Quantifying Error Propagation and Model Collapse in Diffusion Models

arXiv:2602.16601 · stat.ML, cs.LG · Submitted 2026-02-18 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "Quantifying Error Propagation and Model Collapse in Diffusion Models".

Jane: Here's the gist of the paper. Quantifying Error Propagation and Model Collapse in Diffusion Models analyzes how performance degrades when diffusion models are recursively trained on synthetic data by establishing bounds on accumulated divergence in score-based diffusion models The paper provides theoretical characterization of different drift regimes and quantifies error accumulation over generations Model Setup and Error Metrics The study defines a recursive training procedure where the training data in each round is a mixture of fresh samples from the true data distribution and synthetic samples from the current model The effective training distribution for generation i ≥ 1 is defined as the mixture qi:= α pdata + (1 − α) ˆp i, where pdata denotes the true data distribution on R d The goal is to answer how errors accumulate over generations and how the divergence between pˆ i and pdata evolves.

Tom: Two key quantities are tracked, namely the accumulated divergence Di:= χ 2(ˆp i∥pdata) and the intra-generation divergence I i:= χ 2(ˆp i+1∥qi).

Jane: The analysis primarily uses the χ 2-divergence because it yields convenient algebra under the refresh step qi = α pdata + (1 − α) ˆp i.

Tom: Intra-generational Divergence Control The paper establishes bounds on the intra-generation divergence I i by relating it to pathwise score-error energies.

Paper summary: Tom: We started by setting up this recursive training procedure where the training data in each round is a mixture of fresh samples from the true distribution and synthetic samples from the current model > Page two of THIS PAPER — Quantifying Error Propagation and Model Collapse in Diffusion Models <ref:2602.16601#pg3,Error Propagation and Model Collapse in Diffusion Models>

Jane: They then defined two key quantities to track: the accumulated divergence Di, which is how much the learned distribution drifts from the target, and I sub i, which is how different the next generation's data is from what was expected > Page two of THIS PAPER — Quantifying Error Propagation and Model Collapse in Diffusion Models <ref:2602.16601#pg3,Error Propagation and Model Collapse in Diffusion Models>

Lu: The thesis here is to theoretically analyze this phenomenon in score-based diffusion models by obtaining upper and lower bounds on that accumulated divergence between the generated and target distributions > Page one of THIS PAPER — Quantifying Error Propagation and Model Collapse in Diffusion Models <ref:2602.16601#pg1,Quantifying Error Propagation and Model Collapse in Diffusion Models>

Meng: So what are they actually trying to prove with these bounds? Are they trying to show that collapse is inevitable under certain conditions?

Tom: They aren't claiming collapse is inevitable, but rather they are characterizing different regimes of drift based on the score estimation error and the proportion of fresh data used in each generation > Page two of THIS PAPER — Quantifying Error Propagation and Model Collapse in Diffusion Models <ref:2602.16601#pg3,Error Propagation and Model Collapse in Diffusion Models>

Jane: So if a model is trained too aggressively on its own output, it might drift away from what it was originally supposed to learn > Page two of THIS PAPER — Quantifying Error Propagation and Model Collapse in Diffusion Models <ref:2602.16601#pg3,Error Propagation and Model Collapse in Diffusion Models>

Lu: They establish bounds on the intra-generation divergence I sub i by relating it to pathwise score error energies, showing that for small errors, I sub i is equivalent to the score error energy > Page three of THIS PAPER — Quantifying Error Propagation and Model Collapse in Diffusion Models <ref:2602.16601#pg3,Error Propagation and Model Collapse in Diffusion Models>

Meng: That's a nice result because it connects the internal training step directly to those score errors we already know about > Page three of THIS PAPER — Quantifying Error Propagation and Model Collapse in Diffusion Models <ref:2602.16601#pg3,Error Propagation and Model Collapse in Diffusion Models>

Tom: It means we can use the score error energy as a proxy to control how much things get messy in each training step > Page three of THIS PAPER — Quantifying Error Propagation and Model Collapse in Diffusion Models <ref:2602.16601#pg3,Error Propagation and Model Collapse in Diffusion Models>

Jane: So if we control that, we have more confidence that the next model won't be drastically different from what it should be > Page three of THIS PAPER — Quantifying Error Propagation and Model Collapse in Diffusion Models <ref:2602.16601#pg3,Error Propagation and Model Collapse in Diffusion Models>

Conclusion: Tom: So, wrapping up this paper "Quantifying Error Propagation and Model Collapse in Diffusion Models" we've seen that the core idea is about understanding how errors pile up during recursive training > Page three of THIS PAPER — Quantifying Error Propagation and Model Collapse in Diffusion Models <ref:2602.16601#pg3,Error Propagation and Model Collapse in Diffusion Models>

Jane: It’s about establishing bounds on the accumulated divergence between the learned model and the true target distribution over many retraining rounds > Page one of THIS PAPER — Quantifying Error Propagation and Model Collapse in Diffusion Models <ref:2602.16601#pg1,Quantifying Error Propagation and Model Collapse in Diffusion Models>

Lu: The authors provide a framework that allows us to characterize different drift regimes depending on how persistent those score errors are and how much fresh data we mix in each step > Page three of THIS PAPER — Quantifying Error Propagation and Model Collapse in Diffusion Models <ref:2602.16601#pg3,Error Propagation and Model Collapse in Diffusion Models>

Meng: For deployment, this suggests that if we design the training pipeline right, you have a predictable upper limit on performance degradation > Page three of THIS PAPER — Quantifying Error Propagation and Model Collapse in Diffusion Models <ref:2602.16601#pg3,Error Propagation and Model Collapse in Diffusion Models>

Tom: It's about knowing when the model is stable and when it’s heading towards divergence > Page three of THIS PAPER — Quantifying Error Propagation and Model Collapse in Diffusion Models <ref:2602.16601#pg3,Error Propagation and Model Collapse in Diffusion Models>

Jane: So the paper gives us a way to understand the dynamics of these models without needing to read all that math > Page three of THIS PAPER — Quantifying Error Propagation and Model Collapse in Diffusion Models <ref:2602.16601#pg3,Error Propagation and Model Collapse in Diffusion Models>

Naäil B. Khelifa, Richard E. Turner, Ramji Venkataramanan

University of Cambridge

stat.ML, cs.LG

Submitted: 2026-02-18

Updated: 2026-11-04

Comments: Accepted at ICML 2026

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 100/100

The gist: Quantifying Error Propagation and Model Collapse in Diffusion Models analyzes how performance degrades when diffusion models are recursively trained on synthetic data by establishing bounds on…

Key concepts

Accumulated Divergence (Di)
This measures the total divergence between the current model's distribution ($\hat{p}_i$) and the true data distribution ($p_{data}$) over many training rounds. It quantifies how much the model drifts away from reality as it is repeatedly trained on its own synthetic data.
Intra-generational Divergence ($I_i$)
This tracks the divergence between successive models within a single training round. The paper shows this error is controlled by pathwise score-error energies, providing a lower bound that relates the model's discrepancy to the energy of its score errors.
Score Error Energy
This concept measures how much the learned score function (which guides generation) deviates from the true data distribution. The analysis shows that if this energy is summable, accumulated divergence is bounded; otherwise, persistent errors cause divergence to grow indefinitely.
Model Collapse
This refers to the phenomenon where recursive training leads to performance degradation because persistent score errors cause the accumulated divergence ($D_i$) to not converge. The analysis shows that if score errors are large and persistent, the model's performance will eventually be limited by this accumulated error.

Terminology

Summary

Quantifying Error Propagation and Model Collapse in Diffusion Models analyzes how performance degrades when diffusion models are recursively trained on synthetic data by establishing bounds on accumulated divergence in score-based diffusion models The paper provides theoretical characterization of different drift regimes and quantifies error accumulation over generations

Model Setup and Error Metrics

The study defines a recursive training procedure where the training data in each round is a mixture of fresh samples from the true data distribution and synthetic samples from the current model The effective training distribution for generation i ≥ 1 is defined as the mixture qi:= α pdata + (1 − α) ˆp i, where pdata denotes the true data distribution on R d The goal is to answer how errors accumulate over generations and how the divergence between pˆ i and pdata evolves. Two key quantities are tracked, namely the accumulated divergence Di:= χ 2(ˆp i∥pdata) and the intra-generation divergence I i:= χ 2(ˆp i+1∥qi). p. 1

Where the paper says it

We track the evolution of the recursion using two quantities based on the χ 2 -divergence (defined in (5)): • Accumulated divergence, measuring the divergence between the i-th generation model and the true data distribution: Di := χ 2 (ˆp i ∥ pdata). (3) • The intra-generation divergence, measuring the error introduced in one training round: Ii := χ 2 (ˆp i+1 ∥ qi).

The analysis primarily uses the χ 2-divergence because it yields convenient algebra under the refresh step qi = α pdata + (1 − α) ˆp i. p. 1
Where the paper says it

(3) • The intra-generation divergence, measuring the error introduced in one training round: Ii := χ 2 (ˆp i+1 ∥ qi). (4) We work primarily with the χ 2 -divergence because, in the perturbative regime relevant for our analysis, it is equivalent to the KL divergence up to universal constants (see empirical evidence in Figure 4, where our bounds hold for both divergences), while yielding a particularly convenient algebra under the refresh step qi = αpdata + (1 − α)ˆp i that makes the cross-generation recursion explicit.

Intra-generational Divergence Control

The paper establishes bounds on the intra-generation divergence I i by relating it to pathwise score-error energies. Specifically, a lower bound for I i is provided as the product of observability of errors at the endpoint and the energy of the path-wise score error To our knowledge, this is the first lower bound for diffusion models (even without recursive training) quantifying the discrepancy between the learned distribution and the target. p. 2

Where the paper says it

Error Propagation and Model Collapse in Diffusion Models To our knowledge, this is the first lower bound for diffusion models (even without recursive training) quantifying the discrepancy between the learned distribution and the target. Intra-generation Divergence Control via Score Error: Combining the above lower bound with a standard upper bound (Proposition 3.

Furthermore, under certain assumptions (A3 and A4), it is shown that I i is equivalent to the score error energy up to constants. p. 2
Where the paper says it

1), and proving an equivalence between the score error energies on the ideal-path and the learned-path (for small errors), we show in Theorem 3. 5 that in this regime Ii is equivalent to the score error energy (up to constants): χ 2 (ˆp i+1∥qi) ≍ ε 2 ⋆,i.

This equivalence is formalized in Theorem 3. p. 2
Where the paper says it

Intra-generation Divergence Control via Score Error: Combining the above lower bound with a standard upper bound (Proposition 3. 1), and proving an equivalence between the score error energies on the ideal-path and the learned-path (for small errors), we show in Theorem 3.

5, which states that when the perturbative condition ε 2⋆,i ≤ min n 1, η i 8C o holds, then χ 2(ˆp i+1∥qi) is controlled by epsilon 2⋆,i. p. 6
Where the paper says it

Under Assumptions A1- A4, the following holds for generations i ≥ i0. There exist constants c > 0 and C < ∞ (depending only on δ, Cδ, γ, Kγ) such that, when the perturbative condition ε 2 ⋆,i ≤ min{1, ηi 8C } holds, we have: 1 4 ηi ε 2 ⋆,i − C · ε 4 ⋆,i ≤ χ 2 (ˆp i+1∥qi) ≤ 4 ε 2 ⋆,i + cε4 ⋆,i.

Long-horizon Regimes and Accumulation Law

The analysis characterizes the evolution of the accumulated divergence Di over many generations N, identifying a dichotomy based on score error energies. If P i ≥ 0 ε 2⋆,i = +∞, then Di is non-summable and cannot converge to zero. p. 2

Where the paper says it

1 and Theorem 4. 2): (i) if the score errors persist across generations (energies bounded below by a positive constant), then the accumulated divergence DN cannot vanish; (ii) if the score error energies are summable, i.

Conversely, if the score error energies are summable (P i ≥ 0 ε 2⋆,i < ∞), then Di is given by a discounted sum of score error energies from past generations. p. 6
Where the paper says it

We distinguish two cases, assuming small score errors in each generation: if P i≥0 ε 2 ⋆,i = ∞, then we show that (Di)i≥0 are non-summable. On the other hand, if P i≥0 ε 2 ⋆,i < ∞, we show that (up to constants) Di is given by a discounted sum of score error energies from past generations.

The accumulated divergence after a finite number of generations N admits an interpretable geometrically-discounted decomposition up to a bias term. p. 2
Where the paper says it

P k ε 2 ⋆,k < ∞, then DN remains stable (uniformly bounded) with growing N. Moreover, the accumulated divergence after a finite number of generations N admits an interpretable geometrically-discounted decomposition up to a bias term: χ 2 (ˆp N ∥ pdata)+Cbias ≍ N X−1 k=0 (1 − α) 2(N−1−k) | {z } Forgetting (real data) · ε 2 ⋆,k |{z} Accumulation (synthetic data) , which shows that errors from m generations in the past are suppressed by a factor (1 − α) 2m, quantifying the effect of α, the proportion of fresh data used in each training round.

This shows that errors from m generations in the past are suppressed by a factor (1 − α) 2m, quantifying the effect of α, the proportion of fresh data used in each training round. p. 2
Where the paper says it

P k ε 2 ⋆,k < ∞, then DN remains stable (uniformly bounded) with growing N. Moreover, the accumulated divergence after a finite number of generations N admits an interpretable geometrically-discounted decomposition up to a bias term: χ 2 (ˆp N ∥ pdata)+Cbias ≍ N X−1 k=0 (1 − α) 2(N−1−k) | {z } Forgetting (real data) · ε 2 ⋆,k |{z} Accumulation (synthetic data) , which shows that errors from m generations in the past are suppressed by a factor (1 − α) 2m, quantifying the effect of α, the proportion of fresh data used in each training round.

Error Propagation and Model Collapse

The paper derives a lower bound for accumulated divergence Di under persistent errors. If ε 2⋆,i ≥ ε squared > 0 for all sufficiently large i, then lim sup i→∞ Di ≥ α 16(1 + (1 − α) 2) · η ε squared,7. p. 6

Where the paper says it

(ii) A score-error floor implies persistent accumulated divergence. If ε 2 ⋆,i ≥ ε 2 > 0 for all sufficiently large i, then lim sup i→∞ Di ≥ α 16(1 + (1 − α) 2) · η ε 2 , 7

This implies that persistent score errors lead to persistent accumulated divergence. p. 6
Where the paper says it

If P i≥0 ε 2 ⋆,i = +∞, then P i≥0 Di = +∞. (ii) A score-error floor implies persistent accumulated divergence.

Conversely, if the score errors are summable, the accumulated divergence is bounded by a geometrically-discounted sum of square errors. p. 7
Where the paper says it

3, where we derive upper and lower bounds for DN+1 with explicit constants. 2 shows that in the regime where the score errors are small and summable, the accumulated divergence is a geometrically-discounted sum of square errors.

The memory structure shows that at low α, score errors from 10+ generations ago still contribute significantly to the current divergence. p. 8
Where the paper says it

0 Fractional Contribution Memory Structure: How Past Errors Contribute to DN Figure 5. Geometrically-discounted decomposition of accumulated divergence in (23), for pdata = 1 5 P5 k=1 N (µk , σ2 I10) ∈ R 10.

Experimental Validation

The theoretical framework is validated through experiments on a 10-dimensional Gaussian mixture model (GMM), Fashion-MNIST, and CIFAR-10 datasets. In the GMM setting, exact Gaussian closure for the data allows for closed-form expressions for all theoretical quantities. p. 4

Where the paper says it

Consider the linear-Gaussian reverse diffusion dY⋆ i,s = −βY⋆ i,sds + dB¯ s and a linear state-dependent score error of the form ei,s(x) = ai(s)Cix where Ci = C ⊤ i ≻ 0, and ai(s) ≥ 0 is not identically zero. In this setting, one has that Mi T = − R T t0 ai(s)(CiY⋆ i,s) · Bs.

For image datasets, empirical results confirm that memory structure is a universal feature of recursive diffusion training independent of data complexity. p. 1
Where the paper says it

(4) We work primarily with the χ 2 -divergence because, in the perturbative regime relevant for our analysis, it is equivalent to the KL divergence up to universal constants (see empirical evidence in Figure 4, where our bounds hold for both divergences), while yielding a particularly convenient algebra under the refresh step qi = αpdata + (1 − α)ˆp i that makes the cross-generation recursion explicit. Main contributions The recursion induced by training on real-synthetic mixtures involves two competing effects: error mitigation from the fresh data used in each step, and error accumulation from imperfect score learning and sampling.

The observability coefficient η i stabilizes to a positive constant after initial transients, confirming that score errors on distributions closer to pdata project more strongly onto distinguishable directions. p. 2
Where the paper says it

1 and Theorem 4. 2): (i) if the score errors persist across generations (energies bounded below by a positive constant), then the accumulated divergence DN cannot vanish; (ii) if the score error energies are summable, i.

The memory heatmaps test the structural prediction of Theorem 4. p. 8
Where the paper says it

9): Short memory—only the most recent generation dominates, yielding a sharp diagonal structure. The plots are consistent with Theorem 4.

2, showing that the contribution matrix depends only on the structure of error propagation, not on absolute divergence magnitudes. p. 8
Where the paper says it

Error Propagation and Model Collapse in Diffusion Models 0 3 6 9 12 15 18 Source Generation i 0 3 6 9 12 15 18 C u r r e n t G e n e r a tio n N 0 3 6 9 12 15 18 Source Generation i 0 3 6 9 12 15 18 0 3 6 9 12 15 18 Source Generation i 0 3 6 9 12 15 18 0. 0 Fractional Contribution Memory Structure: How Past Errors Contribute to DN Figure 5.

The non-zero values of η i confirm that our lower bound provides meaningful information in the neural network setting. p. 6
Where the paper says it

Proposition 3. 3 provides a matching lower bound in terms of the ideal-path energy ε 2 ⋆,i, up to the observability factor ηi and higher-order terms.

The results are relevant in the regime where the energy of the path-wise score error is small for all but a few generations. p. 2
Where the paper says it

We also provide experiments on both synthetic and realimage datasets to illustrate the validity of our bounds. Our results are relevant in the regime where the energy of the path-wise score error (defined in (10)) is small for all but a few generations.

p. 2
Where the paper says it

We also provide experiments on both synthetic and realimage datasets to illustrate the validity of our bounds. Our results are relevant in the regime where the energy of the path-wise score error (defined in (10)) is small for all but a few generations.

Conclusion and Future Directions

The paper concludes by showing that under square-summability of score errors, there exists an explicit finite constant Dmax such that the accumulated divergence is bounded. The asymptotic bound on Di shows that lim sup i→∞ Di ≤ ρ + √C 0 + K/α cubed 4 − ρ squared. p. 6

Where the paper says it

(ii) A score-error floor implies persistent accumulated divergence. If ε 2 ⋆,i ≥ ε 2 > 0 for all sufficiently large i, then lim sup i→∞ Di ≥ α 16(1 + (1 − α) 2) · η ε 2 , 7

Future work is suggested to establish bounds in the large error regime, account for discretization error, and investigate the limiting distribution of recursively trained models. p. 8
Where the paper says it

Another direction for further work is to take into account the discretization error due to the diffusion being implemented in discrete-time. Finally, a key open question is: is there a limiting distribution to which the model converges when recursively trained, and, if so, how it depends on α and pdata?

The results are relevant in the regime where the energy of the path-wise score error is small for all but a few generations. The paper provides a theoretical framework for understanding model collapse in diffusion models. p. 0
Where the paper says it

Quantifying Error Propagation and Model Collapse in Diffusion Models Na¨ıl B. Khelifa 1 Richard E.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Incorporate an error accumulation management strategy into recursive training pipelines that explicitly accounts for memory decay based on a fresh data proportion alpha. This allows systems to dynamically adjust training schedules or data sampling rates to prevent performance degradation as model collapse occurs, as errors from m generations in the past are suppressed by a factor (1 - α)2m.

  2. Implement score error observability metrics based on the coefficient eta i to diagnose which parts of the learned score error are most likely to manifest in terminal distribution changes. This enables targeted regularization or score refinement during training rounds, focusing computational effort where it yields the highest impact on performance.

  3. Design training protocols that leverage the geometric discounting formula for accumulated divergence, such as Di+1 + Cbias ≥ αη16 SN − (1 − α)2(N−i0)Di0, to establish robust stability guarantees for recursive learning over long horizons. This ensures that the model's performance remains bounded even when trained on its own synthetic output.

Abstract

Machine learning models are increasingly trained or fine-tuned on synthetic data. Recursively training on such data has been observed to significantly degrade performance in a wide range of tasks, often characterized by a progressive drift away from the target distribution. In this work, we theoretically analyze this phenomenon in the setting of score-based diffusion models. For a realistic pipeline where each training round uses a combination of synthetic data and fresh samples from the target distribution, we obtain upper and lower bounds on the accumulated divergence between the generated and target distributions. Notably, to the best of our knowledge, this is the first lower bound on the divergence between the learned and target distributions, even for standard diffusion models. Our results allow us to characterize different regimes of drift, depending on the score estimation error and the proportion of fresh data used in each generation. In a certain regime, the accumulated divergence after several retraining rounds can be expressed as a discounted sum of score estimation errors made at each generation. We also provide empirical results on synthetic data and images to illustrate the theory.

Sources

Related papers