acceptodds
Under review as a conference paper at ICLR 2027

PhysFlow: Physics-Aware Image Editing via Continuous Latent Thoughts

Abstract

Instruction-based image editing has achieved impressive progress in visual quality and prompt adherence, yet maintaining physical consistency remains challenging when edits involve complex dynamics such as refraction, deformation, or object interactions. Existing efforts either rely solely on textual reasoning, which struggles with fine-grained visual changes and lengthy token sequences, or resort to video generative models to synthesize intermediate transformation states, incurring prohibitive sampling overhead. In this paper, we present PhysFlow, an efficient continuous-thought framework that jointly internalizes textual editing planning and visual evolution within a compact latent space. By encoding the underlying transformation process as continuous latent thoughts, this way elegantly bypasses the limitations of both language-only reasoning and multi-frame rendering. Specifically, we train PhysFlow via a three-stage curriculum: first learning from explicit multi-step textual rationales and temporally ordered visual states, then progressively compressing them into continuous latent tokens, and finally steering the image editor with the learned latent editing dynamics. To enable this curriculum, we curate a high-quality dataset (namely PhysTrace) spanning diverse physical phenomena, annotated with hierarchical reasoning chains and dense visual trajectories. Extensive evaluations on PICABench and RISEBench demonstrate that our PhysFlow substantially outperforms competitive baselines in physical plausibility while introducing minor inference latency.

Primary area:
generative models
Keywords:
Image Editing, Latent Chain-of-Thought

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_xuAmOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

PhysFlow is a framework for physics-aware image editing that compresses textual reasoning and intermediate visual states into continuous latent representations. Built on Qwen-Image-Edit-2511, it uses a three-stage curriculum to learn explicit trajectories, progressively compress them, and adapt the image generator. The paper also introduces PhysTrace, a dataset of 11,448 annotated video trajectories. Experiments on PICABench and RISEBench report improvements over the backbone with modest additional inference cost, supported by automated evaluation and a user study.

Critical strengths
  1. The paper presents a coherent framework for incorporating physical transformation processes into image editing.
  2. The reported improvements are supported by complementary evaluation methods.
  3. The framework achieves a useful balance between editing quality and inference cost.
Critical weaknesses
  1. The incremental methodological contribution over existing latent reasoning approaches is not sufficiently established. Monet already explores visually informed latent reasoning and distillation, and is acknowledged in the paper. More directly, EndoCoT uses the same Qwen-Image-Edit-2511 backbone, recurrent hidden-state updates, and intermediate visual supervision, while avoiding intermediate image decoding at inference. Therefore, guiding image generation through continuous latent reasoning without rendering intermediate images does not, by itself, clearly distinguish PhysFlow.
  2. The experimental comparison does not adequately cover a closely related method for physics-aware image editing. PhysicEdit learns latent transition priors from videos for physical image editing without requiring intermediate video generation at inference. This directly relates to PhysFlow’s motivation concerning visual dynamics and computational efficiency. However, the paper primarily uses this work as the source of PhysicTran38K and does not provide a method-level comparison. In particular, the score of 53.82 in Tables 4 and 9 corresponds to PhysFlow trained on PhysicTran38K, rather than PhysicEdit itself. This dataset comparison therefore does not establish an advantage over the existing approach.
  3. The ablations do not sufficiently separate the benefits of dataset adaptation from those of structured latent reasoning. The no-reasoning configuration in Table 3 scores 54.26, matching the original backbone, and is described as the base model. The paper does not clearly report a direct fine-tuning baseline trained on the same PhysTrace source–target pairs with comparable generator adaptation and training budget. PICA100K-SFT cannot substitute for this control because it uses different data. Furthermore, the textual, visual, and joint reasoning ablations support the complementarity of the two modalities, but do not establish whether the progressive compression curriculum is more effective than learning latent conditioning with terminal supervision alone.
  4. The claim that general editing capabilities are preserved is stronger than the category-level results support. Although the overall ImgEdit-Bench score decreases only slightly, from 4.41 to 4.39, some categories show larger declines: Hybrid drops from 3.76 to 3.26, and Extract from 4.16 to 3.91. These results support broadly maintained aggregate performance, but not an unqualified claim of improvement without compromising general editing capabilities. This is a secondary limitation that calls for a more precise discussion of the trade-offs.
Questions for authors
  1. What is the central methodological contribution relative to EndoCoT and Monet?
  2. Was the no-reasoning configuration in Table 3 adapted using the same PhysTrace training data?
  3. How does performance change when the same recurrent latent tokens, final visual block, and trainable modules are retained, but training uses only terminal visual supervision and the generation loss?
  4. Could you provide a method-level comparison with PhysicEdit under a consistent evaluation protocol, explicitly accounting for backbone differences?
3 comments
AuthorsOpenReview ↗

Response to Reviewer xuAm

We sincerely thank you for the time and effort you devoted to reviewing our manuscript and for your detailed, constructive feedback. We appreciate your recognition of our coherent framework, the complementary evaluations supporting our improvements, and our inference efficiency.

We have carefully considered each of your concerns and respond point by point below.


1. Central contribution relative to EndoCoT and Monet.

As suggested, we clarify the key contributions of our PhysFlow relative to EndoCoT and Monet.

1.1 Comparison with EndoCoT.

(a) Our latent representations receive richer process supervision. EndoCoT learns latent tokens for intermediate states through the diffusion loss, with the alignment loss applied only at the final state. In contrast, our PhysFlow progressively replaces explicit reasoning blocks with compact latent tokens, one block at a time. Specifically, these latent tokens receive indirect supervision from text prediction and visual alignment losses on subsequent uncompressed reasoning blocks, together with the diffusion loss computed on the final edited image. This provides richer supervision for learning compact and informative latent representations.

(b) Our diffusion supervision targets only the final edited image. During its initial training stage, EndoCoT learns latent tokens through diffusion losses computed on both the corresponding intermediate and final images, with the alignment loss applied at the final reasoning state. In contrast, PhysFlow supervises intermediate states in feature space and applies the diffusion loss only to the final edited image, reducing the diffusion computation required for intermediate-state supervision.

1.2 Comparison with Monet.

(a) Our textual and visual reasoning share a unified latent space. Monet solely distills visual evidence into latent embeddings while retaining explicit textual reasoning. In contrast, our PhysFlow jointly encodes textual physical explanations and visual trajectories within a unified multimodal latent space, without generating explicit reasoning text at inference.

(b) Internalizing multimodal reasoning requires a non-trivial training strategy. Monet adopts a three-stage supervised fine-tuning strategy: it first adapts the MLLM to interleaved textual-visual reasoning, then generates latent embedding targets in the second stage to supervise latent predictions in the third stage. In contrast, our PhysFlow employs a progressive compression mechanism without requiring a separate stage to generate latent embedding targets. Instead, the compressed latent tokens receive indirect supervision from text prediction and visual alignment losses on subsequent uncompressed reasoning blocks, together with the diffusion loss, to learn compact latent representations.

(c) Our framework targets reasoning-to-editing rather than reasoning-to-reasoning. Monet primarily focuses on enhancing the visual reasoning capabilities of MLLMs by reasoning over visual evidence directly within the latent visual space (reasoning-to-reasoning). In contrast, our PhysFlow aims to transfer physical knowledge from MLLMs into physics-aware image editing by explicitly modeling causal visual state transitions in the latent reasoning process (reasoning-to-editing).

We have included both EndoCoT and Monet in Section 2.3 of Related Work.

AuthorsOpenReview ↗

Response to Reviewer xuAm

2. Comparison with PhysicEdit.

Thank you very much for pointing out the missing discussion of PhysicEdit.

As suggested, we have included a discussion of PhysicEdit in Section 2.3, “Latent Reasoning”, and Summary in Related Work.

Although both our proposed PhysFlow and PhysicEdit leverage textual explanations and visual trajectories to enhance the physical understanding of image editing models, they differ fundamentally in their underlying mechanisms, as detailed below.

(a) Our textual reasoning representation is different. PhysicEdit generates textual reasoning traces in the explicit semantic space of the MLLM, while separately modeling visual representations in the latent space. In contrast, our PhysFlow jointly performs textual and visual reasoning within a unified latent space. This formulation allows latent textual reasoning to guide visual-state prediction without generating explicit reasoning text at inference.

(b) Our visual reasoning explicitly models temporal dependencies. PhysicEdit jointly computes a set of transition-query representations, whereas PhysFlow autoregressively predicts successive visual-state representations conditioned on preceding latent textual explanations and visual states. This causal formulation explicitly captures the temporal dependencies underlying physical transitions.

(c) Learning unified latent reasoning requires a non-trivial training strategy. PhysicEdit trains its transition queries through feature alignment with projected DINOv2 and VAE features extracted from intermediate video frames. In contrast, our PhysFlow adopts a progressive training curriculum. It first supervises the prediction of all textual explanations and intermediate and final visual states using ground-truth text tokens and visual features extracted by the MLLM's visual encoder, respectively. It then progressively replaces the reasoning blocks for textual explanations and intermediate visual transitions with model-generated predictions, one at a time. Each replaced block receives indirect supervision through the original prediction loss applied to subsequent unreplaced positions, jointly with the diffusion loss. This stage encourages the model to learn causal dependencies across successive multimodal reasoning steps. Finally, the image generator is adapted to the learned latent representations using the diffusion loss.

We have added a quantitative comparison with PhysicEdit on PICABench-Superficial and ImgEdit-Bench in Table 1 and 5, respectively. Using the same Qwen-Image-Edit-2509 backbone, our PhysFlow achieves an overall score of 54.12% on PICABench-Superficial, outperforming PhysicEdit (52.35%) by 1.77 percentage points, with higher scores in seven out of eight categories. Moreover, our PhysFlow achieves an overall score of 4.31 on ImgEdit-Bench, compared with 4.05 for PhysicEdit, indicating stronger generalization to general image editing tasks.

MethodLPLSERFLRFRDFMCSLGSTLSTOverall
PhysicEdit (2509)51.6261.6351.5652.9644.0345.8957.6453.0752.35
PhysFlow (2509)52.0162.8257.0253.1443.3246.8161.4253.1654.12

 

MethodAddAdjustExtractReplaceRemoveBackGStyleHybridActionOverall
PhysicEdit (2509)4.314.172.344.454.444.034.743.364.854.05
PhysFlow (2509)4.454.253.544.694.484.214.773.384.824.31

We have also added a comparison of inference latency between our PhysFlow and PhysicEdit in Table 4.

MethodLatency (s)
PhysicEdit52.19
PhysFlow (Ours)49.39
AuthorsOpenReview ↗

Response to Reviewer xuAm

3. Non-reasoning baseline trained on the same PhysTrace source–target image pairs.

Following your suggestion, we train a non-reasoning image editor on the same PhysTrace source–target pairs without reasoning annotations, using the same generator backbone (Qwen-Image-Edit-2511). Both methods use rank-512 generator LoRA and 537 generator updates, matching the generator adaptation configuration and budget. PhysFlow additionally trains its planner for 1,253 updates. We refer to this baseline as Qwen-Image-Edit-2511 + PhysTrace (w/o Reasoning) and report its results in Table 8.

As shown, the new baseline outperforms the original backbone by 0.86 percentage points. Under the same source-target image pair training setting, our curated PhysTrace achieves higher overall performance than PICA100K (i.e., Qwen-Image-Edit-2511 + PICA100K (w/o Reasoning)), highlighting its superior data quality and contribution to training efficiency. Furthermore, our latent reasoning framework further improves the overall score from 55.12% to 56.32% over this non-reasoning baseline, demonstrating the effectiveness of our proposed approach.

ConfigurationLPLSERFLRFRDFMCSLGSTLSTOverall (%)
Qwen-Image-Edit-251153.2962.9856.7857.6644.1645.9959.3755.3454.26
+ PhysTrace (w/o Reasoning)55.0765.1657.9656.6742.7145.3063.2353.8155.12
+ PICA100K (w/o Reasoning)55.5260.3057.6953.0441.4645.9159.1352.1253.38
PhysFlow54.4565.1958.5257.9944.9147.3164.3256.4356.32

4. Comparison between terminal-only supervision and full process supervision.

As suggested, we retain the same six recurrent latent tokens, 16-token final visual block, and trainable modules, and train the planner using only terminal visual supervision and the generation loss. With the same architecture and update budget, PhysFlow achieves 56.32% accuracy versus 55.57% for this baseline (Table 3(b)), a gain of 0.75 percentage points. These results support the benefits of textual reasoning supervision and dense intermediate visual supervision for improving physical consistency in image editing.

Training strategyLPLSERFLRFRDFMCSLGSTLSTOverall
Terminal-only supervision53.8464.4857.7359.1842.5746.6863.5955.7155.57
Full PhysFlow54.4565.1958.5257.9944.9147.3164.3256.4356.32

5. Additional analysis of general editing performance.

We value your suggestion to further investigate the category-level results more closely, particularly the declines in Hybrid and Extract despite the largely preserved overall performance on general image editing tasks.

We attribute these declines to the differences between physical-process editing and the specific requirements of these categories. Specifically, Hybrid involves combining multiple editing operations within a single instruction, while Extract requires isolating a subject while preserving its identity and visual details. In contrast, PhysTrace focuses on modeling causally connected state transitions for individual physical editing operations, guided by textual rationales and intermediate visual states (Sections 3.2–3.4). This mismatch in task focus may contribute to the observed performance degradation in Hybrid and Extract following physics-aware fine-tuning. We further observe that PhysicEdit and ChronoEdit, two other physics-aware editing methods, also exhibit relatively lower scores in Hybrid and Extract compared with other general editing categories in Table 5. We have included this analysis in Appendix C.1.

Reviewer_62MGOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents PhysFlow, a physics-aware image editing framework that integrates high-level physical reasoning, low-level transformation planning, and intermediate visual evolution into a compact sequence of continuous latent thoughts. The framework adopts a three-stage training curriculum to progressively internalize explicit textual and visual reasoning into six latent tokens, which subsequently guide a pretrained image generator. The authors additionally construct PhysTrace, a dataset of 11,448 synthetic physical transformation trajectories with hierarchical reasoning annotations and intermediate visual states.

The paper addresses a meaningful problem and presents a coherent technical framework. Experiments on PICABench and RISEBench demonstrate consistent improvements over the Qwen-Image-Edit-2511 backbone, with relatively modest inference overhead. However, the central methodological contribution is not yet fully isolated from improvements attributable to additional training data, generator adaptation, and training procedures. Stronger controlled experiments and more direct validation of the proposed physical reasoning mechanism would substantially strengthen the paper.

Critical strengths

Well-motivated research problem. The paper identifies meaningful limitations of text-only reasoning and video-based intermediate-state generation for physics-aware image editing. The proposed continuous representation provides a plausible approach to combining causal planning with visual transformation information while reducing inference overhead.

Coherent framework and training curriculum. The three-stage procedure connecting explicit supervision, progressive internalization, and generator adaptation is clearly described (Sections 3.3–3.4; Figure 2). PhysTrace also provides structured textual and visual supervision across multiple physical domains (Section 3.2; Appendix A).

Consistent empirical improvements. PhysFlow improves PICABench-Superficial accuracy from 54.26% to 56.32% and RISEBench accuracy from 8.06% to 9.44% relative to its backbone (Tables 1–2). Component ablations, alternative evaluators, and qualitative comparisons provide additional support for the method (Table 3; Appendix Tables 7–8; Figures 3 and 6). The inference latency increases by only approximately 11.4% relative to the backbone, while remaining lower than the reported explicit-reasoning and video-based alternatives (Table 5).

Critical weaknesses
  1. Insufficient isolation of the contribution of latent reasoning.

The proposed method simultaneously introduces continuous latent reasoning, a curated PhysTrace dataset, a three-stage training curriculum, and additional generator adaptation using rank-512 LoRA (Sections 3.2–3.4; Appendix B). Although Table 3 evaluates textual and visual reasoning components separately, it does not clearly establish whether the improvements persist against a non-reasoning editor trained on exactly the same PhysTrace data with identical generator adaptation, optimization budget, and trainable capacity.

This comparison is essential for distinguishing the contribution of continuous latent physical reasoning from gains obtained through additional supervision, higher-quality training data, or generator fine-tuning.

Furthermore, the dataset comparison in Table 4 and Appendix Table 9 does not fully isolate dataset quality. In particular, PICA100K-SFT uses a different supervision and training procedure, while the trajectory-based comparison uses different numbers of optimization steps. The results support the usefulness of PhysTrace but do not conclusively establish the independent contribution of its proposed supervision design.

  1. Limited validation of the proposed reasoning mechanism.

The core claim is that continuous latent thoughts internalize physically meaningful causal dependencies and intermediate visual evolution. However, several important aspects of this mechanism remain insufficiently validated.

First, the change-uniform frame sampling strategy is motivated in Appendix A.2 but is not quantitatively compared with ordinary temporal-uniform sampling. Second, the choice of four intermediate visual states and the proposed progressive compression sequence are not systematically ablated against alternative token counts or direct compression strategies.

Appendix Figure 7 provides a useful intervention study. Reversing or shuffling the intermediate visual tokens decreases physical-question accuracy by approximately 3.42 percentage points on 64 examples. Nevertheless, this result primarily demonstrates that the model is sensitive to the ordering of its learned representations. It does not establish that the latent tokens encode physically meaningful intermediate states or that their ordering corresponds to valid physical evolution.

Additional representation-level analyses and controlled interventions would provide stronger evidence for the proposed interpretation.

  1. Relatively modest performance gains and limited statistical validation.

Although PhysFlow consistently improves upon its backbone, the absolute gains remain relatively modest. For example, RISEBench accuracy increases from 8.06% to 9.44% (Table 2), corresponding to approximately five additional successful examples out of roughly 360, as inferred from the reported category-level results.

The paper does not report repeated-run variance or confidence intervals for its main benchmark comparisons. Therefore, the robustness of these improvements across training seeds and evaluation samples remains unclear.

Moreover, the central claims concern physical consistency, yet the main evaluations largely rely on multimodal-model judgments. The additional GPT-based evaluations in Appendix Tables 7–8 help address possible Qwen-family evaluator bias, but they do not independently establish objective physical correctness.

Evaluations targeting measurable physical constraints, together with difficult cases and failure analysis, would provide stronger support for the claimed improvements in physical reasoning.

  1. Originality and positioning relative to closely related approaches.

The integration of textual and visual latent reasoning for image editing is interesting. Nevertheless, several underlying mechanisms, including progressive CoT internalization, continuous latent reasoning, and visual feature prediction, have already been explored in related work (Section 2.3).

Questions for authors

Can the authors provide a controlled baseline using identical PhysTrace data, generator LoRA configuration, and training budget, but without continuous latent reasoning? How much of the reported improvement remains attributable specifically to the latent reasoning mechanism?

Can the authors provide stronger evidence that the learned intermediate visual tokens represent physically meaningful state transitions? For example, can their representations be aligned with actual intermediate states or evaluated under controlled physical interventions?

Can the authors clarify how PhysTrace samples are separated from the PICABench evaluation examples, including any source-image or near-duplicate overlap checks? Can they also provide a parameter-count and peak-memory comparison to complement the latency measurements?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 62MG

Thank you very much for your careful reading and detailed, constructive feedback. We sincerely appreciate your recognition of our motivation, framework, and empirical improvements. We address your questions and concerns in the following eleven responses.


1. Isolating the effect of additional supervision in training dataset PhysTrace.

Following your suggestion, we train a non-reasoning image editor on the same PhysTrace source–target pairs without reasoning annotations, using the same generator backbone (Qwen-Image-Edit-2511). Both methods use rank-512 generator LoRA and 537 generator updates, matching the generator adaptation configuration and budget. PhysFlow additionally trains its planner for 1,253 updates. We refer to this baseline as Qwen-Image-Edit-2511 + PhysTrace (w/o Reasoning) and report its results in Table 8.

As shown, the new baseline outperforms the original backbone by 0.86 percentage points. Furthermore, our latent reasoning framework further improves the overall score from 55.12% to 56.32% over this non-reasoning baseline, demonstrating the effectiveness of our proposed approach. Particularly, under the same source-target image pair setting, training on our curated PhysTrace achieves higher overall performance than training on PICA100K (i.e., Qwen-Image-Edit-2511 + PICA100K (w/o Reasoning)), highlighting the benefits of its superior data quality for training efficiency.

ConfigurationLPLSERFLRFRDFMCSLGSTLSTOverall (%)
Qwen-Image-Edit-251153.2962.9856.7857.6644.1645.9959.3755.3454.26
+ PhysTrace (w/o Reasoning)55.0765.1657.9656.6742.7145.3063.2353.8155.12
+ PICA100K (w/o Reasoning)55.5260.3057.6953.0441.4645.9159.1352.1253.38
PhysFlow54.4565.1958.5257.9944.9147.3164.3256.4356.32

2. Comparison between change-uniform and time-uniform sampling.

As suggested, we have compared our proposed change-uniform frame sampling strategy with conventional temporal-uniform sampling on PICABench-Superficial. The results are reported in Table 3(d). As shown, our change-uniform strategy improves the overall score by 1.27 points over temporal-uniform sampling. This improvement can be attributed to its ability to capture richer transition dynamics, whereas temporal-uniform sampling may concentrate substantial visual changes between only a few consecutive sampled frames.

SamplingLPLSERFLRFRDFMCSLGSTLSTOverall
Temporal-uniform53.7263.0659.3057.6443.2946.2162.3553.6555.05
Change-uniform54.4565.1958.5257.9944.9147.3164.3256.4356.32

3. Ablation study on progressive compression with different token counts.

As suggested, we evaluate our proposed progressive compression mechanism using different numbers of tokens per intermediate visual state (1, 2, and 4). The results are reported in Table 9 of Appendix C.7. As shown, using a single token per intermediate visual state achieves the highest overall accuracy (56.32%), and is adopted as the default setting in our paper.

Tokens per stateLPLSERFLRFRDFMCSLGSTLSTOverall
154.4565.1958.5257.9944.9147.3164.3256.4356.32
252.1064.2857.6156.1642.6146.0862.9453.9254.73
452.6564.4156.9356.1641.8347.5463.4052.8454.85

4. Comparison between progressive and direct compression.

As suggested, we additionally compare our proposed progressive compression mechanism with a direct compression strategy that replaces all textual and visual reasoning blocks with latent tokens in a single step. The results are added to Table 3(c). As shown, our progressive compression improves the overall accuracy by 1.80 points over direct compression. This improvement suggests that progressively compressing reasoning blocks allows the model to gradually distill multimodal reasoning information into compact latent representations, while preserving the causal dependencies across successive reasoning steps. In contrast, directly compressing all reasoning blocks at once may lead to greater information loss and more challenging optimization.

Compression strategyLPLSERFLRFRDFMCSLGSTLSTOverall
Direct compression52.8463.8355.6856.3142.5446.8562.4053.9654.52
Progressive compression (paper)54.4565.1958.5257.9944.9147.3164.3256.4356.32
AuthorsOpenReview ↗

Response to Reviewer 62MG

5. Additional representation analysis and controlled interventions for better interpretability.

We thank the reviewer for highlighting the importance of analyzing the learned latent representations for better interpretability. As suggested, we conducted two additional analyses in Appendix C.8 of the revised manuscript: (a) Token-role visualization, and (b) Physical intervention analysis.

(a) Token-role visualization. We visualize latent tokens from PhysTrace test cases across four reasoning roles: high-level textual reasoning, low-level textual reasoning, intermediate visual states, and the final visual state (Figure 7(a)). For each input, we average hidden states within each role and normalize the resulting vectors to fit a shared PCA basis. We then normalize individual token vectors and project them onto the first two principal components using the same PCA basis. As shown in Figure 7(a), the compressed tokens form a well-structured latent space, with distinct clusters corresponding to different reasoning roles.

(b) Physical intervention analysis. We investigate whether intermediate visual latents encode the physical consequences of different interventions, with results reported in Figure 7(b). We construct synthetic scenes across seven tasks (support, torque, buoyancy, outflow, absorption, phase change, and shape recovery), each with two instructions yielding opposite outcomes under idealized physical rules. For each input, we extract four intermediate visual tokens from frozen PhysFlow and train a task-specific linear ridge probe on their averaged features to predict binary outcomes. For each task, we use 256 training pairs, 64 validation pairs, and 128 test pairs. A pair is considered correct only if both intervention outcomes are predicted correctly. As shown in Figure 7(b), the overall pair accuracy across the seven tasks reaches 97.88%, indicating that the learned visual latents encode discriminative information about the physical consequences of different interventions.

We have added the detailed experimental settings and controls to Appendix C.8.


6. Repeated-run variance on PICABench-Superficial and RISEBench.

As suggested, we separately report the mean ± standard deviation (SD) across training and inference seeds.

We repeat PhysFlow training with three different seeds while fixing the inference seed at 42. PhysFlow achieves 56.22 ± 0.37% on PICABench-Superficial and 10.07 ± 1.21% on RISEBench, with mean gains of 1.96 and 2.01 percentage points over the original backbone, respectively.

MethodPICABench-Superficial Mean (%)SD (%)RISEBench Mean (%)SD (%)
PhysFlow56.220.3710.071.21

We also evaluate fixed checkpoints using three different seeds. PhysFlow outperforms the original backbone on both PICABench-Superficial and RISEBench, with a mean improvement of 1.92 and 1.29 percentage points, respectively.

MethodPICABench-Superficial Mean (%)SD (%)RISEBench Mean (%)SD (%)
Qwen-Image-Edit-251154.410.348.710.58
PhysFlow56.330.4010.000.96

7. Evaluation Using an MLLM as a Judge for Physical Consistency.

PICABench evaluates sample-specific physical criteria through human-annotated questions and reference answers, such as whether object removal also removes its shadow. PICAEval uses an MLLM to answer these questions and scores agreement with the references (PICABench, Section 3.3 and Appendix A.3).

Our human study complements this evaluation: 17 participants compare PhysFlow with five methods on 100 randomly sampled PICABench and RISEBench pairs. Their physical-plausibility preferences agree with the benchmark trends. We acknowledge the value of measurable physical constraints and leave their evaluation to future work.


8. Failure Analysis.

As suggested, we inspected PICABench outputs to analyze failure modes and capability boundaries. Other methods sometimes struggle with basic physical relationships involving object pose, support, and liquid motion. PhysFlow better captures these main changes in the inspected comparisons, but coordinating coupled physical processes remains challenging, as each change can trigger further effects that must remain consistent with one another.

For example, tipping a coffee cup onto a fabric cushion requires the cup to fall onto its side, coffee to spill and spread, and the fabric to absorb the liquid. Step1X introduces a second cup and a pouring hand, changing the requested action. ChronoEdit depicts the cup suspended above the cushion. In contrast, PhysFlow shows the cup resting on its side, with a continuous coffee spill connecting its opening to the puddle on the cushion. This result more coherently captures the cup's motion, contact with the cushion, and liquid spillage. However, the puddle retains a sharp boundary, with little evidence of coffee soaking into the surrounding fabric.

AuthorsOpenReview ↗

Response to Reviewer 62MG

9. Originality and positioning relative to closely related approaches.

We acknowledge that the ideas of progressive CoT internalization, continuous latent reasoning, and visual feature prediction have been explored in prior work (Section 2.3). However, effectively adapting these techniques to physics-aware image editing and achieving substantial performance improvements is non-trivial. The key contributions of our PhysFlow are summarized as follows.

(a) Our framework internalizes both textual and visual reasoning into a unified latent space. Existing latent reasoning methods, such as Coconut, CODI, and SIM-CoT, internalize textual reasoning without explicitly modeling the corresponding visual states. Methods such as LVR, Monet, and PhysicEdit combine explicit textual reasoning with visual latent representations for textual answer generation (LVR and Monet) or image editing (PhysicEdit). In contrast, our PhysFlow internalizes both textual explanations and visual-state transitions into a unified multimodal latent space, enabling joint textual and visual reasoning without generating explicit reasoning text or rendering intermediate visual states at inference.

(b) Internalizing multimodal reasoning for image editing requires a non-trivial training strategy. Existing methods learn latent representations through feature alignment or distillation (CODI, PhysicEdit, and Monet) or diffusion-based supervision of visual states (EndoCoT). In contrast, our PhysFlow introduces a progressive compression mechanism tailored to physics-aware image editing, jointly internalizing textual explanations and visual-state transitions by replacing their reasoning blocks with compact latent tokens, one block at a time. These latent tokens receive indirect supervision from text prediction and visual alignment losses on subsequent uncompressed reasoning blocks, together with the diffusion loss for the final edited image, coupling multimodal reasoning compression with the image editing objective.

(c) We curated a high-quality dataset enabling larger performance gains with less training data. Most existing image editing datasets, such as PICA100K, provide only source–target image pairs without intermediate visual supervision. Although PhysicTran38K introduces video trajectories, we observe physical inconsistencies in videos generated by its Wan2.2 backbone. To provide physically grounded process supervision, we use Cosmos 3 Super to generate transition trajectories, filter them for instruction compliance and physical consistency, and annotate them with hierarchical textual reasoning and change-uniformly sampled intermediate frames. Our experiments show that PhysFlow trained on our PhysTrace achieves larger performance gains with fewer training examples than PhysFlow trained on PhysicTran38K, supporting more data-efficient training.


10. Data overlap checks between PhysTrace and PICABench.

We thank the reviewer for raising the question of training and evaluation data separation. As described in Sections 3.2 and 3.4 of the PICABench paper, PICABench uses natural images collected from licensed and public sources, whereas PICA-100K uses synthetically generated source images. As described in Section 3.2 and Appendix A, we construct PhysTrace using image-instruction pairs from PICA-100K, rather than PICABench, ensuring that no PICABench evaluation examples are directly included in our training data.


11. Parameter-count and peak-memory comparison.

As suggested, we have added these measurements to Table 4 to provide a more comprehensive assessment of inference costs.

As shown, PhysFlow maintains nearly the same parameter count and peak memory usage as Qwen-Image-Edit (28.87B vs. 28.85B parameters and 58.11 vs. 58.06 GiB). Despite having a parameter count comparable to PhysicEdit and larger than those of Step1X-Edit and ChronoEdit, PhysFlow achieves lower inference latency than all three methods. Moreover, PhysFlow requires only 49.39 seconds per edit, compared with 60.12 seconds for Qwen-Image-Edit with explicit reasoning. These results highlight the inference efficiency of our compact latent planning design.

MethodParams (B)Memory (GiB)Latency (s)
Qwen-Image-Edit28.8558.0644.34
PhysicEdit29.0056.2552.19
Step1X-Edit20.8241.3459.18
Qwen-Image-Edit + explicit reasoning28.8558.0660.12
ChronoEdit22.8354.81228.68
PhysFlow (Ours)28.8758.1149.39
Reviewer_8uwkOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes using additional supervision from text and video to train a compact latent planner to improve the physical understanding of current image editing models. This paper also curates a new dataset of synthetic transformation trajectories with reasoning annotations and intermediate frames. The proposed method improves performance on both PICABench and RISEBench without degrading general editing performance too much.

Critical strengths
  • I like the motivation behind using both textual explanation and visual trajectory since they are complementary. Text is good at representing "why it will change," while visual state is good at representing "what the change looks like". Besides, this paper considers the latency at deployment and only brings about 11% increase at inference time.

  • The experiments are very comprehensive. This paper presents not only the main comparison on PICABench and RISEBench, but also multiple ablations to validate the effectiveness of each design choice.

Critical weaknesses
  • The main issue is that the core idea of this paper, i.e., curating synthetic video trajectory and textual explanation and encoding them into an implicit visual representation to improve the physical understanding of image editing models, is highly similar to the PhysicEdit [1] paper. However, this paper doesn't include any discussion about the similarity of the high-level idea or comparisons between these papers.

[1] Zhao, Liangbing, et al. "From Statics to Dynamics: Physics-Aware Image Editing with Latent Transition Priors." arXiv preprint arXiv:2602.21778 (2026).

Questions for authors
  • Please refer to the weakness section.
1 comment
AuthorsOpenReview ↗

Response to Reviewer 8uwk

We sincerely thank you for your time, careful review, and constructive feedback. We appreciate your recognition of our motivation, experiments, and inference efficiency. Your suggestion to discuss and compare with PhysicEdit has helped us clarify our contributions and strengthen our evaluation.

Comparison with PhysicEdit

As suggested, we have included a discussion of PhysicEdit in Section 2.3, “Latent Reasoning”, and Summary in Related Work.

Although both our proposed PhysFlow and PhysicEdit leverage textual explanations and visual trajectories to enhance the physical understanding of image editing models, they differ fundamentally in their underlying mechanisms, as detailed below.

1. Our textual reasoning representation is different. PhysicEdit generates textual reasoning traces in the explicit semantic space of the MLLM, while separately modeling visual representations in the latent space. In contrast, our PhysFlow jointly performs textual and visual reasoning within a unified latent space. This formulation allows latent textual reasoning to guide visual-state prediction without generating explicit reasoning text at inference.

2. Our visual reasoning explicitly models temporal dependencies. PhysicEdit jointly computes a set of transition-query representations, whereas PhysFlow autoregressively predicts successive visual-state representations conditioned on preceding latent textual explanations and visual states. This causal formulation explicitly captures the temporal dependencies underlying physical transitions.

3. Learning unified latent reasoning requires a non-trivial training strategy. PhysicEdit trains its transition queries through feature alignment with projected DINOv2 and VAE features extracted from intermediate video frames. In contrast, our PhysFlow adopts a progressive training curriculum. It first supervises the prediction of all textual explanations and intermediate and final visual states using ground-truth text tokens and visual features extracted by the MLLM's visual encoder, respectively. It then progressively replaces the reasoning blocks for textual explanations and intermediate visual transitions with model-generated predictions, one at a time. Each replaced block receives indirect supervision through the original prediction loss applied to subsequent unreplaced positions, jointly with the diffusion loss. This stage encourages the model to learn causal dependencies across successive multimodal reasoning steps. Finally, the image generator is adapted to the learned latent representations using the diffusion loss.

We have added a quantitative comparison with PhysicEdit on PICABench-Superficial and ImgEdit-Bench in Table 1 and 5, respectively. Using the same Qwen-Image-Edit-2509 backbone, our PhysFlow achieves an overall score of 54.12% on PICABench-Superficial, outperforming PhysicEdit (52.35%) by 1.77 percentage points, with higher scores in seven out of eight categories. Moreover, our PhysFlow achieves an overall score of 4.31 on ImgEdit-Bench, compared with 4.05 for PhysicEdit, indicating stronger generalization to general image editing tasks.

MethodLPLSERFLRFRDFMCSLGSTLSTOverall
PhysicEdit (2509)51.6261.6351.5652.9644.0345.8957.6453.0752.35
PhysFlow (2509)52.0162.8257.0253.1443.3246.8161.4253.1654.12

 

MethodAddAdjustExtractReplaceRemoveBackGStyleHybridActionOverall
PhysicEdit (2509)4.314.172.344.454.444.034.743.364.854.05
PhysFlow (2509)4.454.253.544.694.484.214.773.384.824.31

We have also added a comparison of inference latency between our PhysFlow and PhysicEdit in Table 4.

MethodLatency (s)
PhysicEdit52.19
PhysFlow (Ours)49.39

Comments on the paper

2 comments
AuthorsOpenReview ↗

Revised Draft

Dear Reviewers and Area Chairs,

Thank you very much for the time and care you have devoted to reviewing our work. We sincerely appreciate your constructive comments and suggestions, which have helped us improve our manuscript. We have carefully considered every comment and suggestion and revised the paper accordingly, with the main changes marked in blue.

We summarize the main revisions below.

In the main paper:

  1. Expanded the discussion of PhysicEdit, EndoCoT, and Monet to clarify how our PhysFlow differs from these approaches.

  2. Added PhysicEdit and Qwen-Image-Edit-2509 results on PICABench-Superficial.

  3. Added ablation studies comparing full textual and visual process supervision with terminal-only visual supervision, progressive compression with direct compression, and change-uniform sampling with time-uniform sampling.

  4. Added parameter-count and memory measurements alongside inference-latency comparisons.

In the appendix:

  1. Clarified the sources and roles of PICA100K training data and PICABench evaluation examples.

  2. Added comparisons to evaluate the contributions of the PhysTrace dataset and the PhysFlow framework, including fine-tuning on the same PhysTrace data with source–target image pairs.

  3. Added PhysicEdit and Qwen-Image-Edit-2509 results on ImgEdit-Bench.

  4. Expanded the discussion of general editing performance, including the declines in Hybrid and Extract.

  5. Added an ablation on the number of visual tokens per intermediate state.

  6. Added representation analyses including token-role visualization and physical intervention.

We provide detailed responses to each reviewer's comments below. Thank you very much again for your time and helpful suggestions.

Best regards,
The Authors

AuthorsOpenReview ↗

A Gentle Follow-up on Our Rebuttal

Dear Reviewers,

Thank you again for your time and thoughtful reviews. We hope our responses have addressed your concerns. As the discussion period draws to a close, we would be grateful for any feedback you may have when your schedule allows. If any questions or concerns remain, we would be happy to discuss them further.

We sincerely appreciate your time and consideration.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.