acceptodds
Withdrawn from ICLR 2027

CORE: Learning Contextually Grounded Representations for Pixel-Space Diffusion

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Recent self-distillation methods improve latent-space diffusion by aligning noisy representations with cleaner targets, often derived from an exponential moving average (EMA) teacher. However, these gains do not consistently transfer to pixel-space generation, where models must learn spatial structure and fine-grained appearance directly from noisy pixels. We find that high feature similarity can persist even when spatial correspondence deteriorates sharply, revealing that representation alignment alone does not ensure utility for pixel prediction. To address this limitation, we introduce **CORE**, a spatial self-supervision framework for learning contextually grounded representations in pixel-space diffusion. *Correlated Dual Observations* provide paired noisy and cleaner views of the same image with correlated noise for training a shared generator. Built on these observations, *Contextual Representation Grounding* uses two coupled prediction tasks on a shared representation at masked locations. One predicts cleaner features from visible context, while the other predicts correction targets formed by selecting and scaling cleaner-induced prediction changes according to ground-truth pixel residuals. Together, these tasks encourage contextual recoverability while linking representation learning to pixel prediction errors. CORE requires neither an external representation encoder nor an independent EMA teacher and leaves inference unchanged. Experiments on ImageNet at and demonstrate that CORE achieves consistently lower FID than the evaluated self-supervision baselines. At 600 epochs, CORE-L/16 even achieves a lower FID than PixelREPA, which uses external representation supervision. Code will be publicly released.

Primary area:
generative models
Keywords:
Pixel Space Diffusion, Image Generation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_GQniOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies self-supervised representation learning for pixel-space diffusion. The authors observe that self-distillation methods that align noisy features with cleaner targets do not consistently improve pixel-space generation, and further show that high feature similarity can coexist with poor spatial correspondence. To address this, they propose CORE, which combines Correlated Dual Observations (CDO) with Contextual Representation Grounding (CRG). CDO trains a shared generator on paired noisy and cleaner observations, while CRG uses masked contextual prediction through Contextual Feature Prediction (CFP) and Pixel-Guided Correction (PGC) to encourage spatially grounded representations tied to pixel prediction errors. CORE is evaluated on ImageNet using JiT backbones at 256×256 and 512×512 resolutions and is compared with JiT and several self-supervision baselines.

Critical strengths
  1. The paper studies an interesting and relevant problem: representation-alignment and self-distillation methods that are effective in latent-space diffusion do not consistently transfer to pixel-space models such as JiT. The analysis showing that high feature similarity can coexist with poor spatial correspondence provides a useful motivation for revisiting representation supervision in pixel-space diffusion.
  2. Unlike the adapted latent-space self-distillation methods, which do not consistently improve over JiT in pixel space, CORE consistently achieves better performance than vanilla JiT across the evaluated settings.
Critical weaknesses
  1. The current comparisons do not establish training efficiency under a matched compute budget. The methods are mainly compared after the same number of training epochs, although their per-step training costs differ substantially. In particular, CORE requires two full generator evaluations (CDO) and an additional shared-prefix evaluation (CRG), while the compared self-supervision baselines also have different auxiliary training costs. Therefore, comparisons at the same number of epochs do not correspond to matched training cost. Reporting training FLOPs, GPU-hours, or compute-matched performance across methods would provide a clearer assessment of training efficiency.
  2. The ablation suggests that the contribution of CRG is relatively limited compared with CDO. On JiT-B/16, CDO alone improves FID from 4.45 to 4.17, while adding the full CRG further improves it only to 4.09; meanwhile, IS decreases from 260.7 to 257.2. Since CRG is the component directly motivated by contextual grounding, stronger evidence would be helpful to establish that this relatively small incremental gain is robust and meaningful.
  3. There are a few minor reporting and presentation inconsistencies. Some JiT settings in Tab. 1 and Tab. 2 report FID but omit the corresponding IS values. In addition, the “No Image” points in the spatial-retrieval panel of Fig. 1(c) appear vertically misaligned despite belonging to the same intervention row.
Questions for authors
  1. Can the authors provide a compute-matched comparison, particularly between CORE and vanilla JiT, using training FLOPs or GPU-hours rather than primarily comparing methods after the same number of training epochs? Given the additional computation introduced by CDO and CRG, this would help clarify whether CORE still provides an advantage under a comparable training budget.
  2. Can the authors provide additional evidence that CRG contributes meaningfully beyond CDO-only? In Tab. 3, adding the full CRG only improves FID from 4.17 to 4.09, while IS decreases from 260.7 to 257.2.
5 comments
AuthorsOpenReview ↗

Response to Reviewer GQni(Comments 1-2)

We thank the reviewer for the careful review and for recognizing the value of our research question and diagnostics, and CORE's consistent gains over JiT. We have added experiments addressing training cost, component contributions, and result presentation.

1. Compute-matched

To compare the training cost of CORE with other self-supervised methods, we measured steady-state training step times. CORE's context prediction and pixel correction tasks do introduce additional computation per step. However, all losses are jointly optimized rather than updating parameters separately for each branch. Therefore, the number of optimizer updates remains unchanged at the same number of training epochs and effective batch size. Moreover, the auxiliary components are used only during training, so the network architecture, sampling procedure, and computational cost at inference remain unchanged. The comparison is shown below:

MethodStep time ↓ (ms)Throughput ↑ (images/s)
JiT93.1510,992
SRA113.569,018
Self-Flow120.138,524
DTS + AS131.747,773
SRA + DTS + AS136.627,495
CORE165.166,200

Compared with SRA+DTS+AS, CORE adds 28.54 ms per step, approximately 20.9%. This overhead comes from the additional spatial supervision computation.

To further examine the relationship between training cost and generation quality, we use the measured per-step time ratio above as a reference and select 338 epochs as an additional evaluation checkpoint based on 600 × 93.15 / 165.16 ≈ 338. We report the CORE-L/16 results at this checkpoint below:

MethodEpochsReference relative training costFID ↓
JiT-L/166001.00×2.36
CORE-L/16338≈1.00×2.20
CORE-L/166001.77×2.02

At 338 epochs, CORE-L/16 already outperforms JiT-L/16 trained for 600 epochs, demonstrating better generation quality with fewer training epochs and parameter updates. After 600 epochs, its FID further improves to 2.02, surpassing the 2.11 reported by PixelREPA-L/16 with external representation supervision.

MethodExternal representation supervisionFID ↓
PixelREPA-L/16Yes2.11
CORE-L/16No2.02

At a larger scale, CORE-H/16 achieves 1.77 FID with approximately 953M generator parameters, outperforming the 1.82 FID reported by JiT-G/16 with approximately 2B parameters. These results further support CORE's effectiveness as model size increases and its value in improving generation quality without adding inference-time components.

MethodGenerator parametersFID ↓
JiT-G/16≈2B1.82
CORE-H/16≈953M1.77

2. Ablation

Thank you for highlighting the relative contributions of CDO and CRG in the original ablation. The original 200-epoch experiments reflect a shorter training stage and do not fully capture the differences between components under longer training. We therefore extended the JiT-B/16 ablation experiments to 600 epochs:

MethodFID ↓
JiT3.66
CDO-only3.58
CORE3.38

At 600 epochs, CORE achieves 3.38 FID, compared with 3.58 for CDO-only, showing a larger additional gain than at 200 epochs. Thus, the observation that “most gains come from CDO” at the shorter training stage does not persist under longer training: the gains of the complete method cannot be attributed solely to an additional forward pass, and CRG provides more pronounced additional benefits after longer training.

CDO primarily provides CRG with spatially corresponding observations of the same image at different noise levels. The cleaner observation supplies the targets for contextual feature prediction and candidate directions for pixel correction. Both paths also receive generation supervision, keeping the learning of these paired observations connected to the denoising task. CRG further turns this information into context prediction and pixel-error-guided supervision. In other words, CDO provides the observations needed for supervision, while CRG determines how this information is used to shape generative representations.

Moreover, CORE-L/16 reaches 2.02 FID, outperforming the PixelREPA-L/16 with external representation supervision (2.11). CORE-H/16 achieves 1.77 FID with ≈953M parameters, surpassing JiT-G/16 (1.82, ≈2B). These results support CORE's scalability without additional inference components.

AuthorsOpenReview ↗

Response to Reviewer GQni(Comments 3-4)

3. Fig. 1(c)

Thank you for pointing out the misalignment of the “No Image” points. This results from inconsistent vertical placement in the figure: the two points correspond to the same input intervention, not different experimental settings. We will correct their vertical positions and label layout to avoid confusion.

Meanwhile, to show more directly how training supervision responds to spatial information, we added an experiment analyzing the sensitivity of supervision to spatial information. Holding the model weights, queried region, and each method's own supervision targets fixed, we exchange nearby and distant image patches within the context accessible to each method. We then measure the resulting changes in gradient directions at intermediate features and shared backbone parameters.

MethodFeature-gradient changeBackbone-parameter gradient change
JiT0.26370.5396
Self-Flow0.27890.4541
SRA0.3080.4797
SRA+DTS+AS0.06380.2556
CORE0.5650.8508

Under this diagnostic setting, CORE exhibits stronger responses at both levels. This indicates that rearranging spatial context changes not only CORE's feature-level supervision but also, more substantially, the learning signal propagated to the generative backbone.

This analysis more directly reflects our motivation: pixel-space generation requires not merely high feature similarity, but training supervision that can exploit spatial context and establish a connection to pixel prediction. We will use this analysis as the new Fig. 1(c) in the revision and update both the figure layout and the accompanying text.

4. IS reporting and result selection

The missing IS values in Tables 1 and 2 correspond to the 200-epoch settings of JiT-L/16 and JiT-B/32, respectively. The original paper we cite reports only FID for these two configurations, and the corresponding 200-epoch checkpoints have not been officially released. We therefore have not supplemented their IS values. We will explicitly state the reason for these missing entries in the revision.

Regarding the decrease in IS from 260.7 to 257.2 after adding CRG in the 200-epoch ablation, we would like to clarify that we select the evaluation configuration with the best FID and report its corresponding IS, rather than independently selecting the best FID and the best IS. To examine whether this observation persists under longer training, we added the following 600-epoch ablation results. The table below reports results under a common evaluation protocol, including our re-evaluation of JiT, with IS reported at each method's best-FID configuration:

MethodFID ↓Corresponding IS ↑
JiT$^{*}$3.6266269.372
CDO-only3.5845265.4418
CORE3.3784269.4496

∗^{*} JiT is re-evaluated under the same protocol as CDO-only and CORE.

Under the same result-selection rule, adding CRG improves FID from 3.5845 to 3.3784, while the corresponding IS also increases from 265.4418 to 269.4496. This shows that the IS decrease observed at 200 epochs is not an inevitable trade-off introduced by CRG. With longer training, the complete CORE achieves both better FID and better IS than CDO-only. Figure 4 in the original manuscript presents the FID–IS trade-off curves across different CFG values at both 200 and 600 epochs, showing the overall relationship between these two metrics.

Thank you for the additional experiments. Although the compute-matched L/16 result is helpful, I believe compute-matched comparisons with JiT are also necessary for H/16 and the settings in Tab. 2, where the gains at the same number of epochs are smaller. The 600-epoch ablation shows a clearer benefit from CRG, but CDO-only improves FID over JiT by only 0.08 while requiring more computation. Could you compare JiT, CDO-only, and CORE at the same training cost?

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

The paper addresses the challenge of representation learning in pixel-space diffusion models. While self-distillation methods (such as SRA and Self-Flow) effectively improve latent-space diffusion, their performance gains do not reliably transfer to pixel-space generation, where models must directly learn spatial structure and fine-grained visual details from noisy pixels

To solve this issue, the paper proposes CORE, a spatial self-supervision framework consisting of two core components:

Correlated Dual Observations (CDO): Generates paired noisy and cleaner views of the same image with correlated noise to train a shared generator.

Contextual Representation Grounding (CRG): Replaces selected patch embeddings with learned mask tokens in early Transformer layers and evaluates two complementary prediction tasks on a shared representation at masked locations: - Contextual Feature Prediction (CFP): Predicts cleaner target features from visible context to enforce contextual recoverability. - Pixel-Guided Correction (PGC): Predicts correction targets formed by selecting and scaling cleaner-induced prediction changes based on ground-truth pixel residuals, connecting representation learning to pixel prediction errors

Critical strengths

Clear Problem Formulation & Diagnostic Insight: Uncovers a fundamental gap between latent-space and pixel-space self-distillation through controlled input interventions, empirically proving that high feature similarity does not guarantee spatial utility.

Architecture-Native & Lightweight Training: Requires no external vision encoders (e.g., DINOv2 or CLIP) or independent EMA teacher models. It introduces only 0.297M auxiliary parameters (an 87.4% reduction compared to baseline projection heads)

Strong Empirical Results & Scalability: Achieves state-of-the-art FID scores across multiple Transformer backbones (JiT-B/16, L/16, B/32) and resolutions (256 x 256 and 512 x 512), beating external representation supervision (PixelREPA) at 600 epochs

Critical weaknesses

Gains are not compute-matched, and most of the gain comes from the least novel component.

Each CORE step runs two full forward/backward passes plus a KK-block prefix, roughly 2.2–2.4× JiT's training FLOPs. Yet all comparisons are at matched epochs. A rough comparison illustrates the problem: CORE-B/16 at 200 epochs (4.09) costs about as much as JiT at ~450–480 epochs, and JiT at 600 epochs already reaches 3.66.

In the ablation, CDO alone (a second, cleaner-timestep sample of the same image trained with weight η=0.25η=0.25) accounts for 0.28 of the 0.36 FID gain. This looks largely like additional supervision per step.

Questions for authors

What are the per-step training FLOPs and wall-clock time of CORE vs. JiT and SRA+DTS+AS? Can you show FID vs. training compute, or JiT trained for a compute-equivalent number of epochs (~450–480 for B/16)?

Can you report a control that matches CDO's extra supervision without the correlated cleaner view? For example, JiT with two independent timestep samples per image (same ηη), and CDO with independent noise ϵ2ϵ2​ only.

3 comments
AuthorsOpenReview ↗

Response to Reviewer RJUW (Comments 1)

We thank the reviewer for your time and careful review, and for recognizing our clear problem formulation and diagnostic analysis, lightweight supervision design without an external visual encoder, and experimental performance and scalability. Your suggestions regarding training cost and component contributions have been valuable in improving this work. We have added experiments and analyses in response and address your concerns below.

1. Compute-matched

To compare the training cost of CORE with other self-supervised methods, we measured steady-state training step times. CORE's context prediction and pixel correction tasks do introduce additional computation per step. However, all losses are jointly optimized rather than updating parameters separately for each branch. Therefore, the number of optimizer updates remains unchanged at the same number of training epochs and effective batch size. Moreover, the auxiliary components are used only during training, so the network architecture, sampling procedure, and computational cost at inference remain unchanged. The comparison is shown below:

MethodStep time ↓ (ms)Throughput ↑ (images/s)
JiT93.1510,992
SRA113.569,018
Self-Flow120.138,524
DTS + AS131.747,773
SRA + DTS + AS136.627,495
CORE165.166,200

Compared with SRA+DTS+AS, CORE adds 28.54 ms per step, approximately 20.9%. This overhead comes from the additional spatial supervision computation.

To further examine the relationship between training cost and generation quality, we use the measured per-step time ratio above as a reference and select 338 epochs as an additional evaluation checkpoint based on 600 × 93.15 / 165.16 ≈ 338. We report the CORE-L/16 results at this checkpoint below:

MethodEpochsReference relative training costFID ↓
JiT-L/166001.00×2.36
CORE-L/16338≈1.00×2.20
CORE-L/166001.77×2.02

At 338 epochs, CORE-L/16 already outperforms JiT-L/16 trained for 600 epochs, demonstrating better generation quality with fewer training epochs and parameter updates. After 600 epochs, its FID further improves to 2.02, surpassing the 2.11 reported by PixelREPA-L/16 with external representation supervision.

MethodExternal representation supervisionFID ↓
PixelREPA-L/16Yes2.11
CORE-L/16No2.02

At a larger scale, CORE-H/16 achieves 1.77 FID with approximately 953M generator parameters, outperforming the 1.82 FID reported by JiT-G/16 with approximately 2B parameters. These results further support CORE's effectiveness as model size increases and its value in improving generation quality without adding inference-time components.

MethodGenerator parametersFID ↓
JiT-G/16≈2B1.82
CORE-H/16≈953M1.77
AuthorsOpenReview ↗

Response to Reviewer RJUW (Comments 2)

2. Ablation

Thank you for highlighting the relative contributions of CDO and CRG in the original ablation. The original 200-epoch experiments reflect a shorter training stage and do not fully capture the differences between components under longer training. We therefore extended the JiT-B/16 ablation experiments to 600 epochs:

MethodFID ↓
JiT3.66
CDO-only3.58
CORE3.38

At 600 epochs, CORE achieves 3.38 FID, compared with 3.58 for CDO-only, showing a larger additional gain than at 200 epochs. Thus, the observation that “most gains come from CDO” at the shorter training stage does not persist under longer training: the gains of the complete method cannot be attributed solely to an additional forward pass, and CRG provides more pronounced additional benefits after longer training.

CDO primarily provides CRG with spatially corresponding observations of the same image at different noise levels. The cleaner observation supplies the targets for contextual feature prediction and candidate directions for pixel correction. Both paths also receive generation supervision, keeping the learning of these paired observations connected to the denoising task. CRG further turns this information into context prediction and pixel-error-guided supervision. In other words, CDO provides the observations needed for supervision, while CRG determines how this information is used to shape generative representations. The L/16 and H/16 results above also show that the effectiveness of the complete method is not limited to the B/16 scale.

Regarding observation correlation, our paper uses an empirically chosen pairing strategy that facilitates the construction of cleaner observations. Correlated observations support our spatial supervision, while the central contribution lies in connecting supervision more directly to the generation task through context prediction and pixel-error guidance. To examine the practical role of this pairing strategy, we additionally constructed a complete CORE variant using independent timesteps and independent noise as a control, again using JiT-B/16 and training for 600 epochs:

MethodFID ↓
CORE (original correlated observations)3.38
CORE (w/o correlated observations)3.54

The original correlated observations reduce FID from 3.54 with independent timesteps and noise to 3.38, supporting the practical value of this pairing strategy.

AuthorsOpenReview ↗

Updated Manuscript and Responses to Your Concerns

Dear Reviewer,

Thank you again for your time. We have submitted a revised manuscript incorporating the additional experiments and clarifications discussed in our responses to your concerns.

We would appreciate hearing whether these updates address your concerns.

Best regards,

The Authors

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

This paper examines why representation alignment methods do not consistently transfer from latent-space to pixel-space diffusion, showing that high feature similarity can coexist with poor spatial correspondence. It proposes CORE, which combines correlated noisy and cleaner views with masked feature prediction and pixel-error-guided correction supervision. Experiments on ImageNet show improved FID without additional inference cost.

Critical strengths

A strength of the paper is its insightful analysis of representations in latent and pixel spaces. However, it lacks a concrete theoretical explanation for the observed differences.

Critical weaknesses
  1. All experiments use JiT, leaving it unclear whether CORE generalizes to other pixel-space diffusion architectures. Evaluating it on another backbone, such as PixelDiT, would strengthen the claim of broader applicability.

  2. Although the paper evaluates larger models and higher resolutions, experiments remain restricted to class-conditional ImageNet generation. Its effectiveness in more diverse settings, such as text-to-image generation, remains unexplored.

  3. The analysis reveals a discrepancy between feature similarity and spatial correspondence, but does not establish whether this discrepancy explains the limited effectiveness of existing alignment methods in pixel space. A clearer theoretical account or targeted controlled experiments would strengthen this connection.

  4. CORE requires two full generator passes and an additional prefix pass per training step. Comparisons at matched epochs therefore do not isolate the benefits of the proposed method from those of additional training computation.

  5. CORE is motivated by high feature similarity coexisting with poor spatial retrieval, yet the paper does not evaluate CORE under the same spatial retrieval protocol used to establish this issue. Such an evaluation would directly test whether the method addresses its motivating problem.

Questions for authors

See above weakness.

3 comments
AuthorsOpenReview ↗

Response to Reviewer dVGV (Comments 1–3)

We thank the reviewer for your time and careful evaluation, and for recognizing the insights from our analysis of representation differences between latent and pixel spaces. Your specific suggestions regarding applicability, training cost, and validation of our motivation have been valuable in improving this work. We have conducted additional experiments and analyses and respond point by point below.

1. Another baseline

Given the limited rebuttal period, we selected PixelU (ECCV 2026) as another lightweight and representative pixel-space generation backbone. Specifically, we incorporated CORE into PixelU-B/16 and trained it on ImageNet 256×256 for 160 epochs. We retained PixelU's original architecture and configuration, adding only CORE's training supervision without changing the inference architecture.

MethodFID ↓IS ↑
PixelU-B/164.39230.41
PixelU + CORE3.68280.71

Compared with the published results, CORE reduces FID by 0.71 and increases IS by 50.30. This provides further evidence of CORE's applicability to different pixel-space generation backbones and shows that its supervision can complement the backbone's existing representation-alignment mechanism. Because PixelDiT takes longer to train, completing its evaluation within the limited rebuttal period is challenging. We therefore report the completed PixelU-B/16 experiment here. We will continue experiments with PixelDiT and other larger backbones and include the results in subsequent versions.

2. T2I

We further applied CORE to MiniT2I-B/32 to evaluate its effectiveness in text-conditioned generation. The experiments use the same pretraining data and 250K parameter updates as the official setting. Apart from the additional CORE supervision, all other pretraining settings follow the original baseline. Due to time constraints, we report only pretraining-stage results, without subsequent supervised fine-tuning or preference optimization.

The table below reports the GenEval evaluation results:

MethodGenEval ↑
MiniT2I-B/3250.16%
MiniT2I-B/32 + CORE53.11%

CORE improves GenEval from 50.16% to 53.11%. These results indicate that CORE's effectiveness is not limited to ImageNet class-conditioned generation and can extend to text-conditioned pixel-space generation.

3. Relationship between supervision and spatial correspondence

Thank you for this insightful suggestion. In response, we conducted an experiment analyzing the sensitivity of supervision to spatial information. Holding the model weights, queried region, and each method's own supervision targets fixed, we exchange nearby and distant image patches within the context accessible to that method. We then measure the resulting changes in gradient directions at intermediate features and shared backbone parameters.

MethodFeature-gradient changeBackbone-parameter gradient change
JiT0.26370.5396
Self-Flow0.27890.4541
SRA0.3080.4797
SRA+DTS+AS0.06380.2556
CORE0.5650.8508

Under this diagnostic setting, CORE exhibits stronger responses at both levels. This indicates that CORE's supervision is more sensitive to spatial information, with this sensitivity reflected not only at the feature level but also in the gradient signals propagated to the generative backbone.

This analysis more directly reflects our motivation: pixel-space generation requires supervision that uses spatial context and connects to pixel prediction, rather than merely encouraging high feature similarity. We will replace Fig. 1(c) with this analysis in the revised manuscript and update the corresponding text. We appreciate your suggestion, which has helped us articulate this point more precisely.

AuthorsOpenReview ↗

Response to Reviewer dVGV (Comments 4–5)

4. Training computation

To compare CORE's training cost with other self-supervision methods, we measured steady-state training-step latency. CORE's contextual prediction and pixel correction tasks do increase per-step computation. However, all losses are optimized jointly, rather than updating parameters separately for each branch. Therefore, the number of optimizer updates remains unchanged at the same number of epochs and effective batch size. Furthermore, the auxiliary components are used only during training, so the inference architecture, sampling procedure, and computational cost remain unchanged. The comparison is provided below:

MethodStep time ↓ (ms)Throughput ↑ (images/s)
JiT93.1510,992
SRA113.569,018
Self-Flow120.138,524
DTS + AS131.747,773
SRA + DTS + AS136.627,495
CORE165.166,200

Compared with SRA+DTS+AS, CORE adds 28.54 ms per step, approximately 20.9%. This overhead comes from the additional spatial-supervision computation.

To further examine the relationship between training cost and generation quality, we use the measured per-step time ratio above as a reference and select 338 epochs as an additional evaluation checkpoint based on 600 × 93.15 / 165.16 ≈ 338. We report the CORE-L/16 results at this checkpoint below:

MethodEpochsReference relative training costFID ↓
JiT-L/166001.00×2.36
CORE-L/16338≈1.00×2.20
CORE-L/166001.77×2.02

At 338 epochs, CORE-L/16 already outperforms JiT-L/16 trained for 600 epochs, demonstrating better generation quality with fewer training epochs and parameter updates. After 600 epochs, its FID further improves to 2.02, surpassing the 2.11 reported by PixelREPA-L/16 with external representation supervision.

MethodExternal representation supervisionFID ↓
PixelREPA-L/16Yes2.11
CORE-L/16No2.02

At a larger scale, CORE-H/16 achieves 1.77 FID with approximately 953M generator parameters, outperforming the 1.82 FID reported by JiT-G/16 with approximately 2B parameters. These results further support CORE's effectiveness as model size increases and its value in improving generation quality without adding inference-time components.

MethodGenerator parametersFID ↓
JiT-G/16≈2B1.82
CORE-H/16≈953M1.77

5. Spatial retrieval

Thank you for pointing out this issue. The original Fig. 1(c) was intended to examine the relationship between feature similarity and spatial correspondence. Following your suggestion, we evaluated CORE under the same spatial retrieval protocol. Under the context-only condition, retrieval accuracy increases from 5.54% for JiT to 6.09% for CORE, while feature similarity decreases from 0.916 to 0.905. Under the no-image condition, JiT and CORE exhibit nearly identical retrieval accuracy. These results further illustrate that higher feature similarity does not necessarily imply better spatial correspondence.

Upon further reflection, we recognize that this experiment measures spatial retrievability from frozen representations through a linear readout. It provides a representation-level diagnostic, but does not directly measure how training supervision responds to spatial information, nor can it serve as the sole criterion for whether supervision benefits generation.

We particularly appreciate your suggestion in the third comment (Critical Weaknesses 3) to conduct targeted controlled experiments. This motivated our analysis of the spatial sensitivity of supervision, which more directly reflects our central point: our goal is not to obtain higher feature similarity, but to make training supervision respond effectively to spatial context and translate that response into learning signals relevant to pixel prediction. Accordingly, we will use the supervision-sensitivity analysis as the new Fig. 1(c) in the revised manuscript, and refine the corresponding text to better align the motivation, diagnostics, and methodological contributions.

For convenience, we reproduce the supervision-sensitivity results below:

MethodFeature-gradient changeBackbone-parameter gradient change
JiT0.26370.5396
Self-Flow0.27890.4541
SRA0.30800.4797
SRA+DTS+AS0.06380.2556
CORE0.56500.8508
AuthorsOpenReview ↗

Updated Manuscript and Responses to Your Concerns

Dear Reviewer,

Thank you again for your time. We have submitted a revised manuscript incorporating the additional experiments and clarifications discussed in our responses to your concerns.

We would appreciate hearing whether these updates address your concerns.

Best regards,

The Authors

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

Related papers

Discussion (0)

Sign in to comment.