acceptodds
Withdrawn from ICLR 2027

PhyScale: Physically Grounded Object Size Correction in Text-to-Image Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Generated images can appear photorealistic while depicting objects at implausible relative physical sizes. Correcting these errors requires distinguishing physical size from its depth- and pose-dependent appearance. We introduce PhyScale, a framework that formulates size correction as geometric inference and planning, separating physical-size changes from their image-space realization. Using typical real-world object dimensions from a knowledge base and estimated scene geometry, it computes size, depth, and placement commands for an image editor controlled through spatial rotary positional embeddings. To learn the associated appearance changes, we adapt the editor on only 2,468 paired renders from controlled 3D object transformations. The resulting editor achieves the best target-fidelity and identity-preservation metrics among the compared editors on prescribed edits. Combined with our geometric inference and planning front end, it delivers aggregate GenScale correction performance comparable to a state-of-the-art correction pipeline at approximately one-twentieth of its estimated inference cost.

Primary area:
generative models
Keywords:
Text-to-Image Generation, Image Editing, Relative Physical Size Correction, Geometric Reasoning, Rotary Positional Embeddings, Data-Efficient Adaptation

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_AWV7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PhyScale, which aims to correct implausible relative physical scales among objects in text-generated images by formulating the problem as explicit geometric inference, constrained planning, and controlled image editing. Its main contributions include a size observer based on object pose and scene geometry, a planner that jointly determines object size, depth, and placement, and a RoPE-controlled actuator built on Qwen-Image-Edit. Experiments show that PhyScale achieves scale-correction performance comparable to baselines on GenScale, while offering advantages in target fidelity, identity preservation, and inference cost.

Critical strengths
  1. The core insight is both clever and intuitive, cleanly separating physical-scale reasoning from image-space editing.
  2. The method is well-motivated and clearly presented, and the overall paper is polished and easy to follow.
Critical weaknesses
  1. Unintended changes to other objects during scale correction. The proposed method does not always preserve other objects in the scene during editing. For example, in Figure 1(a), although the stated correction is to enlarge the stop sign relative to the trash can, the glasses become noticeably smaller relative to the trash can after editing. It would therefore be useful to quantitatively evaluate whether the sizes of other objects are preserved after editing.
Questions for authors
  1. Could the authors clarify how the proposed method prevents unintended scale changes to objects that are not targeted by the editing instruction? In particular, could the authors quantitatively evaluate the scale preservation of non-target objects before and after editing, for example by measuring changes in their relative sizes? If this concern is satisfactorily addressed, I would be willing to raise my score.
Reviewer_CY4rOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In this submission, the authors proposes PhyScale, a framework a framework connecting size measurement, planning, and editing through explicit geometric commands. The pipeline includes a pose-aware observer estimates size errors, a geometric planner that determines size and placement corrections, and an image editor adapted using synthetic 3D rendering pairs. The main contribution includes the problem formulation, a data-efficient RoPE-controlled image editor, and the system itself.

Critical strengths
  1. Clear motivation and problem formulation. Distinguishing physical size from apparent size is well motivated. The paper is well structured and easy to follow.

  2. Impressive results. The proposed model achieves competitive results on ObjectMover, outperforming the compared baselines on diverse metrics.

  3. Comprehensive experimental analysis. The paper provides extensive comparison experiments, component studies, and statistical analyses.

Critical weaknesses
  1. Over-reliance on fixed category-level size priors.

The method defines correct relative scale using a fixed knowledge base of canonical object dimensions. However, real-world objects can have substantially different valid sizes even within the same category. Therefore, deviations from canonical size ratios do not necessarily indicate physical errors, raising concerns about the general applicability of this formulation.

  1. Potential disruption of original scene composition.

The planner automatically adjusts object positions without user-specified destinations. Although geometric constraints and minimal image changes are considered, these adjustments may still alter the original spatial and semantic relationships between objects, potentially violating the intended scene composition.

  1. Limited novelty.

The framework combines several established techniques, including canonical size priors, geometric planning, and RoPE-based editing with synthetic 3D supervision. The paper would benefit from a clearer discussion of its methodological novelty relative to prior geometry-controlled editing approaches, such as RoPEMover and GIVE.

Reference

RoPEMover: Depth-Aware Object Relocation via Positional Embeddings. Neurips 2026.

Geometry-Instructed Video Editing. Siggraph Asia 2026.

Questions for authors
  1. How does the method distinguish genuine physical-size errors from valid intra-category size variations?

  2. How does the planner ensure that automatic object repositioning preserves the original spatial and semantic relationships? Is there any mechanism for users to constrain or prevent unintended position changes?

Reviewer_4WqgOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

PhyScale addresses relative physical-size errors in generated images using a geometric observer, a constrained placement planner, and a RoPE-controlled Qwen-Image-Edit actuator adapted with LoRA on 2,468 rendered pairs. The system separates physical-size correction from its projected appearance and offers an inspectable alternative to VLM-based correction decisions.

Critical strengths
  1. The distinction between physical size and apparent image size is meaningful. The geometric front end offers explicit constraints and low estimated inference cost.
  2. The matched text-box and RoPE retraining experiment controls the backbone, training data, and budget, providing evidence that the command interface matters beyond fine-tuning alone (Tables 6 and 25).
  3. The paper includes useful component swaps, placement stress tests, and supervision ablations, and acknowledges several limitations of these experiments.
Critical weaknesses
  1. The methodological advance appears incremental. The actuator combines an existing editing backbone, LoRA adaptation, spatial RoPE manipulation, and rendered supervision. The cited concurrent work RoPEMover explores a closely related RoPE-based relocation mechanism with parameter-efficient adaptation and synthetic supervision. PhyScale differs through automatic size diagnosis and planning, as well as its rendered-only training recipe, but the distinct algorithmic contribution remains limited in my assessment. A clearer explanation of this boundary is important. My concern is the extent of the methodological advance, rather than the absence of a direct comparison with concurrent work alone.
  2. The independent downstream benefit of the geometric front end is not convincingly established. With the PhyScale editor fixed, the Gemini front end achieves a lower size-error point estimate than the geometric front end, 0.570 versus 0.654 (Table 4). The adjusted tests do not establish a significant difference; they also do not establish equivalence or noninferiority. Similarly, full-placement versus in-place and predicted-pose versus upright comparisons do not detect downstream size improvements on the development sources (Table 32). The stress tests and cost savings are useful, but provide a narrower justification than demonstrating that the additional geometric machinery materially improves the main correction task.
  3. Evidence for the supervision design is mixed. The matched small-data comparison does not detect a target-DINO advantage for 3D over 2.5D supervision on ObjectMover (Table 24). Human judgments support better physical plausibility with 3D supervision, but with a lower completion rate (Table 29). These results support a specific scene-effect benefit rather than a broadly superior editing recipe. The larger-data comparison also changes the asset pool and command composition, making attribution of the gains less clear.

Minor comment

The teaser could distinguish the source-image generators from the correction method more clearly. The prominent generator names can initially obscure that the corrected outputs are produced by PhyScale.

Overall assessment

The system has practical value and the experimental effort is substantial. However, I remain unconvinced that the methodological novelty and demonstrated incremental benefits are sufficient for acceptance. My main concern is the limited advance beyond existing components, particularly given the inconclusive downstream evidence for the distinctive geometric front end.

Questions for authors
  1. Which component constitutes the principal methodological advance beyond existing geometric editing and RoPE-based control?
  2. Under what conditions does full planning reliably improve over in-place correction on representative, held-out inputs?
  3. Can the benefits of the size–depth training distribution be isolated from those of the command interface and dataset composition?
2 comments
AuthorsOpenReview ↗

Response to Reviewer 4Wqg (1/2) — Methodological Novelty and Geometric Planning

Reviewer 4Wqg — Methodological Novelty and Experimental Evidence

We thank the reviewer for the thoughtful assessment. We would like to clarify three points regarding the methodological contribution and the evidence supporting our design.

W1/Q1. Methodological novelty.

The principal contribution of PhyScale is a physically grounded correction framework that determines what must be corrected and how to execute that correction.

Existing geometry-controlled editors, including RoPEMover and GIVE, primarily execute externally specified transformations. PhyScale instead (1) infers relative physical-size errors while accounting for depth and pose, (2) plans feasible size and placement corrections under physical and image-space constraints, and (3) executes these corrections through an actuator that independently controls physical size and depth.

The actuator is specifically designed for this task. Its size–depth parameterization and supervision include pure-size, pure-depth, joint, and compensating transformations, providing separate commands for physical size and depth. Our matched RoPE-versus-text-box experiment (Tables 6 and 25) isolates the contribution of this command interface using the same backbone, data, and training budget.

We additionally evaluated the official RoPEMover implementation on 188 real-image ObjectMover requests. PhyScale achieves 0.899 versus 0.861 target DINO (paired gain +0.038, 95% CI [0.028, 0.049]). When RoPEMover receives exactly the same edit parameters, its result is 0.866. PhyScale also achieves better identity preservation (0.856 versus 0.802) and DreamSim (0.105 versus 0.139) against the official implementation.

We cited RoPEMover as concurrent work. Its June 25 arXiv preprint was available at our initial submission on September 8. NeurIPS author notifications followed on September 24, and the inference code and pretrained weights were released on October 4. The ICLR 2027 Reviewer Guide makes comparisons with arXiv-only work optional and uses a two-month window for recent peer-reviewed work.

W2/Q2. Value of the geometric front end and full planning.

The geometric front end provides explicit physical reasoning, inspectable decisions, local inference without an external VLM, and substantially lower computational cost. In Table 3, both observers received the reference dimensions and had to identify the object with the size error and estimate its correction within 20%. PhyScale met both criteria on 56.9% of Blender scenes versus 44.8% for Gemini, and on 44.0% of photorealistic scenes versus 38.5%.

On GenScale, PhyScale achieves an aggregate error of 0.640 versus 0.649 for ReScale, while reducing estimated inference cost from $0.696 to $0.037 per image (approximately 19×). The geometric front end alone costs approximately $0.006 versus $0.454 for a Gemini-based front end (Appendix E.3). Table 4 shows that PhyScale can use Gemini as a substitute front end, with a lower size-error point estimate in that comparison.

The planner also addresses a concrete geometric problem that arises when resizing at the original position would violate placement constraints. In our 197-image placement ablation (316 edits), strictly in-place resizing produces 96 actions violating the planner's geometric constraints. Full planning resolves 91 of these cases, including all physical violations under the planner's checks. These conflicts arise when resizing would compromise support, cause overlap or occlusion, extend the object beyond the image boundary, or exceed the editor's range.

On 116 held-out Blender stress scenes, both planners received the same observations and prescribed resizes. Our planner satisfies the evaluated constraints in 87.9% of cases versus 80.2% for Gemini, a reported difference of 7.8 percentage points (95% CI [1.6, 14.5]).

We also measured downstream size error on the 21 development images whose plans changed under either order-preserving variant. With the same actuator, size commands, and seed, mean error was 0.627 with full placement and 0.730 with in-place resizing (in-place minus full +0.103, 95% CI [−0.151, +0.381], bootstrapped by source). Table 32 reports a size-gain difference of +0.038 in favor of full placement on the broader set of 197 images (95% CI [−0.021, +0.104]).

AuthorsOpenReview ↗

Response to Reviewer 4Wqg (2/2) — 3D Supervision and Size–Depth Control

W3/Q3. Contribution of 3D supervision.

The advantage of 3D supervision is physically consistent scene editing, particularly contact and shadows. Target-object feature similarity alone provides an incomplete measure of these scene effects.

We isolate this effect through a controlled comparison using the same 783 training requests, scene assets, commands, model, and training budget, changing only the supervision from 2.5D compositing to 3D rendering.

  • 83.5% tie-adjusted human preference for the physical plausibility of 3D-supervised outputs (95% CI [76.6%, 89.8%]). Among the 84 requests completed by both models, the preference was 88.9% (Table 29).
  • 87.3% preference from an independent Gemini-based evaluation.
  • +1.16 dB contact-shadow PSNR (95% CI [+0.52, +1.79]).

A geometrically correct resize can still leave inconsistent shadows, contact, or occlusion.

To test size and depth control, we held the trained actuator fixed and evaluated nine command combinations across 24 groups (Appendix E.6.7). A response of 1 corresponds to the requested change. We measured responses of 0.910 for size (95% CI [0.798, 0.991]) and 1.014 for depth (95% CI [1.005, 1.030]). The confidence intervals for the cross-responses include zero, consistent with separate control of size and depth in projection. Tables 6 and 25 isolate the effect of the command interface, the matched 783-pair study isolates the effect of 3D versus 2.5D supervision, and the command-response sweep tests whether the trained actuator can respond separately to size and depth commands.

Minor comment. We will clarify the teaser labels to distinguish the source-image generators from the PhyScale correction outputs.

Comments on the paper

1 comment

Dear reviewers,

Thanks for your continued efforts. The authors have provided a rebuttal to the initial reviews. Please review it carefully, along with all other reviews and the corresponding author responses, and consider whether you have any further questions for the authors or whether the rebuttal has addressed your concerns. There are only a few days remaining before the discussion period closes, so please use this time to seek any further clarifications or raise any additional questions about this paper.

Best,

AC

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

Related papers

Discussion (0)

Sign in to comment.