challenge: While self-training degradation appears wasteful, the direction of degradation itself contains information. If this direction is understood, it can be utilized in reverse.

Goal: Can the degradation signals from self-training be converted into self-improvement signals? Can this be supported with theoretical guarantees?

Key Insight: The authors observe that mode-seeking samplers (temperature <1, top-k, finite-step ODE solvers) bias synthetic data toward high-probability regions of the model distribution, causing the population gradients of synthetic and real data to be anti-aligned (\(\cos\varphi < 0\)). Consequently, reversing the self-training gradient is approximately equivalent to optimizing toward the real data distribution.

Core Idea: Self-training makes a model worse, but the "direction of worsening" is precisely the opposite of the "direction of improvement." Therefore, negative extrapolation can improve the model.

Method¶

Overall Architecture¶

Neon converts the degradation direction of self-training into an improvement signal. The pipeline adds three steps after the initial training: first, sample a small batch of synthetic data \(S\) (approx. 1K–6K) using the base model \(\theta_r\); second, briefly fine-tune on \(S\) to obtain a "deliberately worsened" degraded model \(\theta_s\); finally, perform negative extrapolation in parameter space \(\theta_{\text{neon}} = (1+w)\theta_r - w\theta_s\) (\(w>0\)). This process requires no changes to inference and no new real data, with extra computation under 1% of the original training. Two theories (Anti-alignment Theorem and Risk-reduction Theorem) ensure the reverse direction yields improvement, while a U-shaped budget curve determines the optimal degree of degradation.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Train on Real Data<br/>to get Base Model θ_r"] --> B["Sample Synthetic Data S<br/>(1K–6K images,<br/>mode-seeking sampler)"]
    B --> C["Brief Fine-tuning on S<br/>to get Degraded Model θ_s<br/>(Budget 1–2%, U-shaped curve)"]
    C --> D["Negative Extrapolation Weight Merging<br/>θ_neon = (1+w)θ_r − wθ_s"]
    A -->|"Base Weights θ_r"| D
    D --> E["Final Model θ_neon<br/>Zero Inference Overhead"]

Key Designs¶

1. Negative Extrapolation Weight Merging: Moving away from degradation

Self-training pushes the model toward a worse position \(\theta_s\). Neon reverses this by pushing weights in the opposite direction: \(\theta_{\text{neon}} = \theta_r + w(\theta_r - \theta_s)\). Here, \((\theta_r - \theta_s)\) acts as the "anti-degradation" direction vector. Implementation is a single line of merging code: merged[k] = base[k] - w * (aux[k] - base[k]). Unlike methods like SIMS that require running two models during inference (doubling NFE), Neon produces standard weights with zero additional inference overhead.

2. Anti-alignment Theory (Theorem 1): Why the reverse is the right direction

Negative extrapolation relies on gradient directionality. Let the population gradient of the base model on real data be \(r_d = \nabla_\theta \mathcal{L}_{\text{real}}(\theta_r)\) and on synthetic data be \(r_s = \nabla_\theta \mathcal{L}_{\text{synth}}(\theta_r)\). The paper proves that if the sampler is mode-seeking (monotone reweighting, e.g., temperature \(<1\)) and the model error \(|\varepsilon|\) is sufficiently small, these gradients are anti-aligned:

\[\cos\varphi = \frac{\langle r_d, r_s \rangle}{\|r_d\| \|r_s\|} < 0.\]

This explains why self-training (updating along \(r_s\)) increases real data loss, and why reversing \(r_s\) approximates an update along \(r_d\).

3. Neon reduces population risk (Theorem 2): Guaranteed improvement

The paper provides an existence guarantee: there exists a suitable \(w>0\) such that \(\mathcal{L}_{\text{real}}(\theta_{\text{neon}}) < \mathcal{L}_{\text{real}}(\theta_r)\). The optimal \(w\) can be predicted from the degree of gradient alignment, making Neon a theoretically grounded correction rather than a purely empirical trick.

4. U-shaped Training Budget Curve: Optimal degradation

The duration of fine-tuning for \(\theta_s\) (budget \(B\)) determines the accuracy of the extrapolation direction. If \(B\) is too small, \(\theta_s\) lacks a clear degradation trend, leading to high noise. If \(B\) is too large, the displacement exceeds the validity of the first-order Taylor approximation. Performance follows a U-shaped curve with respect to \(B\), with the optimum typically at 1-2% of the base training effort.

Loss & Training¶

The fine-tuning stage uses the original standard training loss for each respective architecture. The parameter \(w\) is typically chosen in the range \([0.5, 1.5]\), with \(w \approx 0.8\text{-}1.0\) recommended. For class-conditional models, \(w\) should be tuned jointly with the CFG scale \(\gamma\).

Key Experimental Results¶

Main Results¶

Model	Type	Dataset	Base FID	Neon FID	Gain
xAR-L	Flow matching	ImageNet-256	1.28	1.02	-20.3%
xAR-B	Flow matching	ImageNet-256	1.72	1.31	-23.8%
VAR d16	Autoregressive	ImageNet-256	3.30	2.01	-39.1%
VAR d36	Autoregressive	ImageNet-512	2.63	1.70	-35.4%
EDM (cond.)	Diffusion	CIFAR-10	1.78	1.38	-22.5%
EDM (uncond.)	Diffusion	FFHQ-64	2.39	1.12	-53.1%
IMM	Moment matching	ImageNet-256	1.99	1.46	-26.6%

Ablation Study¶

Dimension	Key Findings
Training Budget \(B\)	U-shaped curve: optimal at 1-2% of base training.
Merging Weight \(w\)	\(w=-1\) (standard self-training) degrades; \(w \in [0.5, 1.5]\) provides consistent improvement.
Synthetic Samples	1K is effective; returns diminish after 6K.
Cross-architecture	Synthetic data from one architecture can improve another.

Efficiency Comparison¶

Method	FID (EDM, cond. CIFAR-10)	Extra Computation	Synthetic Samples	Inference Overhead
Ours	1.38	1.75%	6K	None
SIMS	1.33	20%	100K	2× NFE
DDO	1.30	12%	800K	None

Key Findings¶

Cross-Architecture Generality: Applicable without modification to diffusion, flow matching, autoregressive, and moment matching models.
Precision-Recall Tradeoff: Neon primarily improves recall (diversity) while slightly decreasing precision, resulting in a net FID reduction.
Mode-seeking vs Diversity-seeking: If the sampler is diversity-seeking (\(\tau > 1\)), gradient alignment flips, and negative extrapolation fails, confirming theoretical boundary conditions.
SOTA: xAR-L + Neon achieves FID 1.02 on ImageNet 256×256 with only 0.36% extra computation.

Highlights & Insights¶

Elegant "Degradation as Signal" Insight: Converting model collapse into a tool by leveraging the information in the degradation direction is a profound conceptual shift.
Minimal Implementation: The method requires only a single line of weight merging code, with no changes to the inference pipeline or extra overhead.
Theoretical Alignment: The anti-alignment theorem accurately predicts the observed U-shaped curve and the requirement for mode-seeking samplers, making it a rare theoretically driven practical method.

Limitations & Future Work¶

Hyperparameter Tuning: \(w\) and \(B\) require some tuning, and there is currently no automatic selection mechanism.
Domain Scope: Only verified for image generation; not yet tested on NLP (where temperature/top-k are also used) or video generation.
One-time Correction: As a local first-order approximation, it cannot be applied iteratively (multiple extrapolations would invalidate the Taylor expansion).
Quality vs. Diversity: It improves the proportion of samples above a quality threshold rather than increasing the peak generation quality.

vs SIMS: SIMS uses the difference between base and self-trained model scores during inference (2× NFE). Neon performs a one-time merge after training with zero inference cost.
vs DDO: DDO requires 16 iterations and 800K samples, involving far more computation than Neon.
Connection to "Why DPO is Misspecified": Both exploit information from misspecified or degraded directions—repurposing "bias signals" for model refinement.

Rating¶

Novelty: ⭐⭐⭐⭐⭐ "Negative exploitation of degradation" is highly original conceptually and theoretically.
Experimental Thoroughness: ⭐⭐⭐⭐⭐ Comprehensive evaluation across 4 architectures and 3 datasets.
Writing Quality: ⭐⭐⭐⭐⭐ Clear theory, intuitive diagrams, and simple implementation.
Value: ⭐⭐⭐⭐⭐ High potential to become a standard post-training step due to generality and low cost.