acceptodds
Under review as a conference paper at ICLR 2027

Cohort-Level Control for Conditional ECG Generation

Abstract

Conditional electrocardiogram (ECG) generation allows synthetic cohorts to be specified by diagnosis and patient attributes. However, diagnostic utility alone does not establish whether the ECGs reflect the requested patients. The evaluated baselines respond weakly to requested age and sex, while standard guidance for partly recorded attributes can transfer properties of the recorded population. We develop a conditional flow model that generates ECGs in an autoencoder's latent space and explicitly represents whether each attribute is supplied. A hybrid sampler contrasts supplied values for height and weight, avoiding the missing-state contrast without retraining. On PTB-XL and MIMIC-IV-ECG, the cohorts achieve train-on-synthetic, test-on-real macro AUROC of 0.903 and 0.910, respectively, and respond to age, sex, height, weight and heart rate. A controlled missingness experiment links the unintended shift to signal-dependent recording. On PTB-XL, hybrid guidance removes the observed 4.2-year age shift after a height request without reducing downstream utility.

Primary area:
generative models
Keywords:
conditional generation, electrocardiogram, synthetic cohorts, flow matching, classifier-free guidance, informative missingness, attribute control

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_GiHUOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper "Cohort-level control for conditional ECG generation" addresses the proper integration of patient metadata into generative models for ECG data. The authors observed that existing models only weakly followed requested age and sex attributes. In addition, the authors demonstrate that when attributes, such as height or weight, are frequently missing in the training dataset, generated samples with missing attributes out of this category are biased towards the training set mean with respect to other attributes such as age. The authors fix this problem by means of a hybrid sampling procedure within a conditional flow model and provide experimental evidence that this approach fixed the latter problem.

Critical strengths
  • The authors provide clear experimental evidence that both problems (attribute-following and bias towards the mean for missing values) exist and describe the mechanism that leads to the latter.
Critical weaknesses
  • Related work: A clear comparison of the proposed approach (apart from the hybrid sampling) in comparison to literature approaches such as DiffuSETS, SE-Diff or SSSD-ECG is missing. This makes it very difficult to see which components might be relevant for performance improvements. Also relevant prior work on conditional ECG generation is missing https://openreview.net/forum?id=iTPbysd9s1
  • The authors put forward to problems with conditional ECG generation. They don't propose a specific solution for attribute-following of non-missing attributes such as age and sex, where they use standard guidance. Nevertheless, their experimental results seem to indicate that their approach shows an improved behavior in this respect.
  • Architectural details: A big weakness is the almost total lack of architectural details of the proposed model, which remains limited to a few sentences in the appendix. This includes details on the actual diffusion model, the latent encoder, training objectives. No code is provided and the submission is from my perspective therefore not reproducible at all.
  • Experimental details: The selected label sets are not documented, neither the subset of rare labels. No preprocessing code is provided, so the statements remain impossible to assess.
  • Narrow scope: Conditional ECG generation is already a very narrow domain. The particular aspect of attribute-following in particular related to missing attributes is even narrower. The identified problems exist, see strength #1, but they are so specific that they are only relevant for a very small fraction of the ICLR community.
  • Comparison to prior work: "We ran the official SSSD-ECG implementation without its optional CUDA extension, which substantially increases its measured latency." Comparing CUDA to non-CUDA implementation does not yield a like-by-like assessment of the proposed method compared to literature approaches. Similarly, the comparison to DiffuSETS involves a conditional encoder that was trained by the authors (and also remains undescribed). No comparison in strict alignment with literature results is performed which would allow me to assess the prospects of the proposed model.
  • Ill-defined comparison: The submission suggest comparability to related work. However, when conditioning on more attributes (e.g. providing demographic attributes in addition to diagnostic labels or providing height and weight in addition to age, sex and diagnostic labels) constrains the samples to be generated more, which might quite naturally lead to better sample quality. From this perspective a comparison between generative models conditioned on different label sets are not reliable.
  • Privacy: The authors do not mention nor investigate the privacy aspect of things. A model that copies training set samples would score great under their "train on synthetic, test on real" metric. Improved model performance could be due to similarity to training samples.

Summary

  • The scope of the submission is very narrow. Architectural and experimental details are almost completely missing and therefore hardly reproducible. Comparison to prior work is partially ill-defined.
Questions for authors
  • How does the overall setup compare to competing diffusion/flow matching approaches on a specific level?
  • Why does their model show better attribute-following also for age and sex, where the proposed hybrid sampling is not used?
  • What is the authors' mechanistic understanding of the improved attribute-following behavior of their model compared to literature approaches?
  • Would the approach be applicable to other problems in order to broaden the impact of the submission?
1 comment

We thank the reviewer for their time and for the detailed list of points.

Comparison with prior approaches (W1, Q1). The adapted DiffuSETS and SE-Diff share our latent space. Our generator differs from them in three components: a flow-matching objective with a transformer backbone instead of a diffusion UNet; diagnoses and attributes as structured inputs with presence indicators instead of a text embedding with raw age, sex and heart rate; and classifier-free guidance on age and sex, where the baselines sample without guidance. SSSD-ECG is a signal-space diffusion model with structured state-space layers and label conditioning. DCDM-ECG is relevant prior work on conditional ECG generation.

Age and sex (W2, Q2, Q3). Age and sex use standard guidance; contribution (i) is the generator. Its response does not come from the presence flags or factor dropout (age response 0.63 without both; Table 6), nor from the latent space, which the adapted baselines share. The baselines sample without guidance, as in their released code; we consider this the main reason, but have not isolated it.

Reproducibility (W3, W4). Objective and sampler: Sec. 4.1; transformer: Appendix B; parameter counts: Table 8. Settings not stated in the submission: an autoencoder latent of 4 × 128; 20k iterations at batch 128 on PTB-XL and 60k at batch 512 on MIMIC-IV-ECG; AdamW with learning rate 2e-4 and weight decay 1e-5; condition dropout 0.1 and factor dropout 0.15; 4 attention heads and patch size 4. Label sources and counts are in Appendix A and Sec. 5.1, and the rare subsets (training prevalence below 2%) in the captions of Tables 4 and 5; the individual statements are not listed, and we will add them. Code will be released on publication.

Comparison with the literature (W6). Cost is compared in FLOPs, which the CUDA extension does not change (16,710 GFLOPs per sample for SSSD-ECG, against 52.5 for ours). The text encoders are off-the-shelf and frozen, not trained by us. DiffuSETS embeds text through its authors' hosted service, and MIMIC-IV-ECG text may not be sent to an external service under its data use agreement, so the ports use text-embedding-3-small on PTB-XL and a local bge-large-en-v1.5 on MIMIC-IV-ECG. Neither DiffuSETS nor SE-Diff reports TSTR; on heart-rate error, which both report, the ports reach 6.8 and 8.2 bpm on PTB-XL against 6.73 and 8.24 in the original papers. SSSD-ECG enters with its published TSTR of 0.840, against 0.903 for ours.

Condition sets (W7). Without demographics our model reaches TSTR 0.904 (0.901 with them; Appendix E), so the additional conditions do not explain the gap.

Privacy (W8). General response, item 4.

Scope (W5, Q4). Eq. 2 holds for any classifier-free guidance over presence indicators, for example image or EHR-conditioned generation with partly recorded demographics.

Summary. The shift arises when height is supplied for patients whose height was not recorded, and it moves samples towards the recorded population (older patients), not towards the mean (Sec. 1, Sec. 5.4).

Reviewer_uCJWOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies conditional 12-lead ECG generation in which both diagnoses and patient attributes (age, sex, height, weight, heart rate) should be respected. It trains a 7.5M-parameter latent flow-matching transformer with a presence indicator for each attribute, and argues that with partly recorded attributes the "absent" condition mixes unrecorded and dropped-out patients, so classifier-free guidance adds a presence contrast that can shift unrequested properties. A hybrid sampler therefore contrasts the requested height and weight with fixed reference values. On PTB-XL and MIMIC-IV-ECG, TSTR macro AUROC is 0.903 and 0.910 against adapted DiffuSETS and SE-Diff, attribute response is measured with frozen predictors, and a 4.2-year age shift after a height request on PTB-XL disappears with the hybrid sampler.

Critical strengths
  1. I like the the missingness insight: the absent condition in classifier-free guidance mixes patients who were never measured with patients whose value was dropped, so guidance carries a presence contrast when recording is informative. This is a real ml point beyond ecg, backed by a controlled intervention, a negative control and an independent signal measurement.
  2. The cohort-level evaluation, with the same frozen instruments applied to generated and real ECGs and response slopes against the real reference, is a useful standard for conditional ECG generation, and the generator is small and cheap compared with the baselines. I liked that.
Critical weaknesses
  1. I find methodological novelty for ICLR is limited. The generator is a standard latent flow-matching transformer, and the hybrid sampler is the established contrast between two condition values, as the authors say themselves. The missingness analysis is a mixture identity, the classical missing-indicator bias carried into guidance. The insight is nice, but it is a small step on the method side. 2, The leak is shown for one attribute in one dataset, PTB-XL height. mimic-iv is a negative control with no leak, the weight leak is not reported, and the dose-response experiment is synthetic (hiding age by signal amplitude. The general claim about partly recorded attributes needs more real cases imo.
  2. at the mean-height request used to show that the 4.2-year shift disappears, the hybrid's height term is exactly zero, so the sampler simply ignores height ((which is acknowledged in sec 4.4). The paper can't both present the removal of the shift as a main result and say that the mean-height experiment cannot establish height control. The shift needs to be measured at non-reference heights, for example on patients with a recorded height, where the real ECG is known. Am I missing something? 4.The scope is narrow for ICLR… ECG at 100 Hz on two datasets. The missingness issue is general to conditional generation with incomplete metadata (medical images with partly recorded demographics, EHR-conditioned generation), so for ICLR it should be shown beyond ECG. As it stands, it fits better applied or engineering venues like ieee.
  3. I find the related work very limited. It misses ECG generation work at ML venues, conditional time-series generation from heterogeneous metadata, and the generative-modelling literature on data missing not at random which is exactly the mechanism studied!!
  4. The comparisons … DiffuSETS and SE-Diff are re-implemented in the authors' latent space (with a local text encoder on MIMICivECG), se-Diff loses its simulator prior and retrieval, and sssd-ECG enters mostly through its published number. The near-zero age response of DiffuSETS, a model designed to take patient information, could well come from the adaptation. Time Weaver is not compared, and the leak fix is compared only with imputation and complete-case training, not with guidance variants that change the unconditional branch (eg autoguidance or independent condition guidance). In the end, this is your selling point. 7.Attribute response is measured only with learned predictors, so the generated ecgs may match the predictors without matching the physiology. Known relations should be checked with standard ECG measurements, for instance longer QTc in women, lower QRS voltage with higher weight, and PR prolongation and left-axis shift with age, and 100 Hz limits the morphology. Memorization is not assessed, although the motivation is replacing restricted clinical data. And on held-out diagnosisage groups the generator misses badly, which undercuts the use case in the introduction, subgroups that are hard to assemble.
Questions for authors

My main concerns are the weaknesses above:

  1. I find methodological novelty for ICLR is limited. The generator is a standard latent flow-matching transformer, and the hybrid sampler is the established contrast between two condition values, as the authors say themselves. The missingness analysis is a mixture identity, the classical missing-indicator bias carried into guidance. The insight is nice, but it is a small step on the method side. 2, The leak is shown for one attribute in one dataset, PTB-XL height. mimic-iv is a negative control with no leak, the weight leak is not reported, and the dose-response experiment is synthetic (hiding age by signal amplitude. The general claim about partly recorded attributes needs more real cases imo.
  2. at the mean-height request used to show that the 4.2-year shift disappears, the hybrid's height term is exactly zero, so the sampler simply ignores height ((which is acknowledged in sec 4.4). The paper can't both present the removal of the shift as a main result and say that the mean-height experiment cannot establish height control. The shift needs to be measured at non-reference heights, for example on patients with a recorded height, where the real ECG is known. Am I missing something? 4.The scope is narrow for ICLR… ECG at 100 Hz on two datasets. The missingness issue is general to conditional generation with incomplete metadata (medical images with partly recorded demographics, EHR-conditioned generation), so for ICLR it should be shown beyond ECG. As it stands, it fits better applied or engineering venues like ieee.
  3. I find the related work very limited. It misses ECG generation work at ML venues, conditional time-series generation from heterogeneous metadata, and the generative-modelling literature on data missing not at random which is exactly the mechanism studied!!
  4. The comparisons … DiffuSETS and SE-Diff are re-implemented in the authors' latent space (with a local text encoder on MIMICivECG), se-Diff loses its simulator prior and retrieval, and sssd-ECG enters mostly through its published number. The near-zero age response of DiffuSETS, a model designed to take patient information, could well come from the adaptation. Time Weaver is not compared, and the leak fix is compared only with imputation and complete-case training, not with guidance variants that change the unconditional branch (eg autoguidance or independent condition guidance). In the end, this is your selling point. 7.Attribute response is measured only with learned predictors, so the generated ecgs may match the predictors without matching the physiology. Known relations should be checked with standard ECG measurements, for instance longer QTc in women, lower QRS voltage with higher weight, and PR prolongation and left-axis shift with age, and 100 Hz limits the morphology. Memorization is not assessed, although the motivation is replacing restricted clinical data. And on held-out diagnosisage groups the generator misses badly, which undercuts the use case in the introduction, subgroups that are hard to assemble.
1 comment

We thank the reviewer for a thorough and engaged review, and for the concrete tests it proposes, which we ran.

Non-reference heights. The submitted test was close to the reference (167 against 165 cm). Both suggested measurements are in the general response, item 1: away from the reference the hybrid's shift is −0.8 / −1.8 y at 175 / 185 cm, against +3.6 / +2.7 y for standard guidance, and for patients with a recorded height its ECGs read 2.0 y younger than their real ECGs (standard guidance: 0.4 y older).

Novelty and real cases. The missing-indicator bias is known from regression; the contribution is its consequence for classifier-free guidance, where the absent state adds a presence contrast that shifts unrequested properties of the samples (Eq. 2), tested interventionally. Since the sampler's operator is credited to earlier work (Sec. 2), the paper's contribution is the presence contrast it exposes and reduces, which no earlier work on guidance describes. PTB-XL weight gives a second, weaker case: supplying 70 kg for patients without a recorded weight shifts age by −1.4, −1.6 and −0.1 y over three runs; weight-recorded patients are younger, as the mechanism predicts. On MIMIC-IV-ECG, supplying the mean weight shifts age by +0.2 to +0.5 y.

Comparisons. The ports use off-the-shelf frozen text encoders for data-use reasons (reply to GiHU, W6). DiffuSETS's patient-information input responds in our port (heart-rate response ratio 0.74), and the ports' PTB-XL heart-rate errors are close to those of the original papers (reply to d6Bv, Q3). As in their released code, the baselines sample without guidance, which we consider the main reason for their weak age response. Independent condition guidance behaves like a fixed reference at the population mean (+0.3 / −0.5 / −1.4 y at 167 / 175 / 185 cm); we did not evaluate autoguidance. Time Weaver's published TSTR (0.85) uses a different protocol.

Related work. We agree that the related work should cover generative modelling under data missing not at random (MisGAN, GAIN, not-MIWAE, MissDiff), guidance for partly missing conditions (CoInD), conditional time-series generation (Time Weaver) and ECG generation at ML venues. The mechanism studied here is closest to the first two; those methods model or impute the missing values, whereas the paper studies how the absent state enters guidance.

Physiology and memorisation. General response, items 3 and 4.

Scope and held-out groups. The evidence is ECG-only, and 100 Hz limits morphology; both are stated limitations (Sec. 6). On held-out diagnosis–age groups the large misses are in age and in two of the five diagnoses (Table 7), and the introduction states the subgroup use case too broadly.

Reviewer_d6BvOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper builds a conditional flow model in an autoencoder latent space that generates ECG cohorts conditioned on diagnoses and five patient attributes (age, sex, height, weight, heart rate), with an explicit presence indicator per attribute. It argues that standard factored guidance for a partly recorded attribute contrasts the requested value against the "absent" state, and that this contrast mixes a value contrast with a "presence contrast" that can transfer properties of the recorded population when recording depends on unconditioned signal features. The proposed hybrid sampler replaces the absent-state contrast for height/weight with a value-to-value contrast against a fixed reference, without retraining. On PTB-XL and MIMIC-IV-ECG it reports TSTR macro AUROC 0.903/0.910 and response to all five attributes; a controlled missingness experiment supports the mechanism, and on PTB-XL hybrid guidance removes a 4.2-year predicted-age shift after a mean-height request without reducing downstream utility.

Critical strengths

(1) The mechanism is tested interventionally, not just observationally: varying a signal-dependent hiding-rate gap produces a monotone shift (0 to -31%) with random-hiding and conditioned-variable controls showing none, MIMIC-IV-ECG serves as a natural negative control, and the shift replicates on a predictor-independent signal quantity (RMS amplitude). (2) Unusually honest reporting: the paper reports results against its own method (worse Fréchet distance under hybrid, weight-response overshoot of 1.67 on MIMIC, failure on held-out age–diagnosis combinations) and compares against imputation and complete-records alternatives that also reduce the shift, not only against standard guidance.

Critical weaknesses

(1) The headline leak experiment is conducted exactly where the hybrid contrast vanishes: the shift is measured when adding population-mean height, while the hybrid reference is 165 cm, and the paper notes that at the reference value the two inputs coincide and their difference is zero — so removing the shift there may reflect applying no height guidance rather than removing the presence contrast. No leak measurement at non-reference heights (e.g. 175/185 cm) jointly with height response is reported. (2) The causal attribution for natural height missingness is not established: the controlled experiment implants signal-dependent hiding for age, while for height the evidence is correlational (recorded patients are older; probe AUROC 0.85–0.88) and may be explained by recorded confounders (site, era, diagnosis mix, BMI) not conditioned on. (3) The utility claim rests on weakened adapted baselines — both text-conditioned baselines were re-implemented in the authors' latent space with components replaced or removed (SE-Diff's TSTR varies from 0.473 to 0.782 with its simulator weight) — while hybrid vs standard guidance itself differs negligibly on TSTR (0.903 vs 0.901), so the utility advantage comes from the backbone, not the proposed control method. Method novelty is also limited: value-to-value contrast is an application of known minimally-different-condition CFG, and the decomposition is an algebraic identity over a standard missing-indicator representation.

Questions for authors

(1) At requested heights of 175/185 cm, what age shift does hybrid guidance produce, jointly with its height response? (2) After conditioning on richer recorded covariates (site/era, diagnosis, sex, weight/BMI), how much of the height probe AUROC and the 4.2-year shift remains? (3) What are the original (non-adapted) baselines' TSTR under this protocol, or will you scope the utility claim to adapted implementations? (4) Do the slope ratios and leak conclusions replicate with an architecturally independent set of attribute predictors? (5) How sensitive are leak, height response and Fréchet distance to w_X and the reference value?

1 comment

We thank the reviewer for the time and care taken with the paper, and for precise questions.

Q1. General response, items 1 and 5. Over 155–185 cm, the height reader rises 2.2 cm under standard guidance and 4.4 cm under the hybrid, against 1.6 cm for real patients of fixed age and sex, so the hybrid applies height guidance away from the reference.

Q2. General response, items 2 and 5. The shift persists within each recorded covariate; joint stratification over all of them leaves too few patients in both groups.

Q3. The utility claim applies to the adapted implementations, as Sec. 6 states. On PTB-XL the ports reach heart-rate errors of 6.8 (DiffuSETS) and 8.2 bpm (SE-Diff), against 6.73 and 8.24 in the original papers. SE-Diff's 0.473 is its released loss weighting in our pipeline (Fréchet distance 193). As Sec. 5.2 states, the utility gain comes from the generator; the hybrid's effect appears when height and weight are supplied (TSTR 0.889 for the hybrid against 0.873 for standard guidance).

Q4. The age reader (trained on CODE) is architecturally independent of the sex, height and weight readers (XResNet on PTB-XL); heart rate and RMS amplitude use no learned predictor. The standard measurements (item 3) agree in sign with the predictor slopes.

Q5. In a single sweep at w_X = 2 / 3 / 5 / 8, the shift at 185 cm is −1.3 / −1.9 / −2.8 / −4.2 y (averaged over the three runs of item 1: −1.8 y at w_X = 5), the height reader at 185 cm gives 172.4 / 173.1 / 174.3 / 175.6 cm, and the Fréchet distance is 19.5 / 19.7 / 20.3 / 21.7 (standard guidance 15.6). Moving the reference to 155 / 175 cm moves the drift with it (−3.6 / −1.7 y at 185 cm), with Fréchet distances of 20.6 / 20.5.

Novelty. Sec. 2 credits the value-to-value contrast to prior work. The contribution is the failure mode it addresses: classifier-free guidance over presence indicators, the standard way to condition on incomplete metadata, transfers properties of the recorded population when recording is informative. The paper derives this from Eq. 2 and tests it interventionally. Since the operator is credited to earlier work (Sec. 2), the paper's contribution is the presence contrast it exposes and reduces, which no earlier work on guidance describes.

Comments on the paper

1 comment

To all reviewers

We thank the reviewers for their time and for detailed, specific reviews. We ran the additional experiments they requested; the results are summarised here, and the individual replies refer to them.

1. Shift at non-reference heights (d6Bv, uCJW). Age shift in years on test patients without a recorded height (n = 600; mean ± SD over three runs), relative to height left absent:

Requested height155 cm167 cm175 cm185 cm
Standard guidance+4.9 ± 0.9+4.2 ± 0.8+3.6 ± 0.8+2.7 ± 1.0
Hybrid (reference 165 cm)+0.7 ± 0.3−0.2 ± 0.1−0.8 ± 0.5−1.8 ± 0.9

Real ECGs at fixed age and sex show no age change with height (+0.01 y per cm, SE 0.04). The standard-guidance shift is present at every height. The hybrid removes it near the reference and reduces it at 155–175 cm. Further from the reference it adds a drift of opposite sign that grows with distance and with w_X; at 185 cm this drift approaches the standard shift in size. For patients with a recorded height, generated at their own heights, the hybrid's ECGs read 2.0 y younger than their real ECGs (standard guidance: 0.4 y older), because its base term is the absent state.

2. Recorded confounders (d6Bv). Height recording in PTB-XL is determined largely by site and device (AUROC 0.98). The standard-guidance shift remains positive within every site, device, era, weight-recorded and diagnosis stratum (+1.8 to +4.6 y, all 95% CIs above zero). Eq. 2 requires only that recording depend on the signal beyond the supplied conditions; site and device are the route.

3. Standard ECG measurements (uCJW). Real against generated (standard / hybrid): QTc, women minus men, +9.0 against +10.4 / +8.6 ms; PR +0.45 against +0.35 / +0.34 ms per year; QRS axis −0.95 against −1.37 / −1.30° per year; QRS voltage −0.39 against −0.76 / −0.97 mV per 10 kg. All signs match; axis with age and voltage with weight are over-expressed.

4. Memorisation (uCJW, GiHU). Generated PTB-XL ECGs lie 1.10× as far from their nearest training record as real test ECGs, and at most 0.006% fall below the real 1st percentile (MIMIC-IV-ECG: 0.95×, 1.3% against 1% expected).

5. Corrections. (a) The height response ratios in Tables 2 and 3 divide by a real slope that mainly reflects the sex mix across heights. With age and sex held fixed, standard guidance follows height at 1.2–1.4× real rather than below it, and the hybrid at w_X = 5 at about 3× (close to real at w_X = 2). The hybrid's effect is the reduction of the presence shift (item 1), and the near-zero shift in the submission is a near-reference result. (b) Within weight-recorded strata the presence probe falls to 0.54: it mainly reads whether weight was recorded, which the generator is given, so it is not evidence for the mechanism. The mechanism rests on the controlled dose experiment and the stratified shifts (item 2).

Contribution. (i) A small latent flow generator with the highest diagnostic utility and age and sex response among the evaluated implementations (TSTR 0.903 / 0.910, against 0.811 / 0.829 for DiffuSETS, the stronger adapted baseline), at a third of the sampling cost. (ii) The identification of a presence contrast in classifier-free guidance over presence indicators, which transfers properties of the recorded population when recording is informative, tested interventionally, with the shift now also observed within every recording stratum. (iii) A training-free sampler that reduces this shift, with its range and costs reported. Items (ii) and (iii) apply components credited to earlier work in Sec. 2 to a presence contrast in classifier-free guidance that no earlier work describes.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.