acceptodds
Withdrawn from ICLR 2027

EZCAM: Data, Infrastructure, and Policy for Embodied Photography

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Embodied photography asks an agent to move a camera, observe the resulting view, and retain a photograph under a limited interaction budget. Studying this capability requires both action-linked supervision and an environment that exposes the consequences of the agent's own decisions. We present EZCAM, connecting photographic trajectories, interactive camera infrastructure, and a learned reference policy. The corpus contains 96,700 video–pose trajectories from games and game-derived reconstructions, with 347,111 auxiliary Internet clips indexed separately. Policy training also incorporates Internet video. Replay audits distinguish improving endpoints from non-monotonic local motion, and a blinded study supports selected data-image pairs. The infrastructure provides 49 scene assets with explicit camera conventions, complete-path checks, and observation accounting. We evaluate seven trained configurations across all 49 scenes, using six starts per scene and 40 action requests per episode. The learned policy achieves the highest mean best-observed raw score gain of 0.5126, compared with 0.2817 for the strongest supplied visual-feature adapter. Its mean endpoint gain is negative, illustrating why retained-image quality and final-pose quality must be distinguished. These results establish a measurable data–interaction–policy pipeline and characterize the trained-model comparison. The evaluation reports training, development, and unseen scenes separately. Controlled ablations measure the effects of branch supervision and recovery; independent policy-output preference remains unverified.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Embodied Photography, Benchmark

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

The paper introduces EZCAM, an integrated resource for embodied photography in which an agent moves a camera and retains an image under a fixed observation budget. It combines 96,700 game and game-derived video–pose trajectories, separately indexed Internet clips, 49 reconstructed scene assets, a camera execution interface with complete-path validity checks, and the Branch-Supervised Camera Policy (BSCP). BSCP predicts camera-action chunks from RGB, executed-action history, and internal score features. Seven trained configurations are compared on 294 shared starts with 40 requests per episode. BSCP obtains mean best-observed aesthetic-score gain of 0.5126 overall and 0.3985 on 36 unseen scenes, but its mean endpoint gain is negative. A separate 37-participant study supports the scorer's ranking of selected data images; it does not evaluate photographs selected by the learned policy.

Critical strengths

The paper’s strongest contribution is a substantial resource that connects photographic demonstrations with executable camera interaction: 96,700 pose-associated trajectories and 49 scene assets support learning and evaluating how camera movement changes composition. Complete-path validity checks and explicit accounting for observations, rejected requests, and executed actions make the interaction budget interpretable and provide a useful foundation for reproducible comparisons. A second strength is the evidence that the learned policy transfers beyond its adaptation scenes under the specified aesthetic-score objective. On the 36 unseen evaluation scenes, BSCP achieves a mean best-observed score gain of 0.3985, compared with 0.2054 for the strongest evaluated adapter, StreamVLN. Data-source ablations additionally show benefits from combining Internet and game/3DGS supervision. Together, these results give the resource practical value: it supports both a concrete experimental protocol and a learned reference policy with measurable gains.

Critical weaknesses

The experimental analysis has two limitations that leave some uncertainty about robustness and attribution. First, the policy comparisons and ablations report point estimates without confidence intervals for paired performance differences or variability across training runs. Although the shared evaluation starts provide a useful control, uncertainty estimates would help assess the consistency of the rankings and component gains across scenes and optimization runs. Second, the data-source ablation groups game-engine and game-derived 3DGS trajectories into a single bundle (Table 2). This establishes their combined benefit alongside Internet data, but does not isolate the contribution of each source. Separate game-only and 3DGS-only comparisons, with training updates and data volume controlled, would better establish whether these sources provide complementary supervision and guide future data collection.

Questions for authors

How consistent is BSCP’s advantage over the strongest adapter across the 36 unseen scenes? Using the existing evaluation records, could you provide a paired uncertainty estimate for the mean improvement and clarify whether the results represent a single training run? This would help assess how reliably the aggregate gain generalizes across scenes.

Reviewer_BQ7FOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper tries to tackle a task where a photographer policy is dropped in a static 3D environment and has 40 moves worth of budget to move around, taking pictures, getting rewards based on aesthetic quality of the pictures it takes.They collect data from gaming engines and real world, construct a proxy aesthetic scorer, and train their policy which is a flow matching model using this data and scorer used as reward.

Critical strengths

None

Critical weaknesses

The most critical weakness of the submission is the presentation and quality of writing. The manuscript is obviously highly written by LLMs, which is not an issue on its own, but currently has made reading the paper extremely hard. Starting from the abstract which contains weird sentences like "independent policy-output preference remains unverified" to all parts of their manuscript, like using abbreviations like "BSCP" in line 74 before introducing it, to over emphasis on trivial software engineering checks in section 5 and 7.1 (nearly two pages!); all in all the tone of the text reads more like a technical audit/report that an LLM would generate rather than a scientific paper which tries to communicate the problem formulation, data, method, experiments, and findings in a well-organized manner. Even though there seems to be substantial technical work done behind the submission, with significant 3DGS data processing from game engines, human evaluations of the proxy aesthetic scorer, and training the policy, the very bad presentation makes it extremely difficult to understand the work.

Other than that, the authors also mention that their endpoint gain is negative, while best gain is positive, which they attribute to agent exploration even after finding a good capture. Then why not also training the policy to predict a "stop" action after finding a good shot? This would help with endpoint gain being negative, and would also be an interesting contribution and feature for the final photographing agent.

Questions for authors

I think the manuscript needs a whole re-write and restructure to be presentable at a top tier conference like ICLR (and honestly, any other respectable venue!)

The bad quality of writing prevented me from understanding all the methodology details, so I have only the following question below:

Repearing the second weakness above, could you please train the policy to predict its ending move? It's fine (and probably preferrable) to terminate earlier than 40 steps if the policy found a good spot to take the picture. It would be appreciated to train the policy in this way and report the endpoint gain in aesthetic proxy score after this improvement.

Reviewer_B2tzOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces EZCAM, a framework for embodied photography combining a dataset, an interactive camera environment, and a learned Branch-Supervised Camera Policy (BSCP). The dataset contains 96,700 game/game-derived trajectories and 347,111 auxiliary Internet clips, while the environment provides 49 reconstructed scenes with explicit execution and observation accounting. BSCP achieves the best proxy-score performance among seven trained models, including a mean best-observed gain of 0.5126.

Critical strengths
  1. Well-integrated data and evaluation infrastructure. The paper carefully defines action budgets, complete-path safety checks, rejected-action accounting, and reproducible episode records. This provides a useful foundation for studying embodied camera control.

  2. Broad and transparent evaluation. The comparison uses identical starts across 49 scenes and reports seen, development, and 36 unseen scenes separately. The human study with 37 participants also provides some support for the aesthetic scorer.

Critical weaknesses
  1. The main metric does not establish autonomous photograph selection. (G_{\text{best}}) uses the highest-scoring frame among all observations, effectively relying on an oracle selector. The policy does not demonstrate selecting or returning to that frame, and its mean endpoint gain is negative (-0.3653). Therefore, the claimed ability to “retain a photograph” is not fully validated.

  2. Insufficient qualitative evidence for trajectory quality. The main paper primarily shows static best-observed frames and proxy-score curves, without sufficient video demonstrations of complete policy rollouts. This is important because the paper reports non-monotonic local motion and a negative mean endpoint gain (-0.3653). Static selected frames cannot reveal whether the camera motion is smooth, stable, or practically usable. More full-episode videos across diverse scenes and failure cases would strengthen the evaluation.

Questions for authors
  1. Can the authors report a blinded human preference comparison between photographs found by BSCP and the strongest baseline?

  2. Does BSCP contain an executable online frame selector? If so, what is its (G_{\text{sel}})? If not, how is the retained photograph chosen without oracle access to all scores?

  3. What are the parameter counts, training updates, and computational budgets of BSCP and the seven adapters? Do the gains remain under a matched-budget comparison?

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

Related papers

Discussion (0)

Sign in to comment.