acceptodds
Under review as a conference paper at ICLR 2027

CamWAM: A World-Action Model for Autonomous Photography

Abstract

Autonomous photography aims to enable embodied agents to capture visually appealing photographs without human camera operation, yet remains challenging due to the subjective nature of photographic objectives. Existing approaches typically determine camera actions without explicitly modeling their resulting visual outcomes, limiting their ability to reason about how camera motion affects photographic quality. In this work, we propose **CamWAM**, a world-action model for autonomous photography that jointly models future visual evolution and continuous executable camera motions. CamWAM cascades a video expert with an action expert through a lightweight adapter, allowing predicted visual evolution to provide informative representations for camera motion prediction. To effectively couple the two experts, we develop a three-stage video-action learning strategy that progressively aligns visual evolution with camera motion. We further introduce a scalable data pipeline that reverses videos synthesized from curated high-quality photographs to construct camera trajectories toward desirable views, while using independently constructed 3D Gaussian Splatting environments for cross-domain evaluation. Extensive experiments demonstrate the effectiveness of CamWAM and its strong generalization across diverse scenarios. Our models, training code, and dataset will be publicly available.

Primary area:
applications to robotics, autonomy, planning
Keywords:
World Action Models, Autonomous Photography, Embodied AI

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_NTgLOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper applies the world-action modeling paradigm to autonomous photography. CamWAM uses a pretrained video model to predict future visual evolution, and an action expert attends to the video features to predict 7-D camera actions. A lightweight adapter is introduced in a third training stage to make the action predictor more robust to imperfect video predictions. The training data are constructed by generating videos that move away from curated high-quality portraits, reversing them, and recovering camera trajectories with VGGT-Ω. Evaluation is performed in closed loop on an independently constructed 3DGS benchmark with 288 cases. Final images are scored by two VLM judges, whose agreement with human raters is also evaluated. The 14B model achieves the strongest overall results among the compared methods.

Critical strengths
  1. The data construction pipeline is clever and practically useful. Starting from a desirable photograph and reversing a generated trajectory provides supervision that ends at a good photographic viewpoint without requiring manually collected camera demonstrations.
  2. The closed-loop benchmark is a meaningful part of the paper. It is constructed independently from the video-generation pipeline used for training, so the model is evaluated in interactive 3DGS environments rather than on the same type of synthetic videos used for supervision.
  3. The evaluation is relatively transparent. The paper validates the VLM judges against human raters and reports per-judge results, scene-level results, paired bootstrap intervals, and qualitative failure cases. This makes it easier to understand where the reported gains come from.
Critical weaknesses
  1. The benefit of future prediction is not cleanly proved. The key evidence is Action-only vs. Joint at 1.3B, but it is unclear whether the visual backbone is trained in the same way in the two settings. If not, the gain may partly come from backbone adaptation rather than access to future visual information. AVC also does not track downstream performance consistently across model scales.
  2. The necessity of Stage 3 is unclear. Its motivation is the train-test gap between clean video features during training and predicted features at inference, but the paper does not compare against the simpler baseline of adding the same conditioning noise when training the action expert, without an adapter. The Stage-3 gain is also small at 1.3B and 5B.
  3. The benchmark may not fully establish the need for a world-action model. The photography instructions are fairly structured, and many cases may be solvable by iterative geometric correction based on subject position and scale. A simple detection-and-servoing baseline would help clarify how much of the gain actually comes from future visual prediction rather than standard closed-loop visual control.
Questions for authors
  1. Is the Wan backbone frozen or fine-tuned in Action-only? Could you provide a matched control using the same trained video expert, but letting the action expert attend only to the current/first-frame features?
  2. How does Stage 3 compare with simply adding the same conditioning-noise augmentation during Stage 2, without the adapter?
  3. How is the normalized translation recovered from VGGT-Ω mapped to metric motion in the 3DGS scenes? How were the execution scale and stopping rule chosen?
  4. The 288 cases are built from only 24 location-pose configurations. Do the Table 7 conclusions still hold if bootstrap resampling is done at the pose level rather than the individual-case level?

I am glad to raise my score if the authors can clearly demonstrate the benefits of future prediction and the adapter design.

Reviewer_Py9ROpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CamWAM is a world-action model for autonomous photography that predicts future visual evolution and continuous camera motion using a video expert, an action expert, and an adapter that aligns their representations. The paper also builds a data pipeline that reverses videos synthesized from curated photographs into camera trajectories and creates an independent 3D Gaussian Splatting benchmark for evaluation. It compares several model sizes with composition, camera-motion, VLM-agent, and VLA baselines, using automated and human ratings of photographic quality and instruction following.

Critical strengths
  1. Jointly predicting visual outcomes and camera actions addresses a limitation of camera policies that optimize actions without modeling their resulting views. The staged training strategy provides a plausible way to make video features useful for camera control.
  2. The 3DGS evaluation is more informative than scoring predicted camera trajectories alone, and includes comparisons with multiple baseline families as well as human scoring. This makes the work a useful task-specific study of autonomous composition.
Critical weaknesses
  1. The training signal is derived from curated photographs and model-synthesized videos that are reversed to produce trajectories toward desirable views. This creates a potential distribution gap between synthetic reverse trajectories and genuine camera operation, and may encode the curation's aesthetic preferences. The independent 3DGS environments help test transfer, but do not establish performance with physical cameras or real subjects in unconstrained environments.
  2. The central architectural claim—that visual prediction improves action selection—is not isolated as the system comparison suggests. The paper would benefit from matched-capacity ablations separating video-feature conditioning, staged alignment, and data scale, as well as stronger validation of the automated evaluators against expert photography judgments. Other methods (e.g., GenDoP[1], PhotoFramer [2]) and general VLA systems are relevant, but these comparisons on a custom 3DGS benchmark may not establish broader generality.

[1] Zhang, Mengchen, et al. "Gendop: Auto-regressive camera trajectory generation as a director of photography." 2025 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2025.

[2] You, Zhiyuan, et al. "Photoframer: Multi-modal image composition instruction." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2026.

Questions for authors
  1. How much of the gain remains when comparing against an action expert with the same training data, parameter budget, and optimization schedule but without video-prediction features or the alignment adapter?
  2. Were the human raters and VLM evaluators blind to model identity, and how were disagreements between aesthetic preference and instruction compliance handled? Does the result persist on real camera captures or a benchmark whose scenes and subject assets are disjoint from data construction?
Reviewer_kLR9OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies instruction-guided autonomous photography, addressing limited camera-motion supervision and the challenge of translating desired visual outcomes into actions. CamWAM constructs training data by reversing videos generated from curated photographs and recovering camera trajectories. It combines video and action experts through a motion adapter and three-stage training. Experiments in independently constructed 3DGS environments show improved instruction following and composition quality, with the largest model outperforming the evaluated baselines. Ablations support the benefits of future visual prediction and staged training.

Critical strengths
  • The method connects future visual prediction with camera control through a well-motivated three-stage training procedure.

  • The reversed-video data pipeline provides a practical way to obtain paired video and camera-motion supervision.

  • Evaluation includes an independent 3DGS benchmark, diverse baselines, ablation studies, and human validation of the evaluators.

  • The paper is clearly organized, with useful implementation details and additional analyses in the supplementary material.

Critical weaknesses
  • Limited validation in real capture settings. The closed-loop results are obtained by executing camera actions in reconstructed 3DGS environments. Although Figure 13 shows predictions starting from real photographs, it does not demonstrate the outcomes of executing those actions with a real camera. A small set of real-camera experiments would help establish whether the predicted motions produce the intended framing and compositional improvements in practice.

  • Limited environmental diversity and robustness analysis. The benchmark covers six locations in two outdoor environments. Performance under indoor scenes, daylight versus low-light conditions, different texture levels, and varying scene scales is not systematically examined. Such changes could affect the visual cues used to infer camera motion. Controlled evaluations would help clarify their influence on predicted translation, orientation, and field-of-view adjustments, as well as the resulting photographic quality.

  • Incomplete evaluation of subject fidelity in the final photographs. Improved composition alone does not guarantee a usable photograph if the subject exhibits unnatural distortion or loses important appearance details. The current evaluation emphasizes composition and instruction following, while explicitly excluding several appearance and rendering defects. An additional subject-preservation metric or human assessment of the final photographs, accounting for legitimate changes in viewpoint, scale, and framing, would provide a more complete evaluation of photographic quality.

  • Additional video demonstrations. As a minor suggestion, the authors could share predicted and executed videos on a project webpage in a future release. These would complement the sampled frames and help readers assess temporal consistency, camera-motion smoothness, and prediction–execution alignment.

Questions for authors
  • Have you tested the predicted camera motions in real capture settings, and do the resulting photographs show similar compositional improvements?

  • How robust is camera control across indoor scenes, different lighting conditions, texture levels, and scene scales?

  • Could you assess subject fidelity in the final photographs, accounting for legitimate viewpoint and framing changes, to complement the current composition-focused evaluation?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.