- 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
- 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.
- 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.
- 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
- 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.
- 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.
- 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
- 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?
- How does Stage 3 compare with simply adding the same conditioning-noise augmentation during Stage 2, without the adapter?
- 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?
- 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.