acceptodds
Under review as a conference paper at ICLR 2027

PanToMime: Imitating Camera Change by Image Analogy

Abstract

Image analogy provides a natural way to edit images by example: instead of describing a transformation, the user simply demonstrates it. However, when the exemplar also contains a change in camera viewpoint, the desired transformation becomes ambiguous. A user may want to transfer the demonstrated camera movement, match the final exemplar view, or preserve the query view while applying only the visual edit. We introduce PanToMime, a camera-aware image analogy method that understands and controls camera changes while transferring the demonstrated visual transformation. To support this setting, we introduce **PanToMime-Bench**, containing paired visual edits under controlled camera changes, together with **Render2Real**, a geometry-grounded pipeline for generating realistic multi-view training data. Experiments show that existing image-analogy methods can reproduce visual edits but struggle with camera changes, particularly when the exemplar and query start from different viewpoints. PanToMime reduces camera orientation error from for the second-best method to , while retaining competitive edit fidelity with an Edit score of . Our results also suggest that image analogy can provide an effective interface for camera understanding and pose transfer.

Primary area:
generative models
Keywords:
Visual Analogy, In-Context Image Editing, Visual In-Context Learning, Exemplar-Based Editing, Image-to-Image Generation, Flow Matching, Diffusion Transformers, Rotary Position Embeddings, Query-Adaptive Conditioning, Positional Conditioning

Reviews

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

The work targets the task of image analogy where a transition is specified through an input image pair (A, A'), and the same transition should be applied to an input image B in order to create a target B'. In this work, the authors highlight a gap in existing analogy work, which struggles when the pair A, A' specify both an attribute change as well as a change in camera pose. Ideally, a user should be able to decide if they want to transfer only the attribute change, only the camera pose change, or both.

To enable this control, the authors create a set of analogy pairs with known camera estimates using a pipeline based on qwen-based multi angle generation for a sparse initial camera grid, followed by a 3d-lifting approach to rotate the viewpoint to additional views. This 3d-lifting pipeline degrades photorealism, which is later restored via a render-to-real editing pipeline.

Finally, this set is used to tune an image editing model (FLUX.2 Klein) such that it can also be conditioned on the camera pose for the input analogy images and the target via additional tokens. Additionally, the authors train a camera posterior prediction model that attempts to predict the poses for the source images and is then used to estimate the right pose for the target.

On the experimental side, the approach is compared to a set of analogy baselines as well as to leading closed-source API models and it shows remarkable performance (even beating the closed-source models on some of the metrics like camera accuracy). There is a large range of metrics and a large set of qualitative examples and a user study to further back them up.

Critical strengths

The paper's core strength, as I see it, is its aim to help with a large issue that image analogy models have: its very hard to get an analogy pair that only differs in the one attribute you actually care about (unless you can easily describe this attribute in text and do image editing on A, but in this case you don't need an analogy pair since you can just edit image B directly). As such, there's a clear benefit to a method that can disentangle different attributes that may have changed in the A, A' pair, and allow the user to select which of them they want to keep. This work takes a step in disentangling the camera pose, which is an important part of this.

Two further strengths, albeit weaker, are the introduction of a new benchmark for analogy-based editing, and the obvious boost in performance that this approach gives over the baselines. Finally, the ablation into how to condition the resulting model on camera poses may be useful for future work on this topic, but please see weaknesses below.

Critical weaknesses

On the technical side, one glaring weakness here is that qualitative results seem to use the synthetic samples from the distribution on which the authors trained, and which often do not actually look like real images. I don't know how well this approach actually works on real images (at any part of the A, A', B triplet). Some parts of the pipeline, like the camera estimators, may be particularly sensitive to this.

A second weakness is that this work is practically a camera conditioning paper, but it is obviously not the first work to condition generation and editing models on camera poses, and I don't know how well this approach fares compared to other camera control alternatives (for example, image to video models with controlled camera trajectories?).

As a corollary to this, while I noted that I like the desire to disentangle different aspects of the analogy pair, this work is very much designed specifically for camera poses and generalizing it to other attributes is non trivial, so I'm not sure this is even a big step towards that goal.

Finally, this is mostly a systems paper, which essentially combines existing ideas in a pipeline. This seems like a much better fit for a conference like SIGGRAPH and less to ICLR where there is a bigger highlight on works that teach the reader something fundamentally new, provide some new theoretical or empirical understanding, or demonstrate new applications. I personally don't feel like I've gained any new understanding from this work, but this is subjective so I will not insist on a negative score only over this point.

Very minor and didn't affect my score, but didn't know where to list it: Some citations need to be changed from /citep to /citet, for example Reed et all in line 116.

Questions for authors

My main request for authors would be to provide us with results using real images (2+ examples, or better yet some numerical results on a small set) or specify which of the conditioning triplets in the existing paper already use real images if these already exist and I missed them.

Beyond this I'd like to see some comparison to alternative camera conditioning approaches from the existing litterature (perhaps MetaView?) or api-based camera conditioned models (even if they are video models and you take the end frame of the trajectory).

1 comment

Real images You did not miss anything: all test images in the paper come from our generation pipeline. We evaluated camera-only analogies built from CO3D, which have real ground truth (App. A.2, Table 3). Our model estimates the cameras, while SEVA and MetaView are given the ground-truth move.

OA2 error is 66.4° for ours, 41.0° for SEVA and 58.3° for MetaView. OA2 is itself noisy on the real ground-truth images, so these numbers should be read with care. On the 168 analogies where our estimated move matched the label, the errors are 63.2°, 39.8° and 59.6°. That is a tie with MetaView and a clear loss to SEVA. Direction is correct 0.73 of the time (SEVA 0.74), but we under-rotate, with a rotation ratio of 0.61 vs 0.88. It gets worse when the estimate is wrong, so you are right that the estimator is a sensitive part. Image metrics are mixed: PSNR/SSIM slightly favour ours, while LPIPS and the VLM judge favour SEVA (judge, mean of Subject and Background: 6.95 vs 5.28).

How we read this: on real photos the model turns the right way but not far enough, and it is behind a dedicated novel-view model. A figure with selected CO3D cases, including one under-rotation failure, will be added to App. A.2 in the final version. App. A.2, Figure 8 shows analogies on our own phone photos, with A, A′ and B all real.

Other camera-control methods We compared with SEVA and MetaView on the CO3D set above (App. A.2, Table 3). We lose to SEVA, and we are behind MetaView overall; the two are only tied on the matched subset. These are camera-only analogies, so they measure only half of our task, since neither baseline transfers the edit. We did not run image-to-video models with camera trajectories.

Beyond the camera Yes, the method is built for the camera: the estimator, the composition rule and the tokens are all camera-specific. The general recipe might carry over: estimate an attribute for A, A′ and B, compose the target explicitly, then condition on it. But we have no evidence for any attribute other than the camera and will not claim it.

Systems paper Our main claim is that camera-aware image analogy needs one extra constraint: the target camera has to be composed explicitly from the estimated cameras of A, A′ and B and then given to the generator. Without it, analogy models fine-tuned on our data learn the edit but not the camera (judge Camera 3.50 and 3.94 against 7.39 for ours; to be added to the appendix in the final version). How the constraint is posed matters more than how it is encoded: as text, the composed absolute target beats the relative move by 11.35°, and text and camera tokens tie. The gap is widest when the query starts from a different view than the exemplar: commercial editors get 33 to 35° worse, ours 6.7°, and even with the oracle relative move in text the best editor stays about 30° behind. The fix is simple but very effective, and we will restructure the paper around it.

Citations Will fix, including Reed et al. at line 116.

MetaView and API models MetaView is in the CO3D comparison above (App. A.2, Table 3). It gets 58.3° and ours 66.4°, with MetaView given the ground-truth move. So we are behind it overall, and it is only a tie on the matched subset. We did not run API video models.

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

The paper studies image analogy (A : A′ :: B : B′) where the exemplar also contains a camera change. Viewpoint is defined relative to each subject on a grid of 8 azimuths × 2 elevations × 2 distances. A DINOv3-based classifier estimates the camera states of A, A′ and B, and the target camera is computed explicitly as s_B ⊕ (s_A′ ⊖ s_A). All four states are fed as per-frame pose tokens to a LoRA-adapted FLUX.2-Klein-base-9B, alongside the three images and text naming the edit. Training data comes from PanToMime-Bench: 448 subjects, 5,803 views and 15,169 training quadruples. It is built from curated multi-view generations plus rough 3D-proxy renders, which Render2Real (a second FLUX.2-Klein LoRA) converts into photo-style images. On 500 held-out analogies, camera orientation error is 30.7°, against 68–78° for eight baselines, with competitive edit fidelity.

Critical strengths
  • Qualitatively, the method clearly works and beats baselines
  • The ambiguity it identifies is real and well motivated, and the formulation is clean: a subject-relative camera plus explicit composition avoids needing a shared scene frame.
  • The camera gain is large, and OA2 orientation error, RoMa motion ratio and aligned IoU all show improvements. Direction accuracy supports pose conditioning in the reduced-setting ablation.
  • Solid evaluation:
    • the VLM judge is checked against ground truth;
    • outlier filtering is identical for every method;
    • the decoder ablation uses 10 paired seeds (constrained marginal decoding raises exact target-camera accuracy from 76.69% to 79.41%);
Critical weaknesses

Experimental validity

  1. The main comparison mixes method with fine-tuning.
    • PanToMime is fine-tuned on PanToMime-Bench, and every baseline runs zero-shot.
    • The corpus is built with Qwen and R2R, and R2R shares PanToMime's backbone, which may favor PanToMime on reference-based metrics.
    • The paper therefore does not separate how much of the headline gain comes from fine-tuning and how much from camera conditioning. Within the ablation, pose tokens clearly help (37.36° → 23.13°), but the no-camera arm's 37.4° is on a different test set and can't be compared with Table 1's 68°+ baselines.
    • Needed: a full-data fine-tuned model without camera conditioning, and an analogy baseline (e.g., RelationAdapter or LoRWeB) fine-tuned on the same data.
  2. No two-step or camera-informed baselines, and no test of the exemplar images' value.
    • Every method that accepts text receives the benchmark's edit instruction (PanToMime's prompt includes "Edit 1: [instruction]"). So the semantic edit is not inferred purely from the exemplar, and the camera target is computed outside the generator. An arm that keeps the edit text and the camera target predicted from A, A′ and B, but removes A and A′ from the generator's input, would show what the exemplar images add to the semantic edit.
    • An obvious pipeline: estimate the camera move, apply the edit with NanoBanana or GPT-Image, then rotate with a camera-controlled novel-view model. Candidates: the authors' own Qwen multi-angle pipeline, SEVA/SV3D, or their own lift-and-R2R pipeline.
    • Baselines are never given the estimated target camera in text, although the paper's own ablation shows text conditioning helps. Giving existing editors the same predicted absolute target camera would show what the specialized generator and token interface add.
  3. The ablations are fragile.
    • They use 100 test samples and 99 training subjects, and results from multiple training seeds are not reported.
    • The relative-token arm (44.8°) and the arm without the B′ token (47.5°) do worse than no conditioning. The authors' explanation, that the model must still compute the target view itself, is not tested.
    • One arm is missing: the composed absolute target camera given in text.
  4. The VLM judge is not validated against humans (the authors acknowledge this). Its camera axis reaches only 8.5 when the ground truth is scored as its own prediction. The authors therefore treat that score as diagnostic and take headline camera results from OA2, but the Edit and Subject scores still rest on the unvalidated judge. The user study is small (22 raters × 15 sets) and rates joint edit and camera fidelity, not realism.

Domain gap and generalization

  1. No evaluation on independently captured photographs. Held-out subjects test one kind of generalization, but every test image comes from the same generation pipeline.
    • Section 3.2 discloses the Qwen generation and R2R, but not where the seed images come from or what share of final train and test images are generated. Calling views "photographic" or "real" does not establish that they are captured photographs. The pretrained backbone and R2R's appearance restoration may help with real images, but transfer to captured photographs is never demonstrated or measured.
    • These evaluations would need adapting, but are reasonable. One real multi-view dataset (e.g., CO3Dv2 or MVImgNet) could test the central camera claim with real ground truth, using camera-only analogies where A′ is A seen from a new view with no edit. This requires aligning examples to a consistent subject-relative "front" and pairing them suitably. A multi-view human capture dataset (e.g., MVHumanNet) could do the same for pose edits, given matched, transferable pose changes.
    • One standard edit-only analogy benchmark (e.g., the VisualCloze or RelationAdapter evaluation set) would show whether fine-tuning hurt ordinary analogies. The paper's test set covers only rotations plus a few distance-only changes.
  2. The camera estimator is not validated out of distribution. It is trained on the corpus's training split (subject count unspecified) and evaluated only on held-out triples from the same corpus. It should be compared with zero-shot Orient-Anything-V2 on a real viewpoint benchmark (e.g., Objectron or Pascal3D+), with a breakdown for symmetric subjects beyond the OA2 symmetry-filtered metric.
  3. Narrow generalization and no breakdown.
    • Some edit types appear in both training and test, and no evaluation on unseen edit types is reported. The reduced ablation shares all 12 of its edit groups between splits.
    • No results are broken down by subject category (people, animals, vehicles, objects) or by lighting.
    • Beyond the stated limitations, failure cases are not analyzed.

Scope and quality

  1. Narrow scope relative to the claims.
    • "Camera" means 32 discrete object-centric states, with elevation only 0° or 30°. Target views must lie on the grid, and continuous or intermediate angles are not shown.
    • Translation is limited to orbit and shot-distance moves; there is no general 6-DoF camera control (unrestricted translation or roll) and no control of intrinsics.
    • Evaluation centers on one prominent subject. Cluttered and multi-object scenes, and how the reference subject is chosen, are not evaluated.
    • Background parallax is not measured: the background judge ignores placement differences caused by the camera move.
    • A subject-relative orientation cannot distinguish the camera orbiting from the subject turning, and pose edits muddy the canonical front.
    • Remaining errors are substantial: 30.69° camera error, a RoMa motion ratio of 0.542, and a Subject score of 7.76 against 9.43 for the strongest commercial baseline.
  2. Physical consistency and realism are not directly evaluated. The background judge mentions lighting and shadows, but nothing specifically evaluates cast shadows, contact or illumination caused by an edit (a hat's shadow, light from a fireball, relighting after a season change). A human realism study would back the qualitative claim about preserving shadows, and preference or aesthetic scores could supplement it.

Novelty, transparency and presentation

  1. The technical novelty is modest.
    • The method combines pose tokens, a LoRA on a frozen backbone, analytic camera composition, probabilistic decoding and a data pipeline. The main insight is the combined edit-and-viewpoint formulation.
    • Rotation by analogy dates to Reed et al. (2015), as the paper acknowledges, so the novelty lies in combining semantic edits with viewpoint transfer across varied subjects.
    • Related work should position the method against reference-based camera transfer in video (e.g., CamCloneMaster), camera-controlled generation used as a pipeline component, and 3D-aware editing.
    • The topic is in scope for ICLR; the issue is how strong the new evidence is.
  2. Too little detail to reproduce the data or the costs.
    • Data:
      • the source of the seed images;
      • the exact editing pipeline;
      • the share of test ground truth from Qwen versus R2R;
      • reproducible curation criteria and the human effort involved;
      • exact subject split sizes. The paper states that evaluation subjects and their views are excluded from training, but it is unclear whether this also covers R2R's 265 training subjects.
    • Curation filters out inconsistent edited views, but how much edit or identity drift remains between A and A′ is never measured.
    • The 15,169 quadruples reuse 5,803 views, which limits effective diversity.
    • R2R is never quantitatively checked for preserving camera, geometry, identity and edit (Fig. 7 is qualitative). A known render camera does not guarantee a correct label after generative translation.
    • Rendering: no renderer, lighting or settings are given for the proxy renders.
    • Compute: Appendix A reports GPU counts and step counts for the generator, the ablations and R2R, but not for the camera estimator. Also missing are the cost of data generation (Qwen generation, reconstruction, alignment search, R2R) and per-sample inference latency, including camera estimation.
    • Release: there is no statement on releasing code or data. Figures include copyrighted characters (e.g., Superman, Mario, Batman), so how they will be handled in a dataset release should be clarified.
  3. Minor.
    • Sec. 3.3 refers to "the camera change rule from Sec. 3.3"; the rule is in Sec. 3.1.
    • The dataset is called the "Multi-View Analogy Dataset" in Sec. 4 instead of PanToMime-Bench.
    • The RelationAdapter and Hertzmann et al. references are each listed twice.
Questions for authors

Questions

  1. How do a full-data FLUX.2-Klein LoRA without camera conditioning, and a baseline fine-tuned on PanToMime-Bench, compare with PanToMime in Table 1?
  2. How do three alternatives perform:
    • the two-step pipeline (estimate, edit, rotate);
    • GPT-Image and NanoBanana given the same predicted absolute target camera in text;
    • PanToMime with A and A′ removed from the generator, keeping the edit text and the camera predicted from them?
  3. Can you report camera-only analogies on one real multi-view dataset (adapted to a subject-relative frame), results on captured (including cluttered) photographs, and one standard edit-only analogy benchmark? How is the reference subject chosen when a scene has several?
  4. How accurate is the camera estimator on a real viewpoint benchmark, compared with zero-shot OA2?
  5. Where do the seed images come from, and which editor produced the edited views? What share of test ground truth comes from Qwen versus R2R, and do R2R's 265 training subjects overlap the test split? How well does R2R preserve viewpoint, geometry, identity and edit?
  6. Do the ablation rankings hold across seeds and with full data? What do data generation and inference cost? Will code and data be released, and how will copyrighted characters be handled? The highest-priority requests are:
  • matched full-data comparisons;
  • existing editors given the same predicted absolute target camera, and an arm without the exemplar images in the generator;
  • an evaluation on independently captured photographs;
  • a quantitative check of the data and R2R pipeline. The other tests and breakdowns are suggested extensions, not all requirements for acceptance.
3 comments

Method vs. fine-tuning Agreed, and we can give part of what you ask for. We fine-tuned LoRWeB and Edit-Transfer without camera input on the same training data as PanToMime (to be added to the appendix in the final version). We used the same VLM judge and test analogies as Table 1:

methodEditSubjectCamera judgeazimuth error
LoRWeB, fine-tuned7.427.703.5069.0°
Edit-Transfer, fine-tuned6.915.923.9470.6°
PanToMime9.017.797.3923.8°

Fine-tuning teaches the baselines the edit (zero-shot Edit was 3.02 and 1.49) but little of the camera. Their OA2 azimuth error is 72.6° and 72.9°. Copying B, which never moves the camera, gives 82.8°, and ours is 28.6°. A checkpoint curve is flat after step 1000, so longer training is unlikely to close the gap. The full-data FLUX model without camera conditioning is not available.

R2R was trained only on training-split subjects, so it never saw evaluation samples (for subject overlap, see our reply on the data). We cannot rule out that a shared backbone helps on reference-based metrics. OA2 uses the ground-truth B′ only for orientation, and appearance does not enter.

Camera-informed baselines and the exemplar arm Exemplar arm (App. A.4, Tables 5, 6): blanking A and A′ while keeping the edit text and camera tokens raises camera error from 29.59° to 48.80°. So the exemplar pixels carry camera information beyond the tokens. Table 5 also reports fidelity and image-quality metrics; we have no Edit or Subject judge scores for this arm. For the edit, a swap test (to be added to App. A.4 in the final version) points the other way: with conflicting text the output follows the text 100% of the time, and with neutral text it follows the exemplar at chance. Where the text under-specifies the appearance, the exemplar sets the details (App. A.4, Fig. 9), but only in a minority of analogies. You are right that the main contribution of the exemplar is the camera change, and we will rewrite the claims to match.

Editors given the ground-truth A→A′ move in text (App. A.3, Table 4): NanoBanana goes from 67.62° to 60.56° and GPT-Image from 72.42° to 65.21°. LoRWeB and Edit-Transfer are not significant. Ours is 30.31°. We did not run the predicted absolute target as text for the editors. Absolute text beats relative text in our ablation, so it could help them more.

We did not run the two-step pipeline. The closest we have is the real-image comparison with SEVA and MetaView given the ground-truth move (App. A.2, Table 3; reply to Qfom). It covers the camera only.

Ablations We ran the missing arm, the absolute target as text (reduced setting, one seed; to be added to App. D.1 in the final version): no camera 37.09°, relative text 32.62°, camera tokens 22.71°, absolute text 21.26°. Tokens and absolute text are tied (−1.45 [−3.56, +0.59]), and absolute text beats relative text by 11.35° [6.2, 16.8]. What matters is giving the model the composed target, not the token interface, and we will present it that way. This supports our explanation of the weak relative arms for the text interface, but it does not directly test the relative-token arm.

The fragility point stands: we have no additional seeds, no full-data ablation, no human validation of the judge and no new user study. We will say next to the Edit and Subject columns that they rest on the judge.

Captured photographs Following your suggestion, we evaluated camera-only analogies on CO3D with real ground truth (App. A.2, Table 3). Ours estimates the camera; SEVA and MetaView get the ground-truth move. OA2 error is 66.4° / 41.0° / 58.3° (ours / SEVA / MetaView), and OA2 is noisy here (52.0° on ground truth). Where our estimated move matched the label, our direction is right 0.73 of the time vs 0.74 for SEVA, but we under-rotate (ratio 0.61 vs 0.88, and 0.28 on the other analogies). We lose to SEVA on real photos and tie MetaView only on that subset. Wrong estimates are a clear failure source. Examples on our own photos are in App. A.2, Figure 8.

Not available: MVHumanNet pose edits, an edit-only analogy benchmark, cluttered photographs, and the estimator against zero-shot OA2 on Objectron or Pascal3D+.

Generalization and breakdowns On test analogies whose edits never appear in training (App. A.1 lists these instructions; the numbers will be added in the final version), camera error is 29.4° (30.31° overall). Edit score is 6.51 vs 8.55 for NanoBanana, so we keep the camera but lose on the edit. When B starts from a different view than A, our error rises from 25.1° to 31.9°, while the editors get 33 to 35° worse. There is no breakdown by subject category or lighting. The consistent failure on real images is under-rotation (ratio 0.61 on CO3D), and a CO3D figure with such a case will be added to App. A.2 in the final version.

Scope This is correct, and we will narrow the claims. "Camera" means 32 object-centric grid states with orbit and shot-distance moves only. There is no 6-DoF control, roll, intrinsics or off-grid target. We evaluate single prominent subjects, and we did not test cluttered scenes or measure background parallax. The subject-relative frame cannot tell a camera orbit from the subject turning, and pose edits make the canonical front less well defined. On the remaining errors, yes. We lose on Subject score to the commercial editors (7.76 vs 9.43), and 30° of camera error is far from solved. Shadow and lighting realism is not evaluated, and we will remove or soften that claim.

Novelty A fair description, and we accept it. The contribution is the combined edit-and-viewpoint formulation and what the experiments show about it. Fine-tuned analogy models learn the edit but not the camera. Giving the model the composed absolute target matters more than how it is encoded. Editors degrade by 33 to 35° when the query starts from a different view; ours degrades by 6.7°. Related work will add reference-based camera transfer in video (CamCloneMaster), camera-controlled generation as a pipeline component, and 3D-aware editing.

Data and reproducibility Seed and edited images are generated with FLUX.2-klein. The human portraits come from Relation252K, and we will document its licence and terms in the paper. Multi-view images come from Qwen-Image-Edit-2511 (Multiple-Angles). Manual curation kept about 25% of the generated views and about 40% of the subjects. App. A.1 now gives the full provenance, curation criteria, split and quadruple statistics, and the proxy reconstruction, renderer, camera and lighting settings.

Split sizes: 448 subjects, 326 train and 122 evaluation. R2R was trained only on training-split samples (its log: 265 subjects from the train split), so it never saw evaluation samples. However, our audit found a leak (to be documented in App. A.1 and App. B in the final version). 10 evaluation subjects reuse a person from a training sample under a different edit, so "subject-disjoint" is wrong. If we exclude the affected analogies (436 analogies remain), ours is 30.87° and NanoBanana 67.80°. The conclusion holds.

Drift between A and A′ (to be added to App. A.1 in the final version): we asked the same Gemini judge, on all 1,294 A/A′ pairs at the same camera, whether the subject is the same individual. It says yes for 99.6% (subject score 9.59 [9.50, 9.68]). As a negative control, A paired with another subject's A′ from the same edit group is flagged as a different individual in 98.5% of 130 pairs.

On diversity you are right. Each training quadruple is a distinct combination of subjects and camera changes, but the visual diversity of the training data is bounded by its 326 training subjects and 4,175 training views, not by the number of quadruples. App. A.1 now states this.

We do not have a quantitative check that R2R preserves camera, geometry, identity and edit. We also do not have the data-generation cost, the estimator training cost or the inference latency.

Release: the code and PanToMime-Bench will be released upon acceptance. Before release we will remove all images of copyrighted characters (e.g., Superman, Mario, Batman) from the dataset and replace them in the paper's figures (App. A.1).

Minor We will fix all three: the Sec. 3.3 self-reference (the rule is in Sec. 3.1), the dataset name in Sec. 4, and the duplicated RelationAdapter and Hertzmann et al. references.

Your four priorities Where we stand on your four priorities (details are in the replies above).

  1. Matched comparison, partly. Two analogy baselines are fine-tuned on the same data and scored with the judge and with OA2. They reach 72.6° and 72.9° azimuth error against the requested move (ours 28.6°, never moving the camera 82.8°). There is no full-data FLUX model without camera.
  2. The exemplar arm is done for camera error (+19.2° without A and A′). Camera-informed editors only partly: they were given the oracle relative move, not the predicted absolute target.
  3. Captured photographs, partly. We have CO3D with real ground truth (App. A.2, Table 3), and there we lose to SEVA.
  4. Data and R2R check, partly. We give provenance, split sizes, a disclosed subject leak with a leak-free re-score, and an identity-drift number. There is no quantitative R2R check.
Reviewer_3AVkOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces PanToMime, a camera-aware image analogy framework that transfers both visual edits and camera transformations from an exemplar pair to a query image. The method estimates object-centric camera poses, explicitly composes the target viewpoint, and conditions a pretrained FLUX model using camera tokens. The authors also introduce Render2Real to construct multi-view training data with controlled camera transformations.

The method achieves substantial improvements in camera orientation accuracy over existing image analogy and general-purpose image editing models.

Critical strengths
  1. The problem of jointly transferring visual edits and relative camera transformations is well motivated.

  2. The proposed data construction pipeline provides controlled supervision for a task where suitable training data is limited.

  3. The results demonstrate substantial improvements in camera accuracy, with reasonably competitive edit fidelity. The camera-conditioning ablations and human evaluation provide additional supporting evidence.

Critical weaknesses

1. Insufficiently controlled baselines.

The main comparisons are against image analogy and general-purpose image editing models that must infer the camera transformation directly from the input images. In contrast, PanToMime explicitly estimates and composes camera poses before generation.

This makes it difficult to determine how much of the improvement comes from the proposed architecture rather than explicit camera conditioning and specialized training. A stronger comparison would provide existing image editors with the same estimated target camera, potentially through text-based camera instructions. The text-conditioning ablation partially addresses this question, but only in a smaller experimental setting.

2. Limited generalization evaluation.

The camera representation is restricted to a discrete set of azimuths, elevations, and distances, and the evaluation is conducted on the proposed benchmark. Although the training and evaluation subjects are disjoint, edit categories are shared, and the dataset relies on a curated, geometry-grounded generation pipeline.

It remains unclear how well the approach generalizes to unconstrained real-world images, unseen edit types, more complex scenes, and viewpoints outside the predefined camera grid.

3. Unclear contribution of visual analogy beyond text-guided editing.

The model already receives text instructions describing the visual edits, while the camera transformation is computed separately. It is therefore unclear how much the exemplar pair contributes to the actual image-generation process.

An important missing ablation is to remove the exemplar images from the generator while retaining the edit instruction and estimated target camera. This would help establish whether the method genuinely benefits from visual analogy or primarily performs text-guided editing with explicit viewpoint conditioning.

Questions for authors
  1. How would strong existing image editors perform if provided with the same estimated target camera used by PanToMime?

  2. How much performance is lost when the exemplar images are removed from the generator while retaining the text edit instruction and camera conditioning?

  3. Does the method generalize to unconstrained photographs and camera viewpoints outside the discrete training grid?

2 comments

Weakness 1 (controlled baselines) We agree that the original comparison did not separate these factors.

For camera information, we gave the editors the oracle relative move A→A′ in text (App. A.3, Table 4). NanoBanana improves from 67.62° to 60.56° and GPT-Image from 72.42° to 65.21°. Ours reaches 30.31° with estimated cameras. We did not run the composed absolute target as text for the editors. In our ablation the absolute target is 11.35° better than the relative move, so the editors may gain more from it.

For specialized training, we fine-tuned LoRWeB and Edit-Transfer on our data without camera input (to be added to the appendix in the final version). They reach Edit 7.42 and 6.91 but Camera 3.50 and 3.94. Ours reaches Edit 9.01 and Camera 7.39. Their azimuth error (72.6° and 72.9°) is close to never moving the camera (82.8°). They learn the edit, not the camera.

The ablation is still small (one seed), but we added an arm that gives the composed absolute target as text (to be added to App. D.1 in the final version). It gets 21.26°, camera tokens get 22.71° and no camera gets 37.09°. Text and tokens are tied (−1.45 [−3.56, +0.59]), so the gain comes from composing the target camera, not from the token interface. The paper will say this.

Weakness 2 (generalization) On 69 test analogies with edit types never seen in training, camera error is 29.4°, against 30.31° on the full set. App. A.1 lists the unseen instructions; these numbers will be added in the final version. Edit quality drops to 6.51 (NanoBanana 8.55). Camera transfer carries over to new edits; edit quality does so only partly. Results on real photographs (CO3D) are in App. A.2 (Table 3).

We have no results for off-grid viewpoints or complex scenes. The model works on a grid of 32 object-centric camera states with one prominent subject, and we will state this as a limitation.

Weakness 3 (contribution of the exemplar) We blanked A and A′ and kept the edit text and camera tokens. Camera error rises from 29.59° (control re-run) to 48.80°, or +19.2° (App. A.4, Tables 5 and 6). The model therefore also reads the camera from the exemplar pixels.

For the edit, we do not want to overclaim. In a swap test (to be added to App. A.4 in the final version) with neutral text, the output follows the exemplar's hat in 20% of cases, which is chance, and when the text conflicts with the exemplar, the output always follows the text. Where the text under-specifies the appearance, the exemplar sets the details (App. A.4, Fig. 9: the mohawk colour, the pendant, the kayak colour), but this happens in a minority of analogies. The main contribution of the exemplar pair is the camera change, and the paper will describe the method this way.

Q1 (editors given the camera) See Weakness 1 and App. A.3 (Table 4). Direction accuracy is 0.68 for the best editor and 0.93 for ours. One correction: App. C states that no method is told the camera in words, but PanToMime does receive the composed target camera. We will fix that sentence.

Q2 (removing the exemplars) Camera error rises by 19.2° (29.59° to 48.80°); see Weakness 3 and App. A.4. Table 5 also reports fidelity and image-quality metrics for this arm; we have no Edit or Subject judge scores for it.

Q3 (real photographs, off-grid views) On CO3D (App. A.2, Table 3), SEVA and MetaView get the ground-truth move and ours estimates it. OA2 error is 66.4° for ours, 41.0° for SEVA and 58.3° for MetaView. OA2 is noisy here, with 52.0° error on the ground-truth images. Where our estimated move matched the label, our direction is about as accurate as SEVA's (0.73 vs 0.74), but we under-rotate (ratio 0.61 vs 0.88). We are behind SEVA on real photos, and we only tie MetaView on that subset. Qualitative results on our own photos are in App. A.2, Figure 8.

Off-grid viewpoints are not supported and were not tested. We agree this is a limitation.

Ethics flag No reason was given, so we answer the most likely concern. The portraits of real people come from Relation252K, and we will document its licence and terms. All other seed and edited images are generated with FLUX.2-klein. If the flag is about something else, please tell us.

Comments on the paper

1 comment

We thank the three reviewers for their careful and detailed reviews. All three found the problem well motivated: transferring a demonstrated camera change together with the edit. 3AVk and ozKF pointed to the large gain in camera accuracy, ozKF to the evaluation protocol, and Qfom to the value of separating the camera from the other changes in an exemplar pair. The reviews were also clear about where our evidence was thin, and most of the new results below were run because of your requests.

Below are the results and corrections that concern all three reviews. Everything here is on the paper's FLUX.2-klein model unless marked. Camera error is OA2 geodesic error. All new numbers are recomputed with one script (ours 30.31°; the paper's table reports 30.69°).

Editors given the camera move. The baselines received the ground-truth A→A′ move in text. This is an oracle relative move, while our model only gets estimated cameras.

(App. A.3, Table 4)

methodno camera textoracle move in text
NanoBanana67.62°60.56°
GPT-Image†72.42°65.21°
LoRWeB76.58°75.79°
Edit-Transfer69.66°70.36°
PanToMime (estimated cameras)30.31°n/a

The two commercial editors gain about 7° (95% CIs [−11.4, −2.75] and [−10.4, −4.18]). The two analogy models do not change significantly. Direction accuracy is 0.93 for ours and 0.68 for the best editor. Editors with the composed absolute target were not tested. The baseline columns are a fresh re-score of the Table 1 images, so they are not identical to the Table 1 values. († GPT-Image is scored on two fewer analogies.)

Baselines fine-tuned on our data. LoRWeB and Edit-Transfer fine-tuned without camera input on the same training data reach Edit 7.42 and 6.91 (ours 9.01), but their OA2 azimuth error against the requested move stays at 72.6° and 72.9°, close to never moving the camera (82.8°). Ours is 28.6°. Details in our reply to ozKF (to be added to the appendix in the final version).

Real-image results (CO3D and our own photos) are in App. A.2 and our reply to Qfom. On real photographs we are behind SEVA.

Corrections to the paper: The splits are not subject-disjoint (to be corrected in App. A.1 and App. B). Our audit found 10 evaluation subjects that reuse a person from a training sample under a different edit. Without the affected analogies (436 remain) ours is 30.87° and NanoBanana 67.80°.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.