acceptodds
Under review as a conference paper at ICLR 2027

WindowFold: Folding Sliding-Window Inference into a Single Forward Pass for Dense CLIP

Abstract

Every training-free method that turns CLIP into a dense predictor is evaluated with sliding-window inference: the image is cut into overlapping crops at the resolution CLIP was pretrained on, each crop is encoded separately, and the predictions are averaged, so the encoder processes the image area four times over. We show the window is a workaround for a train–test mismatch, and that it is doing two separable things. First, attention trained to spread its mass over 196 keys misallocates it over more: a plain high-resolution forward collapses while windows over the same pixels hold up. Second, a 224 crop silently caps how far a surgery may mix features, a cap worth 8.9 mIoU to a single pass on Cityscapes that no reweighting of attention can express. We replace each with a learned correction fit to the window's own output, with no labels, text or external data: a rank-16 LoRA on the attention projections, and a Gaussian footprint on the last block's mixing whose two to six widths are fit by the same distillation. Across three surgeries and four benchmarks, one forward matches or exceeds sliding-window inference in all 12 settings under the published protocol at – fewer FLOPs. On SC-CLIP, the training-free state of the art, one forward through the full method is above the published numbers on four of five benchmarks and on their average (42.68 against 42.4) with a single configuration. The fitted footprints are themselves a finding: five surgeries trained independently place the attention footprint at 5–7 patches, half a window, fixed in patch units rather than scaling with the image; across four backbones it follows the pretraining crop, not a pixel extent, and on L/14 one forward beats the ensemble on all five benchmarks.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
open-vocabulary semantic segmentation, CLIP, dense prediction, sliding-window inference, self-distillation, LoRA, resolution mismatch, training-free segmentation, attention surgery

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_K5P7OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes WindowFold, which replaces sliding window inference in CLIP-based semantic segmentation with a single whole image pass. WindowFold learns small attention updates and spatial weights from a frozen sliding-window teacher using unlabeled images. Empirical results show lower inference cost with comparable accuracy in many settings

Critical strengths
  1. The paper examines how image resolution, object size, and spatial mixing affect the gap between whole-image and sliding-window inference. These experiments help explain why windows are useful and motivate the two proposed components.
  2. The evaluation covers multiple dense-CLIP methods, datasets, resolutions, and backbones, with ablations covering
Critical weaknesses
  1. My main concern is novelty. The central idea has already been presented in SPAR [1]. WindowFold introduces LoRA and learned Gaussian spatial weights, but the paper neither discusses nor compares with SPAR, making it difficult to assess what these additions contribute beyond the existing approach.
  2. I also find the presentation unclear and inconsistent in places. For example, “surgery” is used in the abstract without a clear explanation of what it refers to. The name “WindowFold” also refers to the LoRA-only variant in Table 4 but the full method with Gaussian spatial weights in Table 9. Although the captions explain this distinction, defining terms at first use and naming the variants consistently would make the method and results easier to follow. The author should consider a major polishing of this paper to make it more accessible.
  3. I’m not convinced that the paper has shown that LoRA and the Gaussian weights repair two separate failure modes. Table 1 shows that restricting spatial mixing helps the frozen model, but it does not test the claim that LoRA cannot learn such a correction. The paper then uses the roughly additive gains in Table 8 as evidence for the two mechanisms. Two partial corrections to the same problem could also work better when combined. It would help to measure how LoRA alone changes the spatial range of attention before assigning each component a separate role.
  4. The claim that WindowFold beats sliding-window inference in all 12 settings under the published protocol needs to be qualified. Table 4 reports LoRA alone and uses a short side of 336 pixels for every dataset, including Cityscapes. However, the published Cityscapes setting is 560 pixels, as the paper itself notes in Appendix A. At that resolution, Table 9 shows that even the full method remains below sliding-window inference for SCLIP and ClearCLIP. The abstract and conclusion should make clear which model and evaluation settings the 12/12 result refers to.
  5. Section 5.6 overstates the average improvement from adding the spatial weights. Table 9 shows an average gain of 0.86 mIoU for ProxyCLIP and 0.73 for SC-CLIP, whereas the text claims gains of 4.3 and 4.4. Those larger numbers appear to refer to Cityscapes alone. This should be corrected because it gives a misleading impression of how much the spatial weights contribute across datasets. 0.86 and 0.73 improvements might just be noise, since no standard deviation reported.

[1] SPAR: Single-Pass Any-Resolution ViT for Open-vocabulary Segmentation. CVPR 2026.

Questions for authors

Refer to weaknesses.

3 comments
AuthorsOpenReview ↗

Response to Reviewer K5P7

W1. Novelty and relation to SPAR (CVPR 2026). Thank you for the pointer. SPAR is now cited and discussed in Section 2. Both methods distill a sliding-window teacher into one pass. They differ in what is adapted:

SPAR (default)WindowFold
Trainable parameterslast 2 ViT blocks, full (14.78M on B/16)rank-16 LoRA on all blocks (1.78M) + 2–6 scalars
Attention at test timeunconstrained globalGaussian footprint with a resolution schedule
Analysis—the window does two separable things (Sec. 3), each isolated by a control

We trained a SPAR-style control on ClearCLIP with our teacher, corpus and 12k steps (Table 19):

ClearCLIP B/16paramsVOC20PC59StuffADECity@560avgADE@672City@672
SPAR-style (last 2 blocks)14.78M85.3238.2826.2517.6927.3238.9716.5827.45
LoRA only1.78M83.5837.4225.7817.9230.9539.1317.6932.05
WindowFold1.78M84.3137.2825.5918.0132.1439.4718.0833.82

The SPAR-style student is better on the three object-centric benchmarks at 336, and we say so in the Limitations. As the token count grows it falls behind: −4.8 on Cityscapes@560 and −6.4 at 672, and below the LoRA alone on ADE20K@672. Fine-tuning the last blocks under global attention does not bound how far attention spreads on a larger grid. That is the failure the footprint targets.

W2. Terminology and presentation. "Surgery" and "WindowFold" are now defined once at the end of the first paragraph of the Introduction. Ablations always say "LoRA only" or "footprint only". Following your request for a major polish, Sections 3 and 4 are rewritten in full (C10, in blue).

W3. Evidence that the components repair separable failures. We ran the measurement you suggested: how far ClearCLIP's last-block attention reaches (RMS distance in patches, 200 ADE20K images) for each variant, all trained with the same teacher, 12k steps and multi-resolution grid (new Table 20 in App. I).

RMS attention distance14×1428×2842×42
frozen6.7814.1021.49
LoRA only6.139.6712.41
footprint only5.8810.9615.17
LoRA + footprint5.707.487.77
teacher (224 window)6.86——

You were right to ask: the LoRA alone does contract the range, by 42% at 42×42. It does not hold it fixed, though. Its range still grows from 9.7 to 12.4 patches as the grid grows, while the teacher's range is fixed by the crop. With the footprint the range stays at 7.5–7.8 patches, close to the teacher. We have therefore softened the claim in the Introduction to "a LoRA alone narrows the range but does not hold it fixed". The two components overlap partially, and the footprint supplies the part that does not depend on resolution. This reading is consistent with three other results:

(a) Near-additivity (Table 8): on SC-CLIP, Cityscapes@672, footprint alone +7.9, LoRA alone +12.8, both +18.1. (b) Selectivity: on ClearCLIP, whose single site already sits near σ ≈ 7, the footprint is close to inert; on SC-CLIP, which mixes globally, it closes most of the Cityscapes gap. (c) Budget-matched control (new Table 18): footprint only cannot reach the full method even at 3× the steps, so the LoRA supplies something the footprint cannot.

W4. "12/12 settings". You are right that it was over-stated. It holds for the LoRA alone at short side 336 (Table 4); at 448 it is 10/12, with Cityscapes as the exceptions. The abstract, introduction, Section 5.1, conclusion and Table 4 caption now say this.

W5. Section 5.6 gains. This was a drafting error, thank you. The corrected text reads:

Footprint gain5-benchmark avgCityscapes@560
ProxyCLIP+0.86 (39.83 → 40.69)+4.43 (30.86 → 35.29)
SC-CLIP+0.73 (41.95 → 42.68)+4.32 (34.23 → 38.55)

On variance: across three σ initializations and two seeds, the SC-CLIP five-benchmark average stays within 0.07 (Table 13). This is a small number of runs, so we read it as indicative. As expected for a range correction, the gain is concentrated on the high-resolution, wide-scene benchmark, and Section 5.6 now says so.

We would be glad to answer further questions before the discussion closes.

AuthorsOpenReview ↗

Follow-up (1/2): new evidence on W5 (variance) and W1 (SPAR + footprint)

W3 and W5 led to the two most informative new experiments in the revision. Since our first reply we have completed the two experiments your review asked for most directly. Both are on full validation sets under the paper's protocol, and both are now in the revised PDF (App. I, Tables 21 and 22, in blue).

W5. Is the footprint gain noise? It is not, and we now report where it comes from.

We retrained the LoRA-only and LoRA + footprint models with 3 seeds each for ProxyCLIP and SC-CLIP. For every seed we stored per-image TP/FP/FN, which lets us run a paired image-level bootstrap (B = 2000, same resampled images for both arms).

mean ± std, 3 seedsVOC20PC59StuffADECity@5605-bench avg
ProxyCLIP, LoRA84.65±0.0638.14±0.0226.62±0.0318.85±0.0730.99±0.2239.85±0.03
ProxyCLIP, + footprint84.56±0.2538.05±0.0626.40±0.0618.79±0.0435.30±0.1340.62±0.08
Δ−0.09−0.10−0.21−0.06+4.31±0.33+0.77±0.11
SC-CLIP, LoRA86.20±0.1741.14±0.0727.84±0.0619.99±0.0234.43±0.1941.92±0.04
SC-CLIP, + footprint85.66±0.1041.01±0.0727.80±0.0720.12±0.0238.52±0.1242.62±0.06
Δ−0.53−0.13−0.04+0.13+4.09±0.28+0.70±0.07
  • The five-benchmark gain is +0.77 [+0.59, +0.95] for ProxyCLIP and +0.70 [+0.56, +0.84] for SC-CLIP (paired bootstrap 95% CI, pooled over seeds, p(Δ ≤ 0) < 0.001 for both). It is 7–10× the seed-to-seed standard deviation. The single-seed values in the paper (+0.86 / +0.73) sit slightly above the 3-seed means.
  • The gain comes from Cityscapes@560. For every one of the 6 model pairs, its 95% CI lies inside [+3.2, +5.3].
  • On the four 336-pixel benchmarks the footprint is neutral to slightly negative: −0.12 [−0.24, +0.01] for ProxyCLIP and −0.14 [−0.26, −0.04] for SC-CLIP. Two cells show a consistent cost: SC-CLIP VOC20 (−0.5) and ProxyCLIP COCO-Stuff (−0.2).

This answers your W5 point more sharply than our first reply did. An average over five benchmarks is the wrong summary of the footprint. It is a range correction. It does almost nothing when the test grid is close to the teacher's window, and it gives about +4 mIoU on the wide, high-resolution benchmark where the range problem appears. This matches the attention-range measurement in Table 20. Section 5.6 now gives the 3-seed means and intervals next to the single-seed numbers, and states that the gain is concentrated on Cityscapes with a small cost at 336.

W1. Is the footprint just SPAR plus a trick? We added it on top of a SPAR-style student.

Our first reply compared the SPAR-style student (last 2 blocks fine-tuned, 14.78M parameters) with our method. A fair objection is that the footprint might help only weak students. We therefore trained SPAR-style + footprint on ClearCLIP with the same teacher, corpus and 12k steps (Table 21):

ClearCLIP B/16VOC20PC59StuffADE@336ADE@448ADE@672City@560City@672
SPAR-style (last 2 blocks)85.3238.2826.2517.6917.8716.5827.3227.45
SPAR-style + footprint85.4838.1426.0717.6117.8916.6928.3029.10
Δ+0.16−0.14−0.18−0.08+0.02+0.11+0.98+1.65

The footprint still adds accuracy on top of SPAR-style fine-tuning. The gain grows with the token count (+0.98 at 560, +1.65 at 672) and is near zero at 336. So the footprint is complementary to SPAR. It can be added to SPAR's student at the cost of 2–6 scalars.

The gain is smaller than on the LoRA student (+4), which suggests the fine-tuned blocks already correct part of the range; that it still grows with the grid suggests they do not bound it. We did not measure the attention range of the SPAR-style student directly, so this reading is indirect. This run uses a single seed, and the SPAR-style baseline was evaluated before we added per-image dumps, so this table has no confidence interval. We say so in the appendix.

AuthorsOpenReview ↗

Follow-up (2/2): contribution beyond SPAR, W4, and summary of changes

What the paper now contributes beyond SPAR. In response to your Informativeness score, we state the takeaway plainly. Sliding windows do two separate things for dense CLIP: they correct the features, and they cap how far attention reaches. A single-pass student has to reproduce both. Our evidence that these are two different things:

  • A learned attention correction only narrows the range. Table 20 shows the LoRA's range still grows from 9.7 to 12.4 patches with grid size.
  • An explicit footprint holds the range fixed at 7.5–7.8 patches, close to the teacher's 6.9.
  • The footprint adds accuracy on top of either student (LoRA: +4.1 to +4.3 on Cityscapes; SPAR-style: +1.0 to +1.7), and adds nothing where the range is already right (336).

We think this explains why single-pass students lose accuracy on large, wide inputs. It also suggests a cheap addition to SPAR-style students.

W4, for completeness. The qualified claim is now "12/12 for the LoRA alone at short side 336; 10/12 at 448". At the published Cityscapes@560 setting, the full method still trails the window ensemble for SCLIP (32.65 vs 32.92), ClearCLIP (32.14 vs 32.68), ProxyCLIP (35.29 vs 36.27) and SC-CLIP (38.55 vs 40.96). It is above the ensemble for NACLIP (36.04 vs 35.57). It does this at 1.5–2.4× fewer FLOPs. The abstract, introduction, Sec. 5.1, conclusion and the Limitations section say this explicitly.

Summary of changes for your points.

Your pointWhereStatus
W1 SPARSec. 2; App. I, Table 19 (SPAR-style); new Table 21 (SPAR-style + footprint)discussed, compared, shown complementary
W2 terminology / polishSec. 1 definitions; Sec. 3–4 rewritten (blue)done
W3 LoRA vs. footprint rolesApp. I, Table 20; claims softened in Abstract, Sec. 1, 3.5, 4.3measured as you suggested
W4 12/12Abstract, Sec. 1, 5.1, Conclusion, Table 4 captionqualified
W5 averages / noiseSec. 5.6; new Table 22 (3 seeds + bootstrap CIs)corrected, significance shown

Each of your five points led to a change in the paper. Some of these changes cut against our original claims: the 12/12 result is now qualified, and the footprint is shown to be neutral or slightly negative at 336. If the revision and the new evidence address your concerns, we would be grateful if you would consider updating your assessment. We are happy to run any further check before the discussion closes.

Reviewer_2j8kOpenReview ↗
Rating
3
Confidence
4(+1)
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

WindowFold replaces sliding-window inference in dense CLIP with one forward pass. It learns small LoRA updates and a local mixing range from the frozen model’s window outputs, without labels. Tests cover several CLIP methods, backbones, and segmentation datasets.

Critical strengths

The paper provides useful controls to study why sliding windows help. It separately tests attention adaptation, position embeddings, and local feature mixing.

The method offers a clear speed–accuracy benefit. One pass often matches or improves window-based results, with lower computation. Tests across backbones and training settings support its usefulness.

Critical weaknesses

The data split needs clarification. Some adapters use COCO 2017 validation images for training, which may overlap with COCO-Stuff evaluation images. This affects whether those results measure performance on unseen images.

The claim that the learned mixing range is a property of the encoder is stronger than the evidence. Since the teacher uses windows, the learned range may also reflect the teacher’s window size.

Questions for authors

Do the distillation images overlap with the COCO-Stuff evaluation images? Please report the image-ID overlap and, if needed, repeat the affected tests using separate training images.

With the backbone fixed, does changing the teacher’s window size change the learned mixing range?

2 comments
AuthorsOpenReview ↗

Response to Reviewer 2j8k

Q1. Image-ID overlap with COCO-Stuff. Confirmed and disclosed in Section 4.4. With disjoint ADE20K training images, the LoRA adapters lose 0.44–0.79 mIoU on COCO-Stuff and still beat the ensemble in all six cells (Table 17). For the footprint-bearing models the difference is at most 0.23.

Q2. Is the mixing range a property of the encoder or of the teacher's window? Thank you; this question changed a claim in the paper. In all our backbones the teacher window equalled the pretraining crop, so the two were confounded. We separated them by distilling ClearCLIP B/16 (pretraining crop 224) from a 336 window at stride 168:

ClearCLIP B/16σ @28×28σ @28×56City@560City@672ADE@448
teacher window 224 (paper)7.06.832.1834.2118.15
teacher window 33613.711.827.5030.4418.12

The footprint follows the teacher's window. We revised Section 5.5, the abstract and contribution (iii) accordingly: the width is shared across surgeries and set to about half the teacher's window, not an intrinsic encoder constant. The 336 teacher also gives a worse student at high resolution, so the 224 window remains the right teacher for a 224 encoder.

Why does σ grow 1.95× when the window grows 1.5×? σ is the width of a logit bias, not the attention radius. On a finite 28×28 lattice the realized radius saturates in σ (RMS attention distance 7.1 / 11.3 / 14.1 patches at σ = 7 / 14 / ∞), so a linear change in radius needs a superlinear change in σ. Measured directly, the realized RMS attention distance of the two trained students grows from 7.48 to 11.48 patches (1.53×), matching the teacher's own window attention (6.86 → 10.44, 1.52×) and the window ratio.

Is it caused by the 336 window extrapolating the 224 position embedding? We checked two ways. (a) The 336 teacher is not more diffuse: its attention spread relative to the window side is 0.497 vs. 0.490 at 224. (b) L/14 → L/14@336, where the 336 window is native and no extrapolation happens, shows the same pattern: σ grows 1.80× (8.36 → 15.05) and the implied radius 1.51× (9.04 → 13.66). All of this is in Appendix I.

Reviewer_2j8kOpenReview ↗

Final comment

Thank you for the response. I would like to raise soundness and confidence, and maintain grade.

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

The paper introduces WindowFold, a lightweight method for replacing sliding window inference in dense CLIP predictions with a single forward pass on the full image. WindowFold proposed a low rank correction to attention and a learned Gaussian footprint. Experiments evaluate several CLIP feature surgery methods across segmentation benchmarks and visual backbones. The experimental results show that the adapted single pass matches or exceeds sliding-window accuracy in many evaluated settings, with reductions in computational cost, although performance gaps remain in some Cityscapes configurations.

Critical strengths

1.The paper addresses a practically relevant problem and demonstrates improvements across several existing dense CLIP methods. In Table 4, the adapted single pass outperforms the corresponding sliding-window baseline in all 12 dataset–method combinations at a short-side resolution of 336.

  1. The proposed method only trained a single adapter and reused it across mutiple datasets and resolutions, reducing the need for dataset-specific retraining. These results suggest that offline adaptation can offer a useful accuracy–efficiency trade-off for repeated inference.

  2. The experiments on the spatial extent of feature mixing are informative. In Table 1, restricting feature mixing to a Chebyshev radius of 4 patches in frozen SC-CLIP improves Cityscapes mIoU from 21.05 to 29.90 at a short-side resolution of 672. The SC-CLIP ablations in Table 8 provide further evidence of complementary benefits from attention adaptation and the learned footprint, while Table 13 examines the stability of the learned spatial parameters across different initializations and random seeds.

Critical weaknesses
  1. This paper has a strong AI flavor, with excessive involvement of AI writing. There are numerous short paragraphs, a preference for using semicolons and colons, and a tendency to provide highly detailed evidence in parentheses when presenting experimental analysis. And it's hard for people to read and understand the main methods and experiments due to AI-flavor writing style.

  2. The proposed method is mainly compared to the training-free CLIP-style methods, while the WindowFold itself is not training-free and requring distllation-based training. Even it doesn't require data labels, it still require training a LoRA module. So the comparison against the training-free CLIP-style methods (sliding-window-based ) seems unfair.

  3. In Section 4.4, the authors said "The LoRA-only adapters of MaskCLIP,SCLIP, ClearCLIP, NACLIP and ProxyCLIP are distilled on the 5000 images of the COCO 2017". However, the evaluation of COCO-Stuff uses the same images. Although the labels may not be used, the data leakage problem still exist. Could you evaluate the proposed adapters trained on a disjoint image set?

Questions for authors

Are the component ablations using the same training corpus and optimization budget, with positional grid learning held constant? After reading the paper, It's still fuzzy to identify the contributions of different components.

2 comments
AuthorsOpenReview ↗

Response to Reviewer WbXU(W1-4)

W1. Writing style. We agree. Sections 3 (Diagnosis) and 4 (Method), which carry the main argument and the method, are rewritten in the revision (all in blue): short paragraphs merged, parenthetical and semicolon chains replaced by full sentences, and every number left unchanged. We also define each term once (C7) and replaced over-compressed claims with qualified statements (C5, C6). Sections 5–6 and the table captions will receive the same pass for the camera-ready.

W2. Fairness of comparing a trained adapter with training-free baselines. WindowFold is not training-free and we do not claim it is. Its target is the sliding-window ensemble, which costs 6–55 forward passes per image. Our claim is about test-time cost: one pass matches or beats the ensemble at 1.5–2.4× fewer FLOPs, after a one-time, label-free distillation of 12k steps. To separate "any training helps" from our design, the revision adds two trained controls under the same teacher, corpus and budget: a budget-matched footprint-only model (Table 18) and a SPAR-style last-two-blocks fine-tune (Table 19). The Limitations section now states this scope explicitly (C8).

W3. COCO val2017 vs. COCO-Stuff overlap. You are right; the image sets coincide. No labels are used, but the test images were seen. Section 4.4 now says so. Retraining the LoRA adapters on disjoint ADE20K images lowers COCO-Stuff by 0.44–0.79 mIoU, and all six disjoint cells still beat the ensemble (Table 17, reproduced in the General Response).

W4. Budget of the footprint-only ablation. Correct: it was trained for 4k steps without the grid. We reran it with the full budget (Table 18):

VOC20PC59ADECity@560avg (4)
ClearCLIP, footprint only, 4k84.9535.7816.9122.0239.92
ClearCLIP, footprint only, 12k + grid85.0536.0316.9121.6539.91
ClearCLIP, WindowFold83.7937.1817.7532.1842.73
SC-CLIP, footprint only, 4k86.5140.4819.9829.8444.20
SC-CLIP, footprint only, 12k + grid86.4340.3319.8330.2844.22
SC-CLIP, WindowFold85.3040.9920.0338.9146.31

Tripling the steps and adding the grid moves the average by ≤0.02. The 2–6 footprint scalars converge early, and the remaining gap is the LoRA's contribution.

AuthorsOpenReview ↗

Response to Reviewer WbXU(Q)

Q (same corpus, budget and grid across ablations?). After this rerun, every trained row of the component ablations (Table 8, Table 18, Table 20) uses 12k steps, the same teacher, the multi-resolution crop schedule and the learnable position grid, so the grid is held constant. One difference remains, and we state it here: the LoRA-only rows were distilled on COCO images and the footprint-bearing rows on ADE20K images (Sec. 4.4). The corpus matters little. A ClearCLIP LoRA distilled on ADE20K images scores 17.84 on ADE20K@336, against 17.92 for the COCO one. The new range measurement (reply to K5P7 W3, Table 20) separates the two components directly, without relying on mIoU differences.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response to All Reviewers

We thank the reviewers for their careful reading. We have uploaded a revised PDF. All changed text is in blue, and every new experiment is collected in a new Appendix I (pp. 19–20) so that it can be read on its own. All new numbers are on full validation sets under the protocol of the paper.

Change map

#Concern (reviewer)Where in the revised PDFWhat changed
C1COCO val2017 images = COCO-Stuff val images (WbXU, 2j8k)Sec. 4.4; Table 4 caption; App. I, Table 17Overlap disclosed. LoRA adapters retrained on disjoint ADE20K images: COCO-Stuff drops 0.44–0.79 and still beats the window ensemble in all 6 cells.
C2Relation to SPAR, CVPR 2026 (K5P7)Sec. 2 (new paragraph); Limitations; App. I, Table 19; refsSPAR cited and discussed. New SPAR-style control trained under our teacher, corpus and budget.
C3Footprint-only ablation not budget-matched (WbXU)App. I, Table 18Footprint-only rerun at 12k steps + multi-resolution grid. Result unchanged (ClearCLIP 39.91 vs 39.92; SC-CLIP 44.22 vs 44.20).
C4Mixing range: encoder property or teacher window? (2j8k)Sec. 5.5 (claim revised); Sec. 5.6 "Other backbones"; Abstract; Intro contribution (iii); App. I, last paragraphB/16 distilled from a 336 window: σ follows the teacher's window. The 1.95× σ growth for a 1.5× window is explained and measured (realized radius 1.53×). The claim "a property of the encoder" is removed.
C5"12/12 settings" over-claimed (K5P7)Abstract; Intro contribution (ii); Sec. 5.1; Conclusion; Table 4 captionQualified everywhere: LoRA alone, short side 336, 10/12 at 448, Cityscapes@560 in Table 9.
C6Sec. 5.6 footprint gains misreported (K5P7)Sec. 5.6, first paragraphCorrected from "+4.3/+4.4 average" to +0.86/+0.73 average (+4.43/+4.32 on Cityscapes).
C7Terminology: "surgery", "WindowFold" (WbXU, K5P7)Sec. 1, end of first paragraphBoth defined once; ablations always say "LoRA only" or "footprint only".
C8Scope of the contribution (WbXU, K5P7)LimitationsStates that the method is bounded by its teacher, that gains over the ensemble are 0.3–2 mIoU in most cells, and that a SPAR-style fine-tune is stronger on object-centric benchmarks at low resolution.
C9Does LoRA alone change the attention range? (K5P7 W3, WbXU Q)App. I, Table 20; Abstract; Sec. 1; Sec. 3.5; Sec. 4.3New measurement of last-block RMS attention distance per variant. LoRA alone narrows the range but it still grows with the grid; only the footprint holds it fixed. "LoRA cannot express it" softened to "only partly recovers".
C10Writing style (WbXU W1, K5P7 W2)Sec. 3 and Sec. 4, entire textRewritten in full sentences; numbers unchanged.
C11Footprint on a SPAR-style student (K5P7 W1)App. I, Table 21+0.98 / +1.65 on Cityscapes@560/672; complementary to SPAR.
C12Seed variance of the footprint gain (K5P7 W5)Sec. 5.6; App. I, Table 223 seeds + paired bootstrap; avg gain +0.77 / +0.70, CI excludes 0; gain concentrated on Cityscapes, neutral to slightly negative at 336.
AuthorsOpenReview ↗

Summary of the new evidence

  1. Leakage does not drive the gains (C1). COCO-Stuff, LoRA only, distilled on COCO vs. on disjoint ADE20K images, against the frozen window ensemble:
slide @336COCOADE (disjoint)slide @448COCOADE (disjoint)
MaskCLIP19.3020.5819.7918.6220.5419.92
SCLIP23.1524.0823.6423.3024.4623.98
ClearCLIP24.5925.7825.1024.6726.0925.62

The footprint-bearing models of Table 9 were already distilled on ADE20K. Their COCO-trained counterparts differ on COCO-Stuff by at most 0.23 mIoU.

  1. The gain over footprint-only comes from the LoRA, not from training budget (C3). See Table 18.

  2. SPAR-style fine-tuning (C2) is stronger at low resolution on object-centric data and loses 4.8–6.4 mIoU on Cityscapes at 560–672. See the reply to K5P7.

  3. The footprint follows the teacher's window (C4). See the reply to 2j8k.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.