Neon: Negative Extrapolation From Self-Training Improves Image Generation

ICLR 2026 Conference SubmissionAnonymous Authors
Generative ModelsSelf-ImprovementWeight MergingImage Generation
Abstract:

Scaling generative AI models is bottlenecked by the scarcity of high-quality training data. The ease of synthesizing from a generative model suggests using (unverified) synthetic data to augment a limited corpus of real data for the purpose of fine-tuning in the hope of improving performance. Unfortunately, however, the resulting positive feedback loop leads to model autophagy disorder (MAD, aka model collapse) that results in a rapid degradation in sample quality and/or diversity. In this paper, we introduce Neon (for Negative Extrapolation frOm self-traiNing), a new learning method that turns the degradation from self-training into a powerful signal for self-improvement. Given a base model, Neon first fine-tunes it on its own self-synthesized data but then, counterintuitively, reverses its gradient updates to extrapolate away from the degraded weights. We prove that Neon works because typical inference samplers that favor high-probability regions create a predictable anti-alignment between the synthetic and real data population gradients, which negative extrapolation corrects to better align the model with the true data distribution. Neon is remarkably easy to implement via a simple post-hoc merge that requires no new real data, works effectively with as few as 1k synthetic samples, and typically uses less than 1% additional training compute. We demonstrate Neon’s universality across a range of architectures (diffusion, flow matching, autoregressive, and inductive moment matching models) and datasets (ImageNet, CIFAR-10, and FFHQ). In particular, on ImageNet 256x256, Neon elevates the xAR-L model to a new state-of-the-art FID of 1.02 with only 0.36% additional training compute.

Disclaimer
This report is AI-GENERATED using Large Language Models and WisPaper (A scholar search engine). It analyzes academic papers' tasks and contributions against retrieved prior work. While this system identifies POTENTIAL overlaps and novel directions, ITS COVERAGE IS NOT EXHAUSTIVE AND JUDGMENTS ARE APPROXIMATE. These results are intended to assist human reviewers and SHOULD NOT be relied upon as a definitive verdict on novelty.
NOTE that some papers exist in multiple, slightly different versions (e.g., with different titles or URLs). The system may retrieve several versions of the same underlying work. The current automated pipeline does not reliably align or distinguish these cases, so human reviewers will need to disambiguate them manually.
If you have any questions, please contact: mingzhang23@m.fudan.edu.cn

Overview

Overall Novelty Assessment

The paper introduces Neon, a method that fine-tunes generative models on self-synthesized data and then reverses gradient updates to extrapolate away from degraded weights. According to the taxonomy, this work resides in the 'Negative Extrapolation and Gradient Reversal' leaf under 'Self-Training and Synthetic Data Feedback Mechanisms'. Notably, this leaf contains only the original paper itself—no sibling papers are listed—suggesting this specific combination of gradient reversal and negative extrapolation from self-training represents a relatively unexplored niche within the broader self-training literature.

The taxonomy reveals that the broader parent category 'Self-Training and Synthetic Data Feedback Mechanisms' includes a sibling leaf focused on 'Model Collapse Prevention Through Data Quality Control', which addresses similar autophagy concerns but through filtering rather than gradient manipulation. Adjacent branches explore 'Constraint-Guided Generation With Negative Examples' (using explicit negative constraints in engineering, vision, and recommendation domains) and 'Hybrid Generative-Discriminative Training Frameworks' (combining generative and discriminative objectives). Neon diverges from these by operating purely at the gradient level without external discriminators or hard constraints, positioning it as a distinct approach within the self-improvement paradigm.

Among the three contributions analyzed, the core Neon method examined two candidates and found one potentially refutable prior work, indicating some overlap in the limited search scope of 22 papers. The theoretical proof of anti-alignment between synthetic and population gradients examined ten candidates with none clearly refuting it, suggesting this formalization may be relatively novel. The demonstration of universality across architectures and datasets also examined ten candidates without clear refutation. These statistics reflect a focused semantic search rather than exhaustive coverage, so the apparent novelty should be interpreted cautiously within this bounded exploration.

Given the limited search scope and the paper's placement in an otherwise-empty taxonomy leaf, Neon appears to occupy a distinct methodological position—combining self-training feedback with gradient reversal in a way not directly captured by the examined prior work. However, the single refutable candidate for the core method suggests that related ideas may exist in the broader literature beyond the 22 papers examined. The theoretical and empirical contributions show fewer direct overlaps within this search, though a more exhaustive review would be needed to assess their full novelty.

Taxonomy

Core-task Taxonomy Papers
6
3
Claimed Contributions
22
Contribution Candidate Papers Compared
1
Refutable Paper

Research Landscape Overview

Core task: improving generative models through negative extrapolation from synthetic data. This field explores how generative systems can be refined by leveraging not only positive examples but also synthetic negative or contrastive signals. The taxonomy reveals three main branches. Self-Training and Synthetic Data Feedback Mechanisms encompasses methods that iteratively refine models using their own outputs, often employing negative extrapolation or gradient reversal to steer away from undesirable generations—exemplified by approaches like Self-improving diffusion[1] and Neon[0]. Constraint-Guided Generation With Negative Examples focuses on incorporating explicit negative constraints or contrastive data to shape the generation process, as seen in Negative Data Constraints[2] and Synthetic Negative Samples[6]. Hybrid Generative-Discriminative Training Frameworks blends generative and discriminative objectives, drawing on ideas from earlier work such as Generative via Discriminative[4] and more recent studies like Hybrid Open-Set[3] and Fake and Square[5], to balance likelihood maximization with classification or rejection of poor samples. Across these branches, a central theme is the trade-off between exploiting model-generated data for self-improvement and avoiding the pitfalls of reinforcing model biases or collapsing onto narrow modes. Many studies explore how to construct informative negative examples—whether by perturbing real data, sampling from the model's own distribution, or using auxiliary discriminators—and how to integrate these signals without destabilizing training. Neon[0] sits squarely within the Self-Training and Synthetic Data Feedback Mechanisms branch, emphasizing negative extrapolation and gradient reversal to push the model away from synthetic failure modes. Its approach contrasts with constraint-based methods like Negative Data Constraints[2], which impose hard rules on generation, and with hybrid frameworks such as Fake and Square[5], which interleave generative and discriminative updates. By focusing on gradient-level steering rather than explicit constraints or separate discriminators, Neon[0] offers a streamlined path to refining generative quality through self-generated negative signals.

Claimed Contributions

Neon method for negative extrapolation from self-training

Neon is a post-processing method that improves generative models by first fine-tuning them on self-synthesized data to obtain degraded weights, then reversing the gradient updates via negative extrapolation. This simple parameter merge requires no new real data, works with as few as 1k synthetic samples, and uses less than 1% additional training compute.

2 retrieved papers
Can Refute
Theoretical proof of anti-alignment between synthetic and population gradients

The authors rigorously prove that mode-seeking inference samplers induce a predictable anti-alignment between synthetic data gradients and real data population gradients. This theoretical result explains why reversing the degradation direction through negative extrapolation reduces the true data risk and guarantees Neon's effectiveness.

10 retrieved papers
Demonstration of Neon's universality across model architectures and datasets

The authors empirically validate Neon across diverse generative model families including diffusion models, flow matching, autoregressive models, and few-step generators on multiple standard benchmarks. On ImageNet 256x256, Neon achieves state-of-the-art FID of 1.02 with only 0.36% additional training compute.

10 retrieved papers

Core Task Comparisons

Comparisons with papers in the same taxonomy category

Within the taxonomy built over the current TopK core-task papers, the original paper is assigned to a leaf with no direct siblings and no cousin branches under the same grandparent topic. In this retrieved landscape, it appears structurally isolated, which is one partial signal of novelty, but still constrained by search coverage and taxonomy granularity.

Contribution Analysis

Detailed comparisons for each claimed contribution

Contribution

Neon method for negative extrapolation from self-training

Neon is a post-processing method that improves generative models by first fine-tuning them on self-synthesized data to obtain degraded weights, then reversing the gradient updates via negative extrapolation. This simple parameter merge requires no new real data, works with as few as 1k synthetic samples, and uses less than 1% additional training compute.

Contribution

Theoretical proof of anti-alignment between synthetic and population gradients

The authors rigorously prove that mode-seeking inference samplers induce a predictable anti-alignment between synthetic data gradients and real data population gradients. This theoretical result explains why reversing the degradation direction through negative extrapolation reduces the true data risk and guarantees Neon's effectiveness.

Contribution

Demonstration of Neon's universality across model architectures and datasets

The authors empirically validate Neon across diverse generative model families including diffusion models, flow matching, autoregressive models, and few-step generators on multiple standard benchmarks. On ImageNet 256x256, Neon achieves state-of-the-art FID of 1.02 with only 0.36% additional training compute.

Neon: Negative Extrapolation From Self-Training Improves Image Generation | Novelty Validation