Towards Croppable Implicit Neural Representations

Implicit Neural Representations (INRs) have peaked interest in recent years due to their ability to encode natural signals using neural networks. While INRs allow for useful applications such as interpolating new coordinates and signal compression, their black-box nature makes it difficult to modify them post-training. In this paper we explore the idea of editable INRs, and specifically focus on the widely used cropping operation. To this end, we present Local-Global SIRENs -- a novel INR architecture that supports cropping by design. Local-Global SIRENs are based on combining local and global feature extraction for signal encoding. What makes their design unique is the ability to effortlessly remove specific portions of an encoded signal, with a proportional weight decrease. This is achieved by eliminating the corresponding weights from the network, without the need for retraining. We further show how this architecture can be used to support the straightforward extension of previously encoded signals. Beyond signal editing, we examine how the Local-Global approach can accelerate training, enhance encoding of various signals, improve downstream performance, and be applied to modern INRs such as INCODE, highlighting its potential and flexibility. Code is available at https://github.com/maorash/Local-Global-INRs.

Paper

References (47)

Scroll for more · 35 remaining

Similar papers

Peer review

Reviewer XqfY6/10 · confidence 4/52024-06-28

Summary

This paper proposes Local-Global SIRENs, which partition the space into different regions and fit each region with smaller local INRs, leading to croppable INR by cropping the weights relative to the local regions. The model further proposes to use local and global feature extraction to improve the fitting performance. The experiments show that their method supports cropping INRs for Image, Audio, Video and CT.

Strengths

1. The idea of using partition-based INRs for croppable INRs is novel. Croppable INRs are an interesting application for partition-based INRs. 2. The paper is well-written and easy to follow. 3. The cropping performance looks fancy and the fitting performance is slightly improved compared to SIREN due to the global feature extraction.

Weaknesses

1. Cropping INRs is a natural property of partition-based INRs, which makes the contribution minor. 2. While the authors mention their method supports automatic partitioning, they do not show the detailed implementation of how to automatically partition the space. 3. Even though the authors have conducted an ablation study on partition factors, the range of experimented partition factors is not enough. I recommend the authors to try Partition Factors as low as (2,2) and as high as (512, 512) for the 512 * 512 images and show how the partition affects the fitting performance. 4. Cropping based on simple grids may be impractical for real scenarios. I recommend the authors improve their method by implementing semantic segmentation-based cropping. See partition-based INRs with semantic segmentation [1]. 5. I am confused about the conclusion that enlarging the partition factors leads to faster training while decreasing the partition factors enhances reconstruction accuracy (line 293). Since SIREN is just your Local-Global SIREN with partition factors (1,1), if your conclusion is right, SIREN should have better performance than your LG SIREN. As pointed out in [1], partition should generally improve the fitting performance with larger partition factors. And please provide some explanation about why increasing the number of partitions enhances overall speed. [1] Liu, Ke, et al. "Partition speeds up learning implicit neural representations based on exponential-increase hypothesis." Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023.

Questions

1. Could you please discuss how to automatically partition the region with your method? 2. Could you please discuss whether your method can be extended to semantic segmentation-based cropping?

Rating

6

Confidence

4

Soundness

2

Presentation

3

Contribution

2

Limitations

The limitations and potential negative societal impact of their work have been discussed.

Reviewer AXyi8/10 · confidence 5/52024-07-09

Summary

The paper proposes a method for learning patchwise INRs that are integrated with a global INR. The method is designed with cropping in mind, and this cropping can be achieved by pruning the relevant patchwise INR - similarly, cropping is limited to the pre-defined patches. This allows for novel post-training cropping, where the INR can be cropped in a way that reduces the number of model weights (and therefore the storage space). The method shows benefits beyond cropping - faster training time (or better visual quality at equal epochs). It is also flexible, and can be applied to various MLP-based INRs.

Strengths

[S1] The paper is exceptionally easy to follow, with good reasoning/motivation, well-explained method, and experiments that support the claims of the paper. Last paragraph of the intro, Figure 1, and Figure 3 are especially good in this regard. [S2] While on face the applications seem limited, since the pruned parameters can be discarded, this method could be quite useful in the compression regime, where the MLP INR could be cropped by the end user and size would be reduced accordingly. The faster training time is also beneficial in this setting. [S3] The method has benefits beyond its target problem (cropping) - better quality with less training. [S4] The experiments are very thorough, proving the flexibility of the method with applications across multiple models and domains. [S5] The supplementary even goes beyond what's necessary to provide comparisons with e.g. KD for INR training.

Weaknesses

The patch-based formulation is a little unsatisfying in the following ways. [W1] It doesn't reveal anything very new about the INRs themselves. Rather than discover pruneable parameters that correspond to cropping, or developing some objective that imparts some locality on the representation, the local vs. global distinction is very rigidly enforced as a prior. [W2] The image can only be cropped according to the pre-defined patches. Separately, [W3] Partly connected to W1, the post-training editing is totally restricted to only cropping. This might have applications to compression, but is otherwise at this point more of a novelty, in the sense that it is not very practical.

Questions

What are the practical applications of this method? The main reason my score isn't higher is because I think the impact of this work is somewhat limited to more nice applications that might be of interest to those that study INR, but less so to the broader community. What does this reveal about the nature of INRs? It would also be nice if the paper revealed some deeper insights about INR in general.

Rating

8

Confidence

5

Soundness

4

Presentation

4

Contribution

3

Limitations

Yes.

Authorsrebuttal2024-08-13

We greatly appreciate the reviewer's support for our work. The authors

Reviewer 45rR5/10 · confidence 3/52024-07-14

Summary

This paper proposes a new INR architecture to admit easy cropping of the target datum to a certain partition, allowing one to save memory and inference cost without any retraining. Comparing with training a new INR for the target partition, the approach lets one utilize the global context as well. The idea is to train multiple local networks, and modulate their intermediate features with the features of a global network. Through experiments on image/audio/video encoding, the paper demonstrates that the method enjoys faster and more accurate fitting than training one INR per partition. Also, the paper shows that the method works when combined with many INR architectures, e.g., SIREN and INCODE.

Strengths

- A notable strength of the proposed method is that it is very simple and easy to use, making it likely to be scalable and generally applicable. I believe that the method can also be combined well with the triplane-based neural fields or instant-ngp. - The method is also very clearly presented. In particular, Figure 3 is very effective in delivering how the proposed architecture works. - Lastly, I appreciate the fact that the paper provides experiments on many different modalities, from image to video.

Weaknesses

- **Utility of croppability?** The key weakness of this paper is the motivation. Apart from "saving storage & compute," the practical utility of having a croppable INR is unclear; will local-global INRs also be useful in performing any subsequent "editing" operations? I suspect that this is why authors provide section 4.4, where the authors "extend" the LGS to parameterize the larger image than the one originally considered. However, for such purposes, there are already other good meta-learning-based solutions such as [26]. I am not sure why the model should be croppable for such applications. - **Novelty & Ablations.** The proposed method bears much similarity with [26], which modulates the local model with another global model. The key difference here is how we modulate; this paper uses an extra linear layer to process the local+global features, while [26] uses multiplications (later works, such as functa, used addition). To fully understand what this paper contributes, there should be an explicit comparison with these similar methods as a baseline. - **Evaluation.** If I understood correctly, the evaluation is mostly based on how the model fits the seen coordinates. I wonder how these affect the generalizability of the learned signal to unseen coordinates. As this is one of the key strengths of having a global context, I have enough reasons to believe that LGS will work well. However, an explicit verification is needed. - **Hyperparameter tuning.** I wonder how the hyperparameters for the models and the baselines are selected. In particular, how were the values of "omega" and the learning rate selected? These two are quite critical in determining the fitting speed, so this point should be crystal clear.

Questions

Please see the "weakness."

Rating

5

Confidence

3

Soundness

3

Presentation

4

Contribution

2

Limitations

Yes.

Reviewer XqfY2024-08-10

Official Comment from Reviewer XqfY

I appreciate the author's effort in providing such a detailed rebuttal. After reading all the reviewers' opinions and the rebuttal, I think this method showcases that partition-based INR can be used to crop the image in an elegant manner, which is a great potential application of partition-based INR and may contribute to object detection based on INR. Most of my concerns have been well addressed. I tend to raise my rating to vote for "weak accept".

Authorsrebuttal2024-08-13

We thank the reviewer for considering our rebuttal in such detail. We appreciate the reviewer’s support and the decision to raise the rating. The authors

Reviewer 45rR2024-08-12

Thank you for the detailed response. **Croppability.** The response makes some sense. However, some points are still unclear to me. - The combinability with meta-learning has not been verified experimentally. In fact, this method seems like one specific way to perform meta-learning. - The argument that "enlarging the number of weights is impossible with meta-learning" is exaggerated; one can simply use multiple models (e.g., the one generated by MAML). Instead, I do agree to the point that meta-learning-based solutions may not be as flexible as the proposed one, having a much larger weight-granularity. **Novelty and ablations.** Thank you for the detailed experiment. **Evaluation.** I may have missed section 4.6. Thank you for pointing this out. **Hyperparameters.** Thank you for stating this. --- Many of my concerns have been verified (with some due to my misunderstanding). Although I am still slightly worried about certain points, I am no longer against acceptance of this paper.

Authorsrebuttal2024-08-13

We thank the reviewer for the detailed response, additional comments, and for raising the score. Regarding the additional points: * We agree that definitive claims regarding the applicability of meta-learning techniques on a Local-Global architecture would require explicit experimental verification. While we have not identified specific limitations that would prevent a Local-Global SIREN network from benefiting from a similar meta-learning approach as applied to SIREN, we acknowledge that further research is needed in this area and plan to explore this in future work. * We also agree that stating "enlarging the number of weights is impossible with meta-learning" is not entirely accurate. The rephrasing suggested by the reviewer, noting that meta-learning-based solutions are not "as flexible," is indeed more appropriate. We appreciate the reviewer’s reconsideration and their support towards the acceptance of our paper. The authors

Reviewer AXyi2024-08-12

Rating unchanged

I appreciate the rebuttal. I still think the work is strong, but not groundbreaking, so I keep my rating at strong accept.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC