acceptodds
Under review as a conference paper at ICLR 2027

The FID Lottery: Quantifying Hidden Randomness in Generative Model Evaluation

19 views

Abstract

The Fréchet Inception Distance (FID) is the de facto arbiter of image generation, yet most papers report just a single number from a single trained model using a single sampling seed. How reproducible is that number if we retrain the model, or merely resample from it? In this paper, we treat FID as a random variable on a two-axis panel of training and generation seeds, and measure its variance directly on several hundred SiT networks trained on class-conditional ImageNet 256×256. We report surprising findings: (a) Retraining the model using the same recipe with a different seed moves FID 3.2× more (in Inception feature space) than redrawing samples from a fixed network. (b) That gap is driven by three factors: random initialisation, data ordering, and the per-step Gaussian noise of the flow-matching loss. (c) Increasing compute or model size barely tightens the spread, holding the FID coefficient of variation (CoV) inside a 1–2% band. (d) Per-cell classifier-free-guidance tuning halves the spread but perturbs which seeds rank best, and a lucky training seed reaches the same FID with up to 2× less compute than an unlucky one. (e) The lottery extends beyond ImageNet and Inception features: on a text-to-image model retrained from 10 seeds and scored under 10 evaluation seeds each, the training seed carries 76% of the GenEval and 85% of the DPG-Bench variance. Based on these findings, we recommend a new FID evaluation protocol: evaluate under per-cell optimal guidance, treat any gap between two single runs below ≈4% of the baseline FID as inconclusive (the threshold implied by the ≈1.3% CoV floor we measure for SiT/DiT-family latent flow matching), and report an error bar over several training seeds rather than a single FID number.

Primary area:
generative models
Keywords:
diffusion, flow matching, generative models, reproducibility, uncertainty

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_6xAcOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
3
Soundness
1
Read review
Paper summary

This paper does not propose a new method; instead, it studies an important question in the field of generative models: whether variations in random seeds significantly affect the FID metric. Through extensive experiments, the paper provides several findings regarding fluctuations and randomness. Based on these findings, the paper suggests an evaluation protocol.

Critical strengths
  1. The question studied in the paper uncovers a new blind spot, i.e., how reliable the conclusions produced by the existing FID evaluation metric are, which has strong practical significance.

  2. The experimental design is fairly systematic and quantifies the degree of impact, providing strong reference value.

Critical weaknesses
  1. The generality of the conclusions is questionable. The authors claim to quantify the randomness of generative model evaluation, but the current experiments mainly focus on diffusion models and do not consider other frameworks; even if only diffusion models are considered, the current methods are primarily based on flow matching, without considering other types of frameworks such as score matching. If the authors believe that the fluctuation of FID mainly originates from the per-step noise in flow matching, and that the reliability issue is relatively less severe in other frameworks, then the scope of generative models covered in the title, abstract, and introduction of the paper should be appropriately narrowed, rather than being treated merely as a research limitation.

  2. The suggested evaluation protocol is theoretically reasonable, but it does not fully consider practical situations. Reporting error bars over multiple random seeds implies that large models may need to be trained multiple times, which would multiply the cost. Considering the topic of the paper, there is no need to solve this problem, but it should at least be pointed out; alternatively, the actual reliability gains brought by the increased cost could be further estimated.

Questions for authors

Have the authors tried conducting experiments on other types of generative model frameworks, such as DDPM? If so, what were the results?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 6xAc

Thank you for your review, and for recognising that the paper "uncovers a new blind spot" with "strong practical significance", and especially for noting that you "gained a substantially deeper understanding of the problem or question." Below you address your concerns:

W1: "[…] the current methods are primarily based on flow matching, without considering other types of frameworks such as score matching. If the authors believe that the fluctuation of FID mainly originates from the per-step noise in flow matching, and that the reliability issue is relatively less severe in other frameworks, then the scope […] should be appropriately narrowed […]"

Q1: "Have the authors tried conducting experiments on other types of generative model frameworks, such as DDPM? If so, what were the results?"

We argue that the reliability issue is not specific to flow matching. The per-step noise we identify is shared by score matching [1]. Denoising score matching, DDPM and flow matching all draw a fresh Gaussian ε at every gradient step and regress onto it; they differ only in the noising path and the loss weighting.

Since the submission, we have also added an experiment with Drifting Models [2], a one-step generator trained without any diffusion loss. We find that it varies by 1.54% across 7 training seeds, which remains inside our 1–2% band, which shows a degree of generalization for our findings. See updated manuscript. Our findings, therefore, hold on three different model families (latent space class conditional flow matching, pixel space T2I and class conditional Drifting).

Indeed, the score and applicability of our paper is quite broad. Flow matching, including rectified flow, is the training objective of many of today's leading generators: SD3 and FLUX for images, Movie Gen for video, Voicebox for speech and π0 for robot control. The SiT recipe we study is also the base of many recent ImageNet results. A calibrated noise floor for this family is directly relevant to a large part of the generative-modeling field.

[1] Generative Modeling by Estimating Gradients of the Data Distribution, Song et al. Neurips 2019

[2] Generative Modeling via Drifting, Deng et al. NeurIPS 2026

W2: "Reporting error bars over multiple random seeds implies that large models may need to be trained multiple times, which would multiply the cost. […] it should at least be pointed out; alternatively, the actual reliability gains brought by the increased cost could be further estimated."

Actually, our proposed protocol is much cheaper in practice than our evaluation experiments might suggest. We trained 25 seeds per setting to measure the noise floor, so that others do not have to! Now, everyone can use our estimated noise floor with little cost. A relative gain g needs about (4/g)² runs per method to be significant, so a >4% improvement needs just a single run. Extra runs are only required for close calls, where a single run provides not evidence either way. And because the floor does not change with scale, it can be calibrated on small models, before training the large model. We have added a paragraph clarifying this in the updated manuscript.

Thank you for your detailed response. The current revision is meaningful and has addressed some of the concerns, especially my second concern, which has been largely resolved.

However, regarding the first concern, although the drifting model experiments provide some important results, I still believe that the current experiments cannot be claimed as “broad” in the field of generative models. Moreover, the diverse application scenarios listed by the authors do not change the issue that the categories of generative models covered are still relatively limited.

Considering that raising my score would shift my recommendation from rejection to acceptance, I prefer to keep my rating unchanged for now, even though the paper has been improved to some extent.

Thank you for engaging with our response, and for confirming that the revision resolves your second concern. We would like to clarify the remaining point on scope, because we think the new evidence speaks to it more directly than our previous reply made clear.

The Drifting experiment tests the hypothesis behind your scope concern. Your original W1 asked us to narrow the scope if the FID fluctuation "mainly originates from the per-step noise in flow matching" and the issue is "relatively less severe in other frameworks." Drifting is the model family where that hypothesis predicts a smaller lottery. It is trained without any diffusion or flow-matching loss, so it has no per-step regression noise, and it samples in a single step with no ODE integration. Its training-seed CoV is still 1.54%, inside the same 1–2% band as SiT. This is evidence that the lottery does not come from flow matching. That is consistent with our decomposition: random initialisation and data ordering, two of the three sources we identify, are present in any network trained with SGD, whatever its objective.

Our experiments vary the axes along which generative models differ. The three families we calibrate differ in:

  • Training objective: flow matching vs. a non-diffusion drifting loss
  • Representation: latent vs. pixel space
  • Conditioning: class labels vs. free-form text
  • Sampler: multi-step ODE vs. one-step
  • Metric: Inception FID, GenEval, DPG-Bench

The lottery appears in every one of these settings, with a consistent magnitude wherever we measure FID.

These families account for most current practice. The open-weight models that lead today's benchmarks are trained with flow matching. For example, Qwen-Image is pre-trained with a flow-matching objective, as are SD3, FLUX and the main open video models. In a recent snapshot of the text-to-image arena, FLUX 2 variants alone took four of the top 15 places, and the top fully open-licensed model is Qwen-Image. On ImageNet, the methods competing at the top are SiT/DiT-based flow models and one-step generators such as Drifting, which are exactly the two families we calibrate. We therefore think the title describes the evaluation practice the paper studies. The Limitations section already states which families the ≈4% threshold is calibrated on.

On the scale of the study. We understand your hesitation to change your recommendation, but we would note that the study is already extensive by the standards of evaluation papers in generative modelling:

  • Several hundred SiT networks, trained across 76 cells spanning model sizes S to XL and 200k to 2M training steps, with 25 training seeds per setting.
  • A two-axis panel of training seeds × generation seeds.
  • A text-to-image model retrained from 10 seeds and evaluated under 10 evaluation seeds each.
  • 7 independently trained Drifting models.

We are not aware of any prior study that retrains diffusion or flow models at this scale to measure seed variance. Each additional family requires retraining many models from scratch, which is exactly the cost our measured noise floor spares others. We believe the coverage we report is sufficient to support the paper's claims as stated, and we hope you will take this into account in your final assessment.

Thank you again for your time and constructive engagement.

Reviewer_WwFjOpenReview ↗
Rating
2(−1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work investigates the reproducibility of generative model evaluation by quantifying how training and sampling randomness affect Fréchet Inception Distance (FID). Using hundreds of independently trained SiT models across different model sizes and training budgets, the authors show that training-seed variability contributes approximately 3.2 times more FID variation than sampling-seed variability. They identify flow-matching noise, weight initialization, and data ordering as the primary sources of this variation, while finding that increasing model size or training compute does not eliminate the approximately 1–2% relative FID variability. Although optimizing classifier-free guidance reduces this variability, it can alter model rankings, and favorable training seeds can yield apparent computational speedups of up to 2×. Similar training-seed effects are observed on text-to-image benchmarks, including GenEval and DPG-Bench. The authors conclude that FID improvements below approximately 4% between two single runs may be statistically inconclusive and advocate multi-seed evaluation, uncertainty reporting, and standardized guidance tuning to improve the reliability of generative model comparisons.

Critical strengths
  • Comprehensive empirical investigation of training-induced variability The paper systematically quantifies the impact of training and sampling randomness on generative model evaluation using hundreds of independently trained SiT models across multiple model sizes and training budgets. Its decomposition of training randomness into initialization, data ordering, and flow-matching noise provides valuable insights into the sources of FID variability. The extensive experiments and statistical analyses offer compelling evidence that training-seed variability is a substantial and frequently overlooked factor in generative model evaluation.

  • Practical implications for reliable generative model benchmarking The finding that training randomness can produce FID variation comparable to reported methodological improvements has important implications for reproducibility and fair model comparison. The proposed evaluation recommendations, including multi-seed reporting, uncertainty quantification, and guidance-aware evaluation, provide actionable guidance for improving experimental rigor. The additional validation using alternative metrics and text-to-image benchmarks strengthens the relevance of these findings beyond the primary experimental setting.

Critical weaknesses
  • Limited methodological novelty The paper's central contribution is primarily an extensive empirical characterization of seed-induced variability rather than a fundamentally new evaluation methodology. Multi-seed evaluation, variance decomposition, and statistical significance testing are well-established practices in machine learning. Although applying these techniques systematically to diffusion and flow-matching models is valuable, the proposed evaluation protocol largely builds on existing statistical tools. This raises concerns about whether the work offers sufficient methodological novelty beyond its empirical findings.

  • Limited generalizability of the quantitative conclusions The main experiments focus on SiT models trained with flow matching on ImageNet, leaving uncertainty about whether the reported 1–2% FID variability and approximately 4% significance threshold extend to other prominent generative architectures, training objectives, and datasets. Although the authors include a text-to-image experiment and alternative evaluation metrics, these investigations cover relatively limited configurations, and some metrics exhibit substantially different variance behavior. Consequently, the proposed quantitative evaluation guidelines may be too specific to the studied settings to serve as broadly applicable benchmarking standards.

Questions for authors
  • Generalizability of the FID variability threshold The proposed ≈4% threshold for distinguishing meaningful FID improvements is primarily calibrated using SiT models trained with flow matching on ImageNet. To what extent can this threshold be generalized to other generative architectures and training objectives, such as DDPMs, rectified flow models, or latent diffusion models? Is there evidence supporting a similar relative variability floor across these settings?

  • Novelty beyond established variance analysis Multi-seed evaluation, variance decomposition, and statistical significance testing have been extensively studied in machine learning. What is the key conceptual or methodological contribution of this work beyond applying these established techniques to generative models? In particular, does the proposed evaluation framework offer insights or capabilities that cannot be obtained through conventional repeated-run statistical analysis?

  • Robustness of the proposed evaluation protocol The paper shows that classifier-free guidance optimization approximately halves FID variability, but also changes training-seed rankings. Given that the reported guidance-optimal FID values are subject to selection bias from tuning and evaluating on the same samples, how robust are the variance reduction and ranking conclusions when guidance is selected on independent samples and evaluated on held-out generation seeds?

6 comments
AuthorsOpenReview ↗

Response to Reviewer WwFj

Thank you for a careful and constructive review, for finding the evidence "compelling" and reporting that you "learned something interesting or useful.". We address your questions below:

W2: "[…] these investigations cover relatively limited configurations […]"

Q1: "To what extent can this threshold be generalized to other generative architectures and training objectives, such as DDPMs, rectified flow models, or latent diffusion models?"

Actually, we do report results on rectified flow models and latent diffusion models (Sec 3, L189-192). DDPM is very similar to flow matching and we expect results to be very similar. Since the submission, we have also added an experiment with Drifting Models (Deng et al., 2026), a one-step generator trained without any diffusion loss. We find that it varies by 1.54% across 7 training seeds, which remains inside our 1–2% band, which shows a degree of generalization for our findings. See updated manuscript.

W1:"[…] largely builds on existing statistical tools"

Q2: "[…] does the proposed evaluation framework offer insights or capabilities that cannot be obtained through conventional repeated-run statistical analysis?"

The tools are standard by design, so that no conclusion rests on a new estimator. Here is what our work provide above and beyond repeated-run analysis:

  1. The two-axis design separating the training lottery from the sampling lottery. It shows that the error bars that the generative-model papers report, obtained by resampling one network, cover only 30% of the FID spread.
  2. Noise source decomposition, with a same-seed control, identifying the per-step loss noise (absent from supervised training) as the largest source, and ruling out GPU non-determinism.
  3. A transferable threshold: Because the floor does not depend on scale (76 cells from S to XL, 200k to 2M steps), one measurement yields a threshold that applies to existing single-run results without retraining.

Q3: "[…] how robust are the variance reduction and ranking conclusions when guidance is selected on independent samples and evaluated on held-out generation seeds?"

Thanks for an interesting question. The values in §4.4 are in-sample search minima, which the paper flags as optimistic. Below we show three different analyses of existing data:

  1. Selection pressure. A coarser search on the same 250 cells (9 evaluations instead of 14) gives the same between-seed CoV (0.68% vs 0.67%), the same ratio σ_b/σ_w (1.89 vs 1.87) and the same ranking (ρ = 0.98). The extra search lowers FID by only 0.016, uniformly across seeds.
  2. Disjoint samples. Guided rankings from two disjoint halves of the sampling seeds agree at ρ = 0.93. Unguided and guided rankings on disjoint halves agree at only ρ = 0.72, so the reshuffling is a property of the models, not of selection noise.
  3. A bound. Even if held-out evaluation doubled the within-seed spread, the CoV would only rise to 0.70%, still about half of the unguided 1.26%.

These all indicate that the bias shifts the level of guided FID, but not its spread or the seed ranking.

Thank you for the detailed response. The authors have made meaningful efforts to address my concerns by clarifying the empirical contributions, extending the evaluation to Drifting models, and providing additional analyses of guidance-selection bias.

However, two important limitations remain only partially addressed. First, the generalizability of the proposed 4% FID threshold remains uncertain, as the additional Drifting experiment involves only seven training seeds and a single evaluation per model, without directly validating the threshold across broader generative model families. Second, regarding the robustness of guidance optimization to selection bias, the authors provide useful indirect analyses but do not conduct a direct held-out evaluation in which guidance is selected on one sample set and assessed independently on another.

While I appreciate the additional evidence and clarifications, these unresolved issues still limit the strength and generalizability of the conclusions. Therefore, I lean toward a Weak Reject.

Thank you for your careful reading and for acknowledging the additional experiments. Regarding your latest comments, let us clear up some misunderstandings regarding our experiments.

1. Drifting

"the additional Drifting experiment involves only seven training seeds and a single evaluation per model, without directly validating the threshold across broader generative model families."

Each Drifting run costs about 150 H100 GPU-hours, so the 7 seeds represent roughly 1,050 H100-hours, run during the rebuttal period. The goal of our Drifting experiment is NOT to recalibrate the threshold (the main paper already does this, using several hundred SiT networks). The goal is to address your original question: "To what extent can this threshold be generalized to other generative architectures and training objectives?" Our experiment shows that the threshold remains essentially unchanged (4.3% vs. 4.0%) for Drifting, a one-step generator trained with an objective far from the diffusion family.

One evaluation per model is deliberate. The ≈4% threshold applies to a gap between two single runs, each being one training run plus one evaluation, which is what papers in this field report as standard practice (e.g. [1–6]). The spread across these 7 runs measures exactly that quantity.

2. Selection bias in guidance

"the authors provide useful indirect analyses but do not conduct a direct held-out evaluation in which guidance is selected on one sample set and assessed independently on another."

Please note that selecting guidance on one sample set and evaluating on another is not standard practice in the literature. Models trained on Imagenet (e.g. [4, 7–10]) report FID at the guidance setting found by sweeping that same evaluation. §4.4 reproduces this practice on purpose, and the values are labelled as in-sample in the revision. Our analyses also bound the effect of selection:

  • Weaker search, same result. Going from 14 to 9 evaluations leaves the CoV (0.68% vs. 0.67%), σ_b/σ_w (1.89 vs. 1.87) and the ranking (ρ = 0.98) unchanged.
  • Disjoint samples, same ranking. Rankings from two disjoint halves of the sampling seeds agree at ρ = 0.93.
  • Pessimistic bound. Even if held-out evaluation doubled the within-seed spread, the guided CoV (0.70%) would still be about half the unguided one (1.26%).

Selection therefore shifts the level of guided FID slightly, but not its spread or the seed ranking, which are what our conclusions rest on.

On the rating

The rating was lowered from 3 to 2 after our rebuttal provided additional evidence (e.g. the Drifting experiment), with the claims unchanged. We would be grateful to understand what motivated this change, so that we can address it.

Thank you again for your time.

[1] Ma et al. SiT: Exploring Flow and Diffusion-based Generative Models with Scalable Interpolant Transformers. ECCV 2024.
[2] Deng et al. Generative Modeling via Drifting. NeurIPS 2026.
[3] Peebles & Xie. Scalable Diffusion Models with Transformers. ICCV 2023.
[4] Yu et al. Representation Alignment for Generation: Training Diffusion Transformers Is Easier Than You Think. ICLR 2025.
[5] Li et al. Autoregressive Image Generation without Vector Quantization. NeurIPS 2024.
[6] Geng et al. Mean Flows for One-step Generative Modeling. NeurIPS 2025. [7] Karras et al. Analyzing and Improving the Training Dynamics of Diffusion Models. CVPR 2024.
[8] Kynkäänniemi et al. Applying Guidance in a Limited Interval Improves Sample and Distribution Quality in Diffusion Models. NeurIPS 2024.
[9] Karras et al. Guiding a Diffusion Model with a Bad Version of Itself. NeurIPS 2024.
[10] Fan et al. Fluid: Scaling Autoregressive Text-to-image Generative Models with Continuous Tokens. ICLR 2025.

Reviewer_q5A9OpenReview ↗
Rating
3(+2)
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper argues that random seeds at both the training and sampling stages have a non-negligible impact on evaluation metrics such as FID, and that training-seed-induced variation is much larger than sampling-seed-induced variation. The author separates the “training lottery” from the “generation lottery,” and employs a random-effects model to estimate σ_between and σ_within. Moreover, the paper provides multi-seed reporting guidelines and comparison thresholds.

Critical strengths
  1. The paper is well written and easy to follow.
  2. The core argument is intuitive and convincing, identifying a common but overlooked source of instability in current evaluation protocols.
  3. The conclusions are supported by a large-scale empirical study.
Critical weaknesses
  1. The novelty and contribution of the paper are relatively limited: it only analyses the extent to which random seeds affect performance and provides corresponding handling recommendations.
  2. The proposed evaluation method incurs extremely high computational cost and is difficult to generalize.
Questions for authors

Please see weaknesses

2 comments
AuthorsOpenReview ↗

Response to Reviewer q5A9

Thank you for your review, and for finding the paper "well written and easy to follow" and its core argument "intuitive and convincing".

W1: "The novelty and contribution of the paper are relatively limited: it only analyses the extent to which random seeds affect performance and provides corresponding handling recommendations."

Measuring the extent to which random seeds affect performance is precisely the contribution, and your own review says why it matters: the paper identifies "a common but overlooked source of instability in current evaluation protocols". Image generation selects models and adopts recipes on the basis of single FID numbers, often separated by a fraction of a point. Whether such a gap is real could not be answered before, because training-seed variance is invisible from a single trained model. The only way to measure it is to retrain many models, which we did, at the scale of several hundred networks.

The answer is not what the field assumes. Retraining moves FID 3.2× more than resampling, so the error bars papers report capture only a third of the spread. That spread does not shrink with model size or training length, and a lucky seed alone can make a recipe look up to 2× faster. We are not aware of prior work establishing these facts for diffusion or flow models, and would gladly discuss any we have missed.

This is the kind of contribution the ICLR 2027 Reviewer Guidelines ask reviewers to value: "Submissions bring value to the ICLR community when they convincingly demonstrate new, relevant, impactful knowledge (incl., empirical, theoretical, for practitioners, etc)." Our general response (see above) shows that the reviews convincingly answer the guidelines' four key questions. The results are also already being used: e.g. reviewers of two other ICLR 2027 submissions cite this work to gauge the validity of reported FID gains.

W2: "The proposed evaluation method incurs extremely high computational cost and is difficult to generalize."

Actually, our proposed protocol is much cheaper in practice than our evaluation experiments might suggest. We trained 25 seeds per setting to measure the noise floor, so that others do not have to! Now, everyone can use our estimated noise floor with little cost. A relative gain g needs about (4/g)² runs per method to be significant, so a >4% improvement needs just a single run. Extra runs are only required for close calls, where a single run provides not evidence either way. And because the floor does not change with scale, it can be calibrated on small models, before training the large model. We have added a paragraph clarifying this in the updated manuscript.

Regarding method generality, the setting that we study -- flow matching for both class conditional and text-to-image in both latent and pixel space -- is what is currently being used by the vast majority of methods in practice. So our work will be useful as is, even without any generalization. Having said that, since the submission we have also added experiments with Drifting Models (Deng et al., 2026), a one-step generator trained without any diffusion loss. We find that it varies by 1.54% across 7 training seeds, which remains inside our 1–2% band, which shows a degree of generalization for our findings. We updated the manuscript with this result.

We hope these answers address your concerns, and would be grateful if you reconsider your assessment.

Thanks for your responses. I believe that the responses have addressed my main concerns, so I will upgrade my rating to 3.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response: generality, contribution and cost

Phenomenon vs. new method. The most common criticism (q5A9 W1; WwFj W1) is that the paper measures a phenomenon rather than proposing a new method. But the ICLR 2027 Reviewer Guidelines do not require a novel method. They define significance as "new, relevant, impactful knowledge (incl., empirical, theoretical, for practitioners, etc)", and ask reviewers four key questions, which the reviews themselves answer:

Key questionThe reviews
"What is the specific question and/or problem tackled by the paper?""an important question in the field of generative models" (6xAc)
"Is the approach well motivated […]?""intuitive and convincing, identifying a common but overlooked source of instability" (q5A9)
"Does the paper support the claims?""supported by a large-scale empirical study" (q5A9)
"What is the significance of the work? […]""strong practical significance" (6xAc); "important implications for reproducibility and fair model comparison" (WwFj)

We urge the reviewers to follow the ICLR guidelines and consider our contribution in terms of "new, relevant, impactful knowledge".**** Thank you.

Updates to the manuscript:

  • §1, §4.3 and App. H.2 (Table 6): We retrained Drifting (Deng et al., 2026), a one-step generator trained without any diffusion loss, from 7 seeds (B/2 ablation setup, one 50k-image evaluation each, CFG 2.0):
training seed42172331345mean ± sdCoV
FID8.548.428.398.528.378.448.758.49 ± 0.131.54%

Its CoV lies inside our 1–2% band (implied threshold 4.3%, vs. our 4.0%), so the lottery holds across three model families.

  • §3 (L189–192): our SiT objective named as latent rectified flow;
  • App. F (new paragraph "Cost of the protocol"), the recommendations box and §5: the cost of the protocol;
  • §4.4 and App. J.1: guided-FID values labelled as in-sample, with the selection-bias analyses;
  • §5: Limitations scoped to the calibrated families.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.