acceptodds
Under review as a conference paper at ICLR 2027

How Far Does a Shared Linear Map Go? Probing Feature-Space Manipulability for Image Editing

Abstract

Understanding how image-space transformations manifest in a model's internal feature representations is a longstanding goal in representation analysis. Prior work on model stitching and equivariance has shown that certain geometric transformations — rotations, flips — can be captured by learned linear operators between feature maps. It remains unclear whether this holds for a broader and more practically relevant class of transformations, including photometric edits and open-ended, semantically defined manipulations with no a priori linear structure in input space, such as those produced by prompted diffusion-based image editors. We train probes with increasing capacity, ranging from a single linear map shared across all spatial locations in a feature map to nonlinear per-vector, receptive-field, and global transformer models, to predict the feature-space effects of diverse image manipulations: geometric transforms, photometric edits, local occlusions, and semantic edits (e.g., altering headlights, rim color, body color) generated via diffusion-based editing. We evaluate three vision backbones: two supervised architectures, ConvNeXt and SwinV2, and one self-supervised foundation model, DINOv3. For ConvNeXt and SwinV2, a single shared linear map, without spatial or instance-dependent conditioning, often predicts held-out manipulation outcomes with little loss relative to substantially more expressive nonlinear models. This pattern is less consistently observed for DINOv3. For the supervised backbones, the sufficiency of the shared linear map generally increases with network depth. We further show this holds despite the spatial weight-tying constraint, which is not implied by prior stitching formulations and represents a nontrivial locality claim independent of linearity. We deliberately scope our claims to representational sufficiency for held-out prediction, rather than to intrinsic properties of feature-space geometry. Our results indicate that a remarkably simple, shared linear operator is often sufficient to represent a broad class of image manipulations, with the semantic content of the edits captured by its leading singular components and higher-rank components primarily refining image details

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Deep feature representations, Feature-space mapping, Representation analysis, Image manipulation

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

This paper investigates whether image manipulations can be approximated by spatially shared linear mappings in the feature spaces of pretrained vision models. It evaluates geometric, photometric, masking, and semantic transformations across ConvNeXt, SwinV2, and DINOv3, comparing linear mappings with MLPs, CNNs, and Transformers. The results suggest that simple linear mappings can often approximate complex semantic edits, particularly in deeper feature representations. An additional SVD analysis suggests that leading singular components capture much of the visible editing effect.

Critical strengths

An interesting research question. The paper explores whether complex semantic transformations can be approximated by simple linear operators in pretrained visual representations. This question is relevant to representation learning, interpretability, and feature-space editing.

Critical weaknesses
  1. Potentially unfair comparison between linear and nonlinear mappings. The MLP and Transformer use a fixed hidden dimension of 512, while deeper features have 1024 channels. Without appropriate residual connections, this introduces dimensional bottlenecks that may limit their expressiveness relative to the linear baseline. Consequently, the small performance gap may reflect architectural constraints rather than the sufficiency of linear mappings, potentially also confounding the claimed improvement with feature depth.
  2. Insufficient distinction between feature preservation and actual editing. High similarity between predicted and target features does not necessarily indicate successful editing, particularly when most features remain unchanged. Appendix A.10 reports substantially weaker displacement alignment for localized edits and a higher displacement prediction error for Linear than Transformer (RMSE: 5.99 vs. 4.42). These findings raise questions about whether linear mappings accurately capture editing-induced changes rather than primarily preserving the original representations.
  3. Questionable interpretation of SVD components. The SVD analysis decomposes the complete mapping WW, whereas the actual editing displacement is determined by (W−I)F+b(W-I)F+b. Therefore, the apparent concentration of semantic edits in leading singular components may simply result from the loss of original image details. Analyzing W−IW-I would provide more convincing evidence.
  4. Potential overestimation of semantic editing success. The VLM evaluation adopts permissive criteria, such as accepting faded mirrors as successful removal or indistinct objects as teddy bears. It also allows successful judgments when the desired property already exists in the original image. Without no-edit controls, these criteria may conflate genuine semantic manipulation with reconstruction artifacts or unchanged attributes.
Questions for authors
  1. Fairness of model comparisons. Can the paper clarify whether the nonlinear probes include residual connections and evaluate wider nonlinear architectures to ensure that their performance is not limited by dimensional bottlenecks?
  2. Actual editing effectiveness. How much do the learned mappings improve over an identity mapping? Can the paper report editing-displacement errors and explain the relatively poor directional alignment observed for localized manipulations?
  3. Semantic evaluation reliability. What is the false-positive rate of the VLM evaluator on original images and FeatInv reconstructions without editing? Can the paper provide stricter semantic success rates, ideally validated through human evaluation?
1 comment

Rebuttal to Reviewer Afwe

We thank the reviewer for the careful reading. We have uploaded an intermediate revision of the paper in which all changes are marked in blue. We will upload the final revised version, incorporating all reviewer feedback, before the end of the discussion period.

The reviewer identified important limitations in the original comparison. We now distinguish edit displacement from preservation and report results relative to an identity/no-edit mapping. In feature space, RΔ2=1−∥Fmap−Ftgt∥2/∥Ftgt−Forig∥2R^2_\Delta = 1 - \lVert F_{map}-F_{tgt}\rVert^2 / \lVert F_{tgt}-F_{orig}\rVert^2; RΔ2=0R^2_\Delta=0 is the no-edit reference, 11 is exact displacement recovery, and a negative value is worse than no-edit. Masked RΔ2R^2_\Delta restricts both errors to the edited region. Feature-space no-edit masked cosine is a separate quantity, not zero by definition.

W1/Q1 (probe capacity and residuals)

The reviewer is right that a 512-wide MLP or Transformer without a skip connection creates an identity-path bottleneck when the feature dimension is 1024. We ran an ablation on five semantic edits at ConvNeXt and SwinV2 feat3, comparing the paper-width probes with 512-wide residual probes and full-width probes (MLP width 4d4d, Transformer dmodel=dd_{model}=d), with and without a zero-initialized residual connection. Values below are means of per-edit medians, reported as RΔ2R^2_\Delta / masked RΔ2R^2_\Delta:

ProbeConvNeXt feat3SwinV2 feat3
Linear0.11 / 0.300.20 / 0.28
MLP (paper)0.11 / 0.300.23 / 0.32
MLP, 512 + skip0.36 / 0.400.31 / 0.36
MLP, full width0.37 / 0.450.36 / 0.42
MLP, full width + skip0.42 / 0.460.36 / 0.41
Transformer (paper)0.00 / 0.220.23 / 0.34
Transformer, 512 + skip0.33 / 0.370.30 / 0.35
Transformer, full width0.18 / 0.330.29 / 0.38
Transformer, full width + skip0.36 / 0.390.33 / 0.38
The bottleneck materially limited the original nonlinear probes. On ConvNeXt feat3, the full-width residual MLP exceeds the paper MLP on 97–99% of samples per edit, with decoded LPIPS decreasing from 0.35–0.39 to 0.15–0.21. The nonlinear probes therefore outperform the linear map at these d=1024d=1024 layers; we will revise the paper’s claim that they perform similarly here. Linear still captures a substantial portion of the in-region displacement (masked RΔ2R^2_\Delta 0.30 vs. 0.46 on ConvNeXt and 0.28 vs. 0.42 on SwinV2, relative to the best probe). Importantly, all compared variants produced the requested manipulation, but neither width matching nor adding a skip connection consistently improved manipulation correctness; the more reliable gain was in overall reconstruction fidelity, as reflected by lower LPIPS.

W2/Q2 (edit capture vs. preservation)

Across the 297 dataset/backbone/layer/edit combinations, means of per-combination medians show positive masked displacement capture for semantic edits: masked RΔ2R^2_\Delta is 0.13 for Linear and 0.21 for the paper Transformer, with Linear positive in 80% of combinations. For masking, the corresponding values are 0.21 and 0.46; for geometric edits, 0.43 and 0.45 (with the spatial-only transform as zero reference). Globally, semantic RΔ2R^2_\Delta is −0.17 for Linear and −0.02 for the paper Transformer: small errors on many unchanged tokens can outweigh gains in the edited region. This explains weak global alignment for localized edits and motivates reporting masked alongside global results. In the Cars ConvNeXt feat3 residual-probe results above, global semantic RΔ2R^2_\Delta is positive for every tested edit (0.28–0.58), compared with −0.22–0.46 for the paper MLP.

W3 (WW versus W−IW-I)

For the reported 297 linear probes, the effective rank of W−IW-I for semantic edits is 0.75d0.75d, compared with 0.53d0.53d for WW; it is higher for every semantic probe. The Frobenius norm ratio ∥W−I∥F/∥I∥F\lVert W-I\rVert_F/\lVert I\rVert_F is 0.95 for semantic edits (masking 0.52, photometric 0.68). For masking edits, W−IW-I is more compact (0.54d0.54d versus 0.82d0.82d for WW). In the ConvNeXt feat3 truncation analysis, top-32 WW falls below the no-edit reference for three of four edits (RΔ2=−0.26R^2_\Delta=-0.26 to −0.84-0.84), whereas top-32 W−IW-I retains 46–75% of the full probe’s RΔ2R^2_\Delta; reaching 90% requires k=256k=256–10241024. We will show both decompositions and describe the semantic displacement as distributed across many directions, with a substantial portion in the leading ones.

W4/Q3 (VLM reliability and human validation)

We agree that a permissive judge can overestimate success, particularly when the requested property is already present or an object is indistinct. We will report per-edit rates under the original rubric and include no-edit controls to investigate the False Positive Rate. We will also report blinded human ratings on a subset rather than imply that the VLM alone establishes semantic success.

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

This paper studies probing geometric, photometric, and semantic manipulations like object addition, on existing architectures like ConvNeXt, SwinV2 and DINOv3. The core of this work relies on conditioning diffusion model with features from these backbones (using prior work FeatInv) and understanding the layers that contribute to perform geometric manipulation,

The paper then analyzes how different features and their corresponding layers contribute towards object addition, color change etc. They show object addition like adding lights to police cars, changing color of objects, rotating images, removing blocked patches, and more to measure geometric, local masking and semantic manipulation.

Critical strengths
  1. This work measures the importance of conditioning stable diffusion model on features of existing models like ConvNeXT, SwinV2, and DINO. The training and effect of three feature layers per model is thorough.
  2. Composable feature manipulation results in the appendix seem interesting where chaining multiple edits performs meaningful manipulation to obtain the desired result.
Critical weaknesses
  1. The paper needs rewriting and improving its motivation. It is motivated as a way to probe internal representations of existing models for geometric, local, and semantic manipulations but provides encoded features to FeatInv for generating geometric edits that relies on stable diffusion with ControlNet. The analysis shows ConvNeXT backbone feat layer 3 is helpful in performing some of the semantic and geometric tasks. It is unclear what this result is trying to show? It seems like this feature layer is the most important to preserve the content that can then be manipulated by a stable diffusion model but the reviewer is not sure how to extend this learning to the capabilities of the model. Furthermore, the reviewer’s claim is supported by the experiment that shows adding a teddy bear or police light is more difficult than changing colors of objects as the former requires adding features not present in the image whereas the latter relies on manipulating what exists in the scene.
  2. FeatInv that is the prior work this method depends on is not covered in a background section and is difficult to follow the paper without introducing it in method or just before it.
  3. Images are too small and are not visually appealing. Adding lights on cars is not clear or neat showing bad edit adherence in the main paper. This method seems to only work well in cases where perceptual manipulation is present or easy geometric manipulation that measure how well the features capture the input image.
  4. By providing features to a stable diffusion backbone to then generate edit images it is not clear if the results are the strength of the feature extractor or the generation capability of stable diffusion. This is also likely the reason why linear feature extractors match nonlinear extractors in reconstruction. Is one model trained per geometric/perceptual manipulation? Or is one model trained for all manipulations for a particular feature extractor and layer?

Post Rebuttal Comments

  1. The authors have resolved my queries about readability of the paper, added important background + pseudocode and explanation of their results. Hence I have updated the score to 3 (weak accept).
Questions for authors

I am inclined to weak reject this work. I request authors to answer the following during rebuttal:

  1. A clear motivation about this work and what we can learn from its results for future research. How can the community benefit from these observations? This is missing in the current work.
  2. A comment about disentangling the capabilities of stable diffusion vs importance of features. Is it possible to tease out if the diffusion model is biased towards features of one architecture? For example, using the feature of SD VAE/feature layers from it may provide high accuracy as it is already aligned with the SD model. But that doesn’t make it a rich feature extractor compared to other works.
  3. How is the editing done in inference? Is there an edit prompt or is controlnet alone handling the specific edit that needs to take place based on the training data it was trained on?
  4. It is unclear why a stochastic generative decoder (ControlNet + Stable Diffusion) is used to evaluate mapping precision, rather than a deterministic decoder (e.g., a U-Net trained with perceptual loss) to understand the importance of features. This would ideally avoid any bias from the diffusion model.
  5. The reviewer appreciates the rigor in experimentation but believes that such analysis would greatly benefit pretrained diffusion and flow matching models that are directly used for editing where layer-wise importance is still an active area of research for models such as FLUX, Qwen Image Edit, Wan video models and more!
7 comments

Rebuttal to Reviewer Zhgf:

We thank the reviewer for the careful reading. We have uploaded an intermediate revision of the paper in which all changes are marked in blue. We will upload the final revised version, which incorporates all reviewer feedback, before the end of the discussion period.

The edit is performed in feature space: for each dataset, manipulation, backbone, and layer, we train a mapping from original to edited image features. FeatInv is a fixed decoder used only for visualization; it receives no edit instruction or information about the manipulation, only the mapped features and the same generic prompt (“a high-quality, detailed, and professional image”). We study whether an edit-induced feature change can be predicted by a shared linear operator, not whether FeatInv or a diffusion model can edit images.

Motivation and takeaways (W1, Q1)

Prior work on stitching and equivariance (Lenc & Vedaldi; Bansal et al.) showed that geometric transformations act approximately linearly on CNN features. We ask whether this extends (G1) to photometric and semantic edits, which have no linear structure in pixel space; (G2) to a single operator shared across spatial positions; and (G3) to ViT and self-supervised backbones. Takeaways:

  • Representation analysis. For supervised backbones, a single spatially shared linear map recovers a substantial part of the edit-induced change even for semantic edits (masked RΔ2R^2_\Delta, deepest layer: 0.30 / 0.41 on ConvNeXt, 0.28 / 0.33 on SwinV2, Cars / LSUN), measured against the no-edit reference. Width-matched residual probes recover more (up to 0.43), so the linear map is strong but not complete. That a position-independent operator works at all is a nontrivial locality property not implied by stitching.
  • Feature-space control. Such operators are cheap, composable and inspectable: the visible edit already emerges from the top 32 SVD directions of W−IW-I.
  • Where linearity breaks. Insertions (teddy bear) benefit from spatial context: on SwinV2 and DINOv3 the Transformer clearly exceeds the linear map in masked feature similarity (0.44 → 0.77, 0.63 → 0.87). On DINOv3, the linear map does not beat the no-edit reference for semantic edits on Cars (RΔ2R^2_\Delta −0.58 to −1.50), whereas residual nonlinear probes do (0.09–0.28), though these edits change DINOv3 features comparatively little (no-edit masked cosine 0.89–0.96).
  • Methodology. Decoder-free feature-space metrics measure how much of the edit-induced displacement a mapping recovers relative to a no-edit reference, globally and inside the edit mask, separating edit capture from preservation. Decoding with FeatInv makes edits visible for semantic judgment (VLM, human ratings); every decoded result is compared with the decoded unedited features, so decoder priors are shared across conditions.
  • Relation to linear encoding of semantics. We do not test the linear representation hypothesis, which concerns static directions within one representation; we study the transformation between representations. If an attribute were encoded linearly, changing it would be an approximately position-independent affine shift, i.e. our operator class. Our results are consistent with this for supervised backbones and show where it fails (insertions, DINOv3): a complementary, operator-level step, not a test.

FeatInv and the role of the decoder (W2, W4, Q2–Q4)

We add a concise description of FeatInv to Methods. Within a fixed backbone and layer, the same FeatInv model decodes all mapped outputs and the no-edit features, so decoder behavior is shared within comparisons. We do not treat decoded-image quality across backbones or layers as decoder-independent evidence; our feature-space metrics do not use FeatInv and provide the primary evidence. Any learned decoder would impose priors; FeatInv is a fixed visualization tool, not the object of study.

Figures and reconstruction quality (W3)

We will enlarge and simplify Fig. 2, showing fewer models/layers and moving the rest to the appendix. Reconstructions can be soft because FeatInv decodes compressed features; this equally affects no-edit reconstructions. We keep difficult insertion examples because they expose the difference between position-wise and spatially contextual mappings.

Extension to generative models (Q5)

Applying the framework to FLUX, Qwen-Image-Edit, or Wan is an interesting direction, but requires separating a learned feature-space mapping from the model’s own editing capability, which is beyond this paper’s scope.

We thank the reviewer for taking the time to read our response. To make sure we’ve addressed your concerns as fully as possible before the discussion period ends, the revised PDF now includes the following changes:

FeatInv (W2). FeatInv is now introduced in Methods (“Feature inversion”), with full details in a new appendix section. These cover its architecture, its training (once per backbone and stage, never on manipulated or mapped features) and its role in our pipeline.

Features vs. decoder (W4, Q2, Q4). The new appendix on probe capacity decodes the outputs of all probes for the same test samples with the same FeatInv model. These are the linear map, width-matched nonlinear probes and residual variants. On ConvNeXt and SwinV2, an edit that fails does so for every probe alike, whereas the probes differ in how well they preserve unedited content. This matches the decoder-free feature-space metrics. In the edited region, masked MdnCS differs between probes by at most 0.030.03 in most settings, while the displacement score RΔ2R^2_\Delta differs mainly through unedited regions. Since all conditions share the same decoder, whether an edit is reproduced is determined by the backbone representation, not by the probe or the decoder.

Edit adherence (W3). We validated the VLM judge against human annotations. On the outputs rated by both, they agree in 74% of cases (Cohen’s κ=0.48\kappa = 0.48), and the VLM reports no edit on unedited originals. Per-edit success rates for both are reported in Fig. 4a. Within the same configuration, outputs judged successful have higher masked MdnCS, which does not depend on the decoder.

If any of your concerns remain unresolved, we would be grateful if you could let us know which one, so that we may address it directly.

Hi Authors,

I am still evaluating your response and will update if I have further questions. The featinv section addition does improve readability, but I need to revisit some of the other experiments in the paper and I will get back to you within a few hours.

Reviewer_dorAOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how image edits change a vision model's internal features. It compares a shared linear map with richer probes across several backbones and edit types. The main finding is that the linear map often performs similarly, especially in deeper supervised features.

Critical strengths
  1. I like the extension to semantic edits. Several color and lighting examples are convincing. The differences across backbones including DINOv3 are informative too.

  2. The additional probe baselines and reconstruction controls make this a useful representation study. The paper is also careful about the scope of its linearity claim.

Critical weaknesses
  1. The semantic-sufficiency claim is harder to assess. Even within the affected regions, similar final features may mostly reflect unchanged content. I would like to see how much of the actual edit is captured, compared with doing nothing to the features. The displacement analysis in A.10 covers one backbone at one depth.

The VLM judge is quite lenient for some edits. A localized shape on the bed can count even when it does not clearly resemble a teddy bear (A.7). Per-edit success rates would be more informative than the pooled metric comparison.

Questions for authors
  1. How do the linear map, transformer, and unchanged-feature baseline compare on the actual feature change and on success for each semantic edit?
2 comments

Rebuttal to Reviewer dorA

We thank the reviewer for the positive assessment and the concrete suggestions. We have uploaded an intermediate revision of the paper in which all changes are marked in blue. We will upload the final revised version, which incorporates all reviewer feedback, before the end of the discussion period.

We distinguish (1) similarity to the target features within the edited region from (2) recovery of the feature displacement caused by the edit. Let Δt=Ftgt−Forig\Delta_t=F_{tgt}-F_{orig} and Δm=Fmap−Forig\Delta_m=F_{map}-F_{orig}. We use RΔ2=1−∑i∥Δm,i−Δt,i∥2/∑i∥Δt,i∥2R^2_\Delta=1-\sum_i\lVert\Delta_{m,i}-\Delta_{t,i}\rVert^2/\sum_i\lVert\Delta_{t,i}\rVert^2, with no-edit equal to zero by construction. The all-token score is deliberately strict; masked RΔ2R^2_\Delta restricts the sums to the edit region. Neither score is a semantic-success rate.

W1 (edit captured versus doing nothing)

The expanded analysis covers 297 dataset/backbone/layer/edit combinations. For semantic edits, the reported means of per-combination medians are 0.13 (Linear) and 0.21 (paper Transformer) for masked RΔ2R^2_\Delta; Linear is positive in 80% of combinations. Globally, the corresponding values are −0.17 and −0.02. Thus, the mappings move features toward the target inside the edited region, while small errors over unchanged tokens can make global displacement recovery weak or negative. We will report both views rather than infer edit capture from target similarity alone. Q1 (per-edit feature-space comparison). The following table reports masked median cosine similarity (MdnCS) to target features at feat3, as no-edit / Linear / Transformer, median over samples. These are feature-space similarities, not VLM success rates or displacement RΔ2R^2_\Delta.

EditConvNeXtSwinV2DINOv3
Cars: headlights0.83 / 0.79 / 0.790.75 / 0.73 / 0.720.96 / 0.91 / 0.93
Cars: police lights0.64 / 0.77 / 0.780.54 / 0.63 / 0.740.96 / 0.91 / 0.93
Cars: body blue0.75 / 0.83 / 0.810.71 / 0.82 / 0.830.97 / 0.93 / 0.94
Cars: rims red0.73 / 0.81 / 0.800.70 / 0.75 / 0.790.95 / 0.93 / 0.94
Cars: no mirrors0.84 / 0.83 / 0.820.81 / 0.84 / 0.850.95 / 0.92 / 0.93
LSUN: teddy bear0.15 / 0.72 / 0.750.16 / 0.44 / 0.770.54 / 0.63 / 0.87
LSUN: bedding blue0.67 / 0.80 / 0.800.65 / 0.79 / 0.800.88 / 0.92 / 0.92
LSUN: bed unmade0.65 / 0.73 / 0.730.60 / 0.73 / 0.750.81 / 0.88 / 0.89
LSUN: lamps on0.75 / 0.82 / 0.810.67 / 0.76 / 0.760.93 / 0.91 / 0.93

On this feature-space metric, both probes exceed no-edit on seven of nine listed edits for ConvNeXt and eight of nine for SwinV2. Linear and Transformer are generally close in these values; the Transformer’s clearest listed gains over Linear are for police lights and teddy bear on SwinV2. For the Cars edits, DINOv3’s no-edit similarities are already high (0.95–0.97), indicating that these edits induce comparatively small changes under this metric; this should not be read as evidence of semantic edit success. Per-edit displacement scores and semantic success rates are separate measures and will be reported separately when finalized.

W2/Q1 (VLM success rates and human validation)

We will replace the pooled semantic comparison with per-edit rates to identify differences by manipulation categories. No-edit controls will estimate the judge’s false-positive rate; those rates are distinct from the feature-space no-edit row above. We will also report blinded human ratings on a subset rather than imply that the VLM alone establishes semantic success.

We thank the reviewer for taking the time to read our response. To make sure we’ve addressed your concerns as fully as possible before the discussion period ends: the revised PDF now includes the analyses announced in our rebuttal.

W1. The displacement score RΔ2R^2_\Delta (no-edit =0= 0) and masked MdnCS are now reported for all three backbones, per manipulation group. They cover the linear map, width-matched nonlinear probes and residual variants. Inside the edited region, the probes differ by at most 0.030.03 in masked MdnCS in most settings; masking is the exception, where the transformer clearly helps. The differences in RΔ2R^2_\Delta come mainly from how well unedited content is preserved, not from the edit itself.

W2/Q1. The pooled comparison has been replaced by per-edit success rates for both a revised VLM judge and a human annotator (Fig. 4a). The judge uses a pairwise protocol: it never sees the target image, it can answer “SAME”, and each comparison is asked in both image orders. On unedited originals, it reports no edit (human: 1–5%). On the 2,757 outputs and 150 originals rated by both, human and VLM agree in 74% of cases (κ=0.48\kappa = 0.48). Agreement is high for color edits (rims red κ=0.81\kappa = 0.81). For subtle or localized edits, the VLM under-detects them, and it over-calls police lights. Because human and VLM rates are shown per edit, such differences are directly visible. The remaining annotations are being completed and will be added before the end of the discussion period.

If any of your concerns remain unresolved, we would be grateful if you could let us know which one, so that we may address it directly.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.