NeurIPS 2020

DeepSVG: A Hierarchical Generative Network for Vector Graphics Animation

Review 1

Summary and Contributions: The paper builds an auto-encoding deep neural network to process SVG programs. By modelling the programs as a set of tokens using a transformer-based architecture, the authors propose a variational auto-encoder that is able to efficiently process simple SVG drawings. Furthermore, the authors propose a novel dataset to train such models and (hopefully) plan to release it along with the work. I believe the contribution is highly relevant to NeurIPS and deserves clear acceptance.

Strengths: The considered task is significantly novel with many applications such as image vectorization and generation of animations, and the presented model is also novel, being in essence a combination of the well-known building blocks, which suggests their stability. The authors conduct an extensive experimental validation building on the novel SVG dataset. The interpolation, generative, and other qualities of the model are clearly demonstrated.

Weaknesses: 1. The model is only able to process simple SVG drawings as of now, which indicates its limitation but does not diminish the novelty. 2. The paper is somewhat difficult to read (see below).

Correctness: The experimental methodology is overall valid but a few things need clarification: 1. Specifically, would SVG-VAE qualify for a baseline? If not, why? 2. What is named as "baseline" in the experimental session? 3. How is the baseline different from “One-stage feed-forward”?

Clarity: Overall the paper is challenging to read, however with a few areas that represent particular difficulty. Specifically, Page 5 is extremely difficult to comprehend. I suggest moving Lines 159—162 describing Transformer modules to before L139, describing Encoders.

Relation to Prior Work: Related work seemingly needs to include along with a short discussion of the relation between the two. Other than that, I have found to gaps in the related literature.

Reproducibility: Yes

Additional Feedback:

Review 2

Summary and Contributions: The paper discusses a method for modeling simple vector drawings such as icons or fonts. Differently to previous works the method is not autoregressive, in which the next step is predicted after previous steps were generated. In the current work, different paths are considered separately and then overlaid together with in the decoder. To model vector graphics the model uses a subset of SVG language with a command of fixed size. The commands and arguments are first encoded using encoding matrices. Then the transformer-based encoder is used. The latent is modeled by a gaussian distribution via VAE.

Strengths: The paper is an interesting piece of work. I agree, modeling vector is an under explored problem. 1. The paper proposed an interesting method of modeling vector graphics, based on a small yet flexible set of svg commands. The paper addressed several problems of how to deal with such entity in machine learning. 2. The paper proposed a dataset of SVG icons which might be interesting for research community. 3. It is also exciting to see that there is some latent space algebra available.

Weaknesses: When reading the paper, I've got the impression that the paper is not finished with couple of key experiments missing. Some parts of the paper lack motivation. Terminology is sometimes unclear and ambiguous. 1. Terminology. The paper uses terms "animation", "generative", "interpolation". See contribution 1 in L40-42. While the paper reported some interpolation experiments, I haven't found any animation or generation experiments. It seems the authors equate interpolation and animation (Section 4.2) which is not correct. I consider animation is a physically plausible motion. Like a person opens a mouth, car moves, while interpolation is just warping one image into the other. Fig 7 shows exactly this with the end states being plausible states of the system. The authors should fix the ambiguity to avoid misunderstanding. The authors also don't report any generation results. Can I sample a random shape from the learnt distribution? If not the I don't think it's correct to say the model is generative. 2. Motivation. It's not clear why the problem is important from practical standpoint? Why one would like to interpolate between two icons? Motivation behind animation is more clear, but in my opinion, the paper doesn't do animation. I believe from a practical standpoint letting the user to input text and be able to generate an icon would also be important. Again, I have hard time understanding why shape autoencoding and interpolation is interesting. 3. Experiments. Probably the biggest concern with the paper is with the experiments. The paper reports only self comparisons. The paper also doesn't explain why this is so, which adds to the poor motivation problem. In a generative setting comparisons with SketchRNN could be performed.

Correctness: Given the motivation and terminology issues I feel the authors should correct the paper to remove potential misunderstanding.

Clarity: I think the paper is written very well.

Relation to Prior Work: I think it is discussed sufficiently well. Comparisons with the related work are missing.

Reproducibility: Yes

Additional Feedback: I think it's a good paper and overall I'm positive. The paper presents an interesting method. There are some issues with the paper. It also true that they try to tackle a somewhat new area. I, however, think that in the present form the paper is slightly below the bar as discussed in the weaknesses section, explaining my rating.

Review 3

Summary and Contributions: The paper proposes a new method called DeepSVG that tackles the ambitious task of generation and interpolation of complex vector graphics. It does so by first concentrating on a dataset the authors collected from 'icons8', and design a variational autoencoder-based architecture that handles the data hierarchically. The sequential mechanism is based on Transformers. The generative network first encodes every path (shape) independently, indexed by E1, and combined with E2 into the final representation (bottleneck) which is then similarly decoded back to the input. Finally, a matching between the input & output is done through Ordered Assignment which was found in an ablation study (human ranking) to give good results. Once reconstruction is achieved, interpolation can be done at the latent space. The authors provide very nice figures in the paper, but even more in the supplementary material, including an interactive tool that let the viewer choose the interpolation between two shapes. The authors will release the collected dataset ('SVG-Icons8'), which is to be appreciated as well. I would encourage the authors also the release the code of their method to allow future work to be able to compare to, since the architecture would be hard to genuinely re-implement.

Strengths: 1. Instead of handling the SVG language as a standard language, the paper proposes a thoughtful way of embedding SVG draw-commands, converting tokens into higher dimensional representation through trainable weights. 2. The method is not autoregressive hence more accessible than alternative methods. 3. Achieves better reconstructions and smoother interpolations than alternative baselines based on the ablation human ranking study.

Weaknesses: 1. Measure of success is limited only to human raters (Table 2). This is a challenge with generative models but I would have liked to know more about how well the structures are reconstructed, perhaps rasterized-based measures could be used. 2. Clearly, scalable vector graphics can be extremely complicated (beyond icon8 dataset) so it's unclear how well the method would perform in other datasets, either more complicated ones or out of domain generalization. What happens when the number of paths exceeds X -- in that case would E2 performs favorably?

Correctness: Ablation study is provided.

Clarity: The paper is well written. The method is described in a clear way. The figures are very descriptive, and the supplementary part has many more examples.

Relation to Prior Work: Sufficient but not extensive.

Reproducibility: No

Additional Feedback:

Review 4

Summary and Contributions: The paper propose a hierarchical generative network architecture, which learns to generate scalable vector graphics. The paper also introduce a dataset of SVG icons, and it shows some successful interpolation and manipulation result with the learned generative model.

Strengths: Directly working on rendering commands of svg icons instead of rasterized images is a more complicated problem which the paper shows a sophisticated solution. The general idea of having a hierarchical VAE encoding/decoding a set of commands in SVG is intuitionly a reasonable solution, which is backed by the experiments.

Weaknesses: A few confusions I had for the paper: 1. The generative model is said to be hierarchical, but the factorization shown in Eq.2 is a single layer model with the variables being independent to each other conditioned on the z. 2. The description of the baseline in sec.4.1 is not abundantly clear, the author should either give a detailed description in the paper/supp, or refer to some reference if the model is borrowed. 3. I don't really see noticable difference between a one-stage model v.s. the hierarchical one in Fig.5. The subjective result in Tab.5 is not enough to convince me the quality difference between the two. 4. also, although it is clear to have the decoder output rendering commands, it is not clear why the encoder should directly take input from rendering commands instead of rasterized images, as in [10]. The paper misses comparison in this regard, which a fair comparison against [10] is needed. 5. No reconstruction error on hold out set is reported, therefore we have no idea about the generalization ability of the trained AE.

Correctness: The description of the method is mostly correct, except Eq.2. which I'd like the author to comment. Additional experiment (see above) is in need.

Clarity: The paper is well written.

Relation to Prior Work: It has a clear discussion of related works.

Reproducibility: Yes

Additional Feedback: