LLaNA: Large Language and NeRF Assistant

Multimodal Large Language Models (MLLMs) have demonstrated an excellent understanding of images and 3D data. However, both modalities have shortcomings in holistically capturing the appearance and geometry of objects. Meanwhile, Neural Radiance Fields (NeRFs), which encode information within the weights of a simple Multi-Layer Perceptron (MLP), have emerged as an increasingly widespread modality that simultaneously encodes the geometry and photorealistic appearance of objects. This paper investigates the feasibility and effectiveness of ingesting NeRF into MLLM. We create LLaNA, the first general-purpose NeRF-language assistant capable of performing new tasks such as NeRF captioning and Q\&A. Notably, our method directly processes the weights of the NeRF's MLP to extract information about the represented objects without the need to render images or materialize 3D data structures. Moreover, we build a dataset of NeRFs with text annotations for various NeRF-language tasks with no human intervention. Based on this dataset, we develop a benchmark to evaluate the NeRF understanding capability of our method. Results show that processing NeRF weights performs favourably against extracting 2D or 3D representations from NeRFs.

Paper

References (86)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer aHSF6/10 · confidence 4/52024-06-21

Summary

In this submission the authors propose a new pipeline that enables Large Language models to interact with trained object-centric NeRF models. They achieve this by utilising a pretrained meta-network that ingests the weights of a NeRF MLP and outputs a low-dimensional feature vector. This low-dimensional feature vector is then transformed to the language models input space via a projection layer similar to LLaVA. The method is trained on a dataset of 40K NeRF models which the authors extend with 240K paired text descriptions. Experiments on this dataset show favourable performance compared to prior works that solely work with images or point clouds.

Strengths

1. The proposed problem and method of interacting with NeRFs via a LLM is novel and very interesting, enabling potential new applications in robotics and AR. 2. The experiments performed in the submission show promising results, where the proposed method is achieving impressive results over the chosen baselines. 3. The dataset/benchmark and code will be released upon acceptance, which will make follow-up works and comparisons much easier. 4. The paper is very well written and concepts are easy to understand.

Weaknesses

In order of severity: 1. The comparisons that were performed in the experiments seem unfair towards the baselines. The baselines are not trained or fine-tuned on the dataset proposed in the submission and it is assumed that they would generalise on this new domain since they were trained on millions of images or hundreds of thousands of 3D shapes. Since the proposed dataset consists only of ShapeNet objects there will be a domain gap for the baselines to overcome, since they were trained on the Objaverse (Deitke, Matt, et al. "Objaverse: A universe of annotated 3d objects." CVPR 2023) or ModelNet40 (Wu, Zhirong, et al. "3d shapenets: A deep representation for volumetric shapes." CVPR 2015) Datasets in the case of the GPT4Point[58] and PointLLM[77]. The same logic applies to the BLIP2 and LLaVA baselines, which will not have seen many synthetic object centric images during their training. This weakens the overall argument to consider NeRFs as an input modality to LLMs because the presented comparisons are flawed. 2. While it is a valid argument that view selection is an issue for 2D-LLMs, as stated in line 123, NeRFs can render arbitrary viewpoints after training. It would therefore be possible to render multiple images from varying viewpoints and use all of them as input to modern Multimodal LLMs jointly, e.g. by concatenating their text-tokens after the projection layer. This would provide a more balanced comparison than the single-view baselines that are chosen in the submission, since a lot more information can be passed onto the LLM with multiple views. Another avenue to make a fair comparison to 2D-LLMs like LLaVA would be to encode 2D images in an MLP (as shown for example in: Sitzmann, Vincent, et al. "Implicit neural representations with periodic activation functions." Advances in neural information processing systems 33 (2020): 7462-7473.) and then use these weights as input to the proposed method. Both of these points will probably be infeasible to address in a rebuttal but would be interesting experiments to support the use of implicit representations as input to LLMs. 3. It is a bit unclear from the description in line 215 if the test set contains only object classes not seen during training or if they were seen before. This should be clarified. 4. In table 5 it is not discussed what a ‘hard’ view is and what constitutes its ‘hardness/complicatedness’. This makes the table confusing and not self-explanatory. 5. Stating in line 90 that Ballerini et al. [5] are the first to utilise NeRFs as an input modality is not correct, what about NeSF (Vora, Suhani, et al. "Nesf: Neural semantic fields for generalizable semantic segmentation of 3d scenes." arXiv preprint arXiv:2111.13260 (2021)) or Nerf-RPN (Hu, Benran, et al. "Nerf-rpn: A general framework for object detection in nerfs." CVPR 2023)? It probably refers to being the first to directly utilise NeRF MLP weights as an input modality when considering language tasks. This sentence should be re-written.

Questions

To focus the discussion about the weaknesses raised above here are some questions for the authors: 1. Please further justify not finetuning the baselines on the proposed dataset. How is this evaluation done in prior work and what justifications do they give for why they did or did not do this? Do these arguments also apply to this submission? 2. Is it possible to train a limited amount of NeRF models on Objaverse or ModelNet40 and then use the proposed pipeline without finetuning on them and compare to PointLLM and GPT4Point? This would be a very interesting experiment to showcase the generalisation ability of the proposed method. 3. Can you evaluate the 2D baselines with more views ? (related to Weakness 2)

Rating

6

Confidence

4

Soundness

3

Presentation

4

Contribution

3

Limitations

The authors addressed the technical limitations of their work quite well but limitations with regards to prior work are not discussed, as they do not exist in the current evaluation scheme. Societal Impact is briefly discussed in the paper checklist.

Reviewer ZJvT6/10 · confidence 4/52024-07-12

Summary

The authors propose a method to ingest NeRFs and project them into a language model's latent space for question answering and chat applications on NeRFs directly.

Strengths

S1. Very clear abstract, which nicely frames the paper S2. Strong related work section S3. The proposed dataset may be useful to others working on related investigations S4. Many different (standard) metrics considered S5. As NeRFs become a more standard 3D representation, dealing with them in the context of chat applications is increasingly relevant and forward thinking

Weaknesses

W1. Consider adding basic statistics about the proposed dataset in the intro (e.g., the dataset size in number of samples). W2. Consider adding information about the kind of \emph{generalization} that can be expected from the method. Can the method generalize to novel classes? To different instances of fixed classes with held out attributes, colors etc.? W3. ShapeNet seems like a good starting point, but datasets like Objaverse and Objaverse-XL may be interesting to probe generalization. How does the proposed method, trained on ObjectNet generalize to more complex NeRFs? W4. I dont quite follow the addition on L141. Consider giving a more high-level description of this and deferring details to the Appx. W5. L181-182 about retaining safety of the model reads an unsupported claim. Either point to an experiment to this effect, run this experiment, or remove this claim. W6. Figure 3 does not stand alone. Just looking at it I am not able to understand what I should take away. Consider adding more information to the caption so it is clear how to interpret the figure and what the takeaway is. W7. I am not 100% clear on how the training dataset is split in L216. Are there some held out classes? Or is there some data that is held out per class? W8. The proposed method uses fine-tuning while the baseline methods do not apply fine-tuning. Hence, I do not consider these fair comparisons calling into question the validity of the baselines and the gains of the prosed method for NeRFs over traditional 3D and 2D input representations. Adding fine-tuned models as baselines seems important to contextualize the results.

Questions

See weaknesses for specific questions.

Rating

6

Confidence

4

Soundness

2

Presentation

3

Contribution

3

Limitations

Yes the authors address limitations in a specific section.

Reviewer W4u15/10 · confidence 4/52024-07-12

Summary

This work proposes LLaNA, a multimodal large language model (MLLM) that is capable of aligning language with 3D scene fields embedded with (learned) NeRF models. By projecting the NeRF fields into an LLM’s embedding space, LLaNA utilizes a meta encoder to transform fields seamlessly into the token space of an LLM (Llama model in this work) and hence can be adapted to downstream higher-level reasoning tasks (centered around 3D understanding). This work also features a newly created dataset called NeRF-language, which has the emphasis on leveraging NeRF features in answering various visual questions.

Strengths

- The proposed LLaNA method that directly projects a learned radiance field weights into LLM’s space is an interesting idea. - The training on multi-turn 3D-aware question answering style makes good use of both fields and LLM’s capabilities. - A new challenge is proposed with the emphasis on multi-turn QAs on 3D-awareness.

Weaknesses

- A more detailed statistical analysis of the dataset is required, such as detailed sizes, diversities of the questions, human performance or judgment on the dataset, type-token ratios, and some frequent word analysis. - The questions introduced in NeRF-language do not seem to require a very thorough understanding of 3D scenes. - There is an increasing amount of work aimed at marrying the benefits of radiance fields with the strong ability of LLMs, just to name a few [1,2,3]. How does this work compared to them? - Additionally, the 3D-LLM [1] also presents question answering capabilities from 3D input scenes, at least a comparison of the LLaNA method with it is required. - Minor but not of least importance: additional standard VQA baselines can provide more in-depth analysis on the dataset, as well as language-only baseline models to gauge the potential spurious patterns of the curated question-answer relationships. [1] Hong, Yining, et al. "3d-llm: Injecting the 3d world into large language models." NeurIPS 2023. [2] Kerr, Justin, et al. "Lerf: Language embedded radiance fields." ICCV 2023. [3] Qin, Minghan, et al. "LangSplat: 3D Language Gaussian Splatting." CVPR 2024.

Questions

- Are there any questions that exactly depend on the variant of the views where the NeRF fields would benefit? - And if so, what is the detailed percentage analysis of these types of questions, and what do they look like in general?

Rating

5

Confidence

4

Soundness

2

Presentation

3

Contribution

2

Limitations

- The authors addressed some limitations of the work in the manuscript.

Reviewer ZJvT2024-08-11

Thanks for the detailed rebuttal and new experiments! I maintain that as NeRF representations become more standard as a 3D representation, it is important to understand how they can be ingested into causal transformers. While the proposed method does seem susceptible to distribution shifts, I consider adding this result for transparency a huge strength not a weakness. I am raising my score to a 6.

Reviewer aHSF2024-08-11

Thank you for the detailed rebuttal and performing the additional experiments that were requested. This must have been quite a lot of work and I appreciate it. Seeing the fine-tuned results on the ShapeNeRF dataset and experiments on Objaverse, I believe this work presents an interesting approach of combining NeRFs with LLMs that should be presented to the community. I will raise my score to Weak Accept (6).

Reviewer W4u12024-08-12

Thanks for the detailed responses and the additional experiments, they are well-appreciated. The statistical analysis part is more complete now (though still lacking a few items), but they should've been there during submission, and hence I felt the manuscript wasn't ready at its current state. The additional experiments eased some of my doubts, but my W2 still stands, it's more of a matter of how one design the view sampling unless majority of your data requires question askers to examine and be aware of the whole 3D structure of objects. Nevertheless, I raise my score to 5 to give credits to the whole rebuttal process and clearing some of my questions.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC