acceptodds
Withdrawn from ICLR 2027

Rethinking Feed-Forward Visual Prediction with Meta-Learned Adaptation Priors

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Feed-forward networks efficiently predict geometry, cameras, and appearance, but can lose accuracy under domain shift. Standard supervised training optimizes prediction without explicitly preparing models to benefit from limited target-domain supervision. We introduce MetaFeed, a framework that learns adaptation priors for supervised few-shot domain calibration. Through first-order meta-learning, MetaFeed optimizes a shared initialization for prediction after support-set updates to a task-dependent subset of existing weights. Geometry tasks adapt encoder attention projections and feed-forward layers, whereas appearance tasks adapt decoder convolutions. Calibration uses a small labeled support set once per target domain, after which the model processes held-out scenes with fixed weights and ordinary feed-forward inference. Controlled comparisons with supervised training and target fine-tuning under matched data and adaptation protocols show improved cross-domain prediction with competitive in-domain performance. Multi-source controls further show that the calibrated meta-trained model outperforms matched target fine-tuning on a domain unseen during source training. Experiments with VGGT reveal two distinct gains in reconstruction quality: meta-training reduces Chamfer distance before calibration relative to matched supervised training, and supervised calibration using a single support episode from a different scene within the same target domain further reduces it on disjoint target scenes.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Meta-learning, Few-shot domain adaptation, Dense visual prediction, Feed-forward models, 3D reconstruction, Inverse rendering.

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_6SRwOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper present MetaFeed that aims on supervised few-shot domain calibration for dense geometry and appearance prediction. MetaFeed utilizes first-order meta-learning that optimizes an initialization for support-set updates to selected existing parameters for multiview feed-forward methods, and the paper updates the encoder attention/FFN weights for geometry and decoder convolutions for appearance in the pipeline. Calibration is performed once per target domain, followed by fixed-weight inference on held-out scenes. The authors provide extensive experiments to validate the effectiveness of the method and adaptation studies to validate the core components. The supplementary video additionally illustrates geometry and appearance improvements during adaptation.

Critical strengths
  1. The proposed meta-learning framework offers a practical deployment strategy with unchanged inference architecture of current feed-forward multiview models. Calibration updates selected existing parameters once per target domain, after which the model uses fixed weights on held-out scenes. The authors also provide detailed results of training and calibration costs in the paper, which helps assess when this additional preparation may be worthwhile.

  2. The results demonstrate clear improvements, which also persist beyond the single-source training setting. The multi-source experiment retains a substantial advantage on unseen datasets and strengthens the contribution that the method offers more than compensation for narrowly trained source models.

  3. The paper provide detailed ablations. The ablations distinguish initialization gains from additional calibration gains, providing more informative evidence than a comparison between adapted MetaFeed and an unadapted baseline alone.

Critical weaknesses
  1. The ablation section of the paper can be improved. The appearance experiments do not separate improved initialization from improved adaptation. Table 1 and Table 5 omit the quantitative uncalibrated MetaFeed results, while Figure 6(a) uses Static-Sup as its zero-step reference. Consequently, the appearance gains cannot be attributed to the support update as clearly as the geometry gains in Table 3. I suggest clarifying this in the paper.

  2. The comparison does not establish that ordinary fine-tuning receives an independently suitable adaptation configuration. FT and MetaFeed share learning rates, whereas the sensitivity study primarily evaluates MetaFeed; the deterioration of InteriorVerse normal error from 3.77° to 14.98° after FT-1 makes baseline-specific tuning important for interpreting the reported advantage. As a method that includes meta-learning the pipeline, the comparison of direct finetuning and the proposed method is important and should be precise. I suggest the authors further discuss this in the paper.

  3. There are some ambiguous presentation in the paper: a) The support-scene definition is ambiguous. In Appendix C.1, the paper defines KK as distinct support scenes, yet Table 11 reports K=10K=10 and K=20K=20 on 7-Scenes, the sampling unit and actual scene/sequence separation therefore require clarification. b) The noisy-support result differs between the text and table. In Appendix E.3, the paper reports CD 0.185 at noise level 0.10, whereas Table 9 reports 0.182 for that setting. Although it does not change the qualitative conclusion, this should be corrected for better presentation.

Questions for authors

During the comparison experiments, are FT and LoRA allowed their own learning-rate and step-count selection using the same validation budget as MetaFeed? Most importantly, does the advantage persist under independently selected adaptation configurations? From the perspective of training feed-forward models, the training dataset ratio can affect the final performance of the method, so from my perspective, providing precise comparison to demonstrate the effectiveness of the paper is very important. Since this year does not have "Borderline" option, I vote for the current score first and I'm more than willing to see the authors provide stronger evidence either by detailing the comparison settings or more experiments on different datasets during the rebuttal section, and I would increase the ratings afterwards.

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

This paper presents MetaFeed, a meta-learning framework for improving the cross-domain adaptability of feed-forward visual prediction models. Instead of directly optimizing prediction accuracy through standard supervised training, the method adopts a first-order meta-learning objective that optimizes model parameters for performance after a few supervised adaptation steps. The authors further propose task-dependent partial adaptation, updating selected encoder parameters for geometry prediction and decoder parameters for appearance prediction. After one-time calibration on a small labeled support set from the target domain, the model retains standard feed-forward inference.

The paper evaluates the approach on both geometry reconstruction and appearance prediction across multiple datasets. The results demonstrate consistent improvements over standard supervised training and conventional fine-tuning, with particularly encouraging performance under domain shifts.

Critical strengths

1. Practical and well-motivated problem formulation.

The paper addresses an important limitation of current feed-forward visual prediction models: their performance degradation under domain shifts. Learning a model initialization that can be efficiently calibrated with limited target-domain supervision is a meaningful direction, especially since the adapted model retains standard feed-forward inference without requiring per-scene optimization.

2. Comprehensive experimental evaluation with promising results.

The experiments cover both geometry reconstruction and appearance prediction across multiple in-domain and out-of-domain datasets. The improvements over standard supervised training and conventional fine-tuning are substantial on several benchmarks. For example, the controlled geometry experiments show a considerable reduction in reconstruction error on 7Scenes, while the appearance experiments demonstrate clear improvements on MatrixCity. The evaluation on VGGT further demonstrates the applicability of the framework to a larger-scale feed-forward model.

Critical weaknesses

1. Limited methodological novelty.

My main concern is the limited methodological novelty of the proposed framework. The core approach largely follows established first-order MAML, including support/query episodic training and optimization of a shared initialization for post-adaptation performance. Partial parameter adaptation is also a well-established technique.

The primary task-specific design appears to be the selection of different adaptable parameter subsets for geometry and appearance prediction. While this choice is reasonable and empirically supported, it is relatively straightforward. Overall, the contribution seems to lie primarily in applying and integrating existing meta-learning techniques into feed-forward visual prediction, rather than introducing a substantially new learning algorithm.

2. Insufficient clarity and validation of the supervised fine-tuning baselines.

The effectiveness of MetaFeed is primarily established through comparisons against conventionally trained and fine-tuned models. However, some important configuration details of these baselines are insufficiently specified.

In particular, for the VGGT experiments, the paper does not clearly report the independently selected training config for VGGT-FT or provide sufficient evidence that the conventional fine-tuning baselines have been adequately tuned.

This is particularly relevant because VGGT-FT performs worse than the original public VGGT checkpoint on 7Scenes and DTU, although the different training data prevent this comparison from directly establishing a problem with the baseline. Furthermore, conventional fine-tuning and MetaFeed use the same target-domain adaptation protocol, even though the latter is explicitly optimized for rapid adaptation. It remains unclear whether independently tuning the conventional fine-tuning baselines could narrow the reported performance gap.

Additionally, meta-training introduces non-negligible computational overhead. A comparison under matched training compute budgets would provide further evidence that the improvements arise from the meta-learning objective rather than additional optimization.

This concern directly affects my confidence in the main empirical claims, as the reported advantages over conventional training may depend partly on the baseline configurations. Clarifying and validating these settings is therefore important for assessing the actual contribution of MetaFeed.

Questions for authors

1. Baseline configuration and fairness. Could the authors clarify the initialization, learning rate, optimizer, and training schedule used for VGGT-FT and MetaFeed? In particular, are both initialized from the same pretrained checkpoint? Were the conventional fine-tuning baselines independently tuned for their adaptation learning rates and step counts? Could the authors provide results with stronger tuned baselines or comparable training compute budgets?

2. Contribution of meta-training versus calibration. Tables 1 and 2 compare standard supervised training, conventional fine-tuning, and calibrated MetaFeed, but do not report the performance of MetaFeed before target-domain calibration. Could the authors provide these uncalibrated results, as already done in Table 3? This would help distinguish improvements in direct cross-domain generalization from improvements specifically introduced by calibration.

3. Ablation of the proposed training design. Could the authors provide additional comparisons between intra-scene-only, cross-scene-only, and mixed episodic training? It would also be helpful to evaluate alternative adaptable parameter subsets under comparable parameter budgets, particularly given the different adaptation capacities of the encoder and decoder configurations. These experiments would help clarify the contributions of the proposed task-dependent design choices.

4. Generality to additional prediction tasks. The current evaluation primarily focuses on domain transfer within fixed prediction tasks. Have the authors considered extending the framework to additional dense prediction tasks, such as semantic segmentation? While cross-task adaptation is beyond the current scope, discussing its feasibility or the modifications required would help clarify the generality of the framework and potentially broaden its practical value.

Reviewer_3CENOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MetaFeed, a method for adapting feed-forward geometry and appearance models to new domains with a meta learning formulation. It uses first-order meta-learning to train an initialization that can be calibrated on a small labeled support set. The method adapts the encoder parameters for geometry and decoder parameters for appearance. It then predicts other scenes without further optimization. Experimental results show that the method outperforms standard fine-tuning: on MatrixCity, the albedo PSNR improves from 14.53 to 18.27, and the VGGT experiments show better 3D reconstruction across three benchmarks.

Critical strengths
  1. The problem setup is practical and well motivated. Being able to adapt to new domains efficiently is valuable.
  2. The experiments cover both geometry and appearance, and also contain meaningful controls.
Critical weaknesses
  1. The experiment setting for fine-tuning seems a bit problematic and needs further investigation. Fine-tuning the supervised model often makes it worse, even on its own training domain. For example, on InteriorVerse, fine-tuning on a single in-domain scene roughly quadruples the normal error. LoRA also gets worse as it trains longer. A supervised model should not degrade this much when fine-tuned on labeled data from a domain it already knows. This suggests the baselines are not well tuned. The paper reuses MetaFeed’s adaptation settings (plain SGD with a large learning rate for 5 steps), which the meta-trained model was specifically trained to handle, but the supervised model was not. Since the main claim is that the meta-learned model adapts better, the paper should also tune the hyperparameters (optimizer, learning rate, steps) for the baselines chosen on held-out target scenes, to make the comparison more fair and informative.

  2. It is unclear where the gains come from. MetaFeed’s improvement could come from two places: the meta-training itself, or the adaptation on target data that it is supposed to enable. The paper does not separate the two. In the controlled experiments, the meta-trained model is never evaluated before adaptation, so we cannot tell how much of the gain comes from adaptation. In the one experiment where this is reported (VGGT), most of the improvement is already there before any adaptation. This suggests meta-training may simply act as a better regularizer, or benefit from its extra training compute (about 1.5 to 1.7 times more per step), rather than producing a model that is especially good at adapting. To support the paper’s main claim, the authors should report the unadapted meta-trained model in every table and compare against a supervised baseline trained with the same compute.

Questions for authors
  1. How were the FT and LoRA hyperparameters chosen? Can you report them with their own learning-rate, optimizer, and step sweeps, selected on held-out target scenes? Why does FT-1 degrade in-domain performance so sharply?
  2. What are the zero-step MetaFeed results in Tables 1, 2, and 5, and how does MetaFeed compare with a compute-matched supervised baseline?
  3. During meta-training, every episode’s support and query come from source datasets the model is trained on, so the inner loop only practices adapting within familiar domains. At deployment, the same procedure must adapt the model to an unseen domain. Could the authors comment on this mismatch? For example, does the gain from adaptation shrink as the target domain moves further from the training data, and would episodes that simulate stronger shifts (such as heavy appearance or camera augmentations on the support set) help?

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

Related papers

Discussion (0)

Sign in to comment.