HeadSculpt: Crafting 3D Head Avatars with Text

Recently, text-guided 3D generative methods have made remarkable advancements in producing high-quality textures and geometry, capitalizing on the proliferation of large vision-language and image diffusion models. However, existing methods still struggle to create high-fidelity 3D head avatars in two aspects: (1) They rely mostly on a pre-trained text-to-image diffusion model whilst missing the necessary 3D awareness and head priors. This makes them prone to inconsistency and geometric distortions in the generated avatars. (2) They fall short in fine-grained editing. This is primarily due to the inherited limitations from the pre-trained 2D image diffusion models, which become more pronounced when it comes to 3D head avatars. In this work, we address these challenges by introducing a versatile coarse-to-fine pipeline dubbed HeadSculpt for crafting (i.e., generating and editing) 3D head avatars from textual prompts. Specifically, we first equip the diffusion model with 3D awareness by leveraging landmark-based control and a learned textual embedding representing the back view appearance of heads, enabling 3D-consistent head avatar generations. We further propose a novel identity-aware editing score distillation strategy to optimize a textured mesh with a high-resolution differentiable rendering technique. This enables identity preservation while following the editing instruction. We showcase HeadSculpt's superior fidelity and editing capabilities through comprehensive experiments and comparisons with existing methods.

Paper

Similar papers

Peer review

Reviewer Gcqm6/10 · confidence 5/52023-06-26

Summary

This work proposes a text-to-avatar creation pipeline, building upon dreamfusion and magic3d. To alleviate the geometric ambiguity, the authors replace the vanilla stable diffusion with a landmark-conditioned diffusion model finetuned with controlnet. They also use textual inversion to obtain specific token for back view. Extensive experiments show that the proposed method achieves SOTA performance.

Strengths

- The generation results are impressive and compelling. As shown in Fig. 1, the method can generate avatars in different domains including photo-realistic, anime. - The writing is easy to follow.

Weaknesses

- The generated avatars lack details and suffer some artifacts. The results in Fig. 1 suffer obvious noises especially when zoomed in. Hair and clothing lack details. - How about the generation diversity? Given the same text prompt, can the model generate multiple avatars that align with the text? - One key aspect of avatars is animation. How could the generated avatars be driven? - Fig. 2 is a bit confusing for me. It would be better to clean it. - As in Fig. 3, the editing performs well for frontal view, but fails on side and back views. - Are results in the main paper cherry-picked? Could you show some randomly sampled results? and more visualization for the geometry?

Questions

- The method used DMTet in the second stage, which is a mesh-based representation. But the mesh has difficulties in representing hair, beards, and so on. How do you solve these parts? - How do you train landmark-conditioned controlnet? training data?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

Authors have claimed the limitations.

Reviewer Gcqm2023-08-18

Thanks to the authors for their responses. After reading the rebuttal and comments from other reviewers, I would keep my original ratings.

Reviewer thuE6/10 · confidence 5/52023-07-03

Summary

Existing text-driven 3D generative models could have many problems, such as geometric artifacts (e.g., Janus problem), and visual inconsistency. This work focuses on 3D head avatar generation, which utilizes FLAME model to incorporate human geometric priors into generation, hence resolving those problems in 3D generation. The proposed framework utilizes a coarse-to-fine framework, and it first generates a coarse face via the neural radiance field (NeRF) and then performs refinement/editing using tetrahedron mesh (DMTet). As a result, the proposed method could generate diverse 3D human faces, and it also enables identify-aware editing with the help of both ControlNet and InstructPix2Pix.

Strengths

(1) This work proposes a text-guided 3D generation, specifically for human heads. It could successfully generate various human heads by using or modifying many existing methods, such as FLAME, ControlNet, Instruct-Pix2Pix, and DMTet. (2) This method also studies identity-aware editing by defining a trade-off between the original appearance and the desired editing.

Weaknesses

The proposed text-guided 3D generation method focuses on human heads, but it seems like combining 3D head priors with the existing general 3D generation or style-editing methods. A few problems associated with 3D human heads have not been deeply studied, and please refer to the following detailed weaknesses: (1) This method effectively addresses the limitations encountered in current text-guided 3D generation approaches through the integration of 3D head priors with the FLAME model. By leveraging the FLAME model for 3D head generation, it is expected that this method could facilitate intricate expression variations. However, the current generated head avatars seem to be static only, and the manuscript did not provide many expressions except for the neutral expression. Enabling expressions is important for 3D head avatars, and a few examples could be found in the FLAME model, ICT-FaceKit, and the very relevant work DreamFace. (2) The identity-aware editing has not been well-justified. Based on my understanding, the current method simply mixes the style of the original appearance and the desired editing to generate the facial texture based on the ControlNet-based InstructPix2Pix. I am wondering if it is possible to do a few specific editing of human heads, such as changing the hairstyle, wearing different glasses, and perhaps a few other geometric adjustments (e.g., getting fatter or getting thinner).

Questions

Besides the questions listed in the weakness, I would like to ask: (1) Is the generated head avatar capable of performing various expressions? Or is it possible to perform different expressions by using simple post processes? (2) Is there a clear definition of the so-called identity-awareness of the editing process? Merely blending styles appears to be an unsatisfactory solution. (3) Is it possible to generate/edit human head avatars with different shapes?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

Yes.

Reviewer ztub6/10 · confidence 4/52023-07-03

Summary

The paper proposes a novel method for performing text-to-3D generation for specifically human (or humanoid) heads. These generated heads can be further edited and refined using more fine-grained detailed text prompts while still preserving the identity of the generated asset. This is accomplished with two main contributions: using a landmark-aware score distillation loss in order to ensure that the generated heads roughly align with a template head mesh (FLAME), and introducing a loss function which balances text-based editing with a new description with the original description used to generate the asset. The generated results fix a number of artifacts with existing text-to-3D generation methods specifically for heads. For example, the generated heads do not suffer from the multi-face problems, and the strong prior given by a parametric head model ensures that the generated results are geometrically correct. Additionally, the results can be edited while still maintaining the identity of the original generated results. This is shown in comparison to a number of baseline methods, which are outperformed. AFTER REBUTTAL: I have read the rebuttal and appreciate the detailed response. I appreciate the additional quantitative evaluations, as I believe this was a limitation of the method that has been addressed. However, I still think the quality is a bit limited, but this is the case for most text-to-3D methods. The additional comparisons weakness has been addressed as well. However, I don't see justification for the the more stable training - I would suggest removing this claim without verifying it significantly, while Fig 11 is nice, I am not sure this is enough to make an entire claim based on.

Strengths

In my opinion, the strengths of the paper are that: 1. The presentation of the paper is extremely high quality. The introduction and related work sections are comprehensive, motivate the problem well, and clearly delineate the contributions provided in the paper. The methods section is very clearly described and is simple to follow for those familiar with the field of generative 3D. I view this as very important because it is much easier to glean information from the paper, and future work in text-to-3D (potentially for avatar generation) would be much more likely to build off of this method. 2. The contributions are clearly stated, and to the best of my knowledge they are novel, and they seem like they could be relatively significant. - The idea of integrating ControlNet together with a score distillation loss makes intuitive sense for gaining better control over generated 3D assets, and I have not, to the best of my knowledge, seen it proposed elsewhere. Maintaining control over the structures which are generated from text-to-3D, including ensuring they have higher quality and less noisy geometry, is a very important problem in text-to-3D generation and I view this solution as potentially having a large impact on those working in this field. - While balancing losses between an editing prompt and original prompt seems simple, it does not seem to be used in other text-to-3D methods and seems like it could be useful for text-based editing of already fitted representations in other applications. 3. I find the ablation study to be high quality, and ablate all of the parts of the method which are relevant: introducing ControlNet into the SDS loss, the representation chosen and coarse-to-fine optimization, and the hyperparameters for balancing the editing losses. These are all clearly ablated, along with other smaller contributions such as textual inversion for generating the back of the head.

Weaknesses

In my opinion, there is only one main weakness with the paper, which is that the evaluations and comparisons to baseline methods do not seem complete. 1. There are no quantitative comparisons. While the user study is insightful, methods like DreamFusion proposed the CLIP-R metric. Is there a reason why this was not used in this paper? Does the metric do a poor job at quantifying what it is trying to measure? Additionally, some quantitative results for evaluating the editing quality would be useful: CLIP-R (or something like it, if CLIP-R is not useful) can be used to ensure that the identity is still preserved despite editing with an additional prompt (and perhaps also show that the rendered images are pushed closer to a desired fine-grained editing prompt). 2. The baseline methods are not evaluated entirely fairly. For example, I don’t see any of the baseline methods evaluated for editing quality. What happens if one of the generated examples (ex: those from Fig 4) go through a fine-grained text-based edit? How much will the identity be degraded or the edit not be done sufficiently? I think this is extremely important to show because without it, I don’t know what is the state-of-the-art that is currently being improved on and thus don’t know if the proposed method is even better. 3. It is mentioned that the optimization of the proposed method is “more stable” (L244-245) than baseline methods. However, this claim is not justified anywhere. This would be an extremely important contribution for text-to-3D (where optimization is notoriously brittle), so some quantification of this would be very impactful if it could be shown. For example, ranges of hyperparameters trained with which still lead to a good solution, or error bars on generated results, or failure cases.

Questions

I do not have additional questions brought up (see weaknesses for evaluations questions). The paper is exceptionally clear in describing the method and contributions.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

The paper adequately addresses the limitations of the proposed method. The first limitation, which mentions that the results are “non-deformable” is insightful as this was a question which is natural to ask considering that the FLAME template is a deformable model for heads. Additionally, the generated results still are not completely photorealistic, such as those representations which have been fit from captured data. Some additional failure cases of the method would be interesting to see in order to understand better what the limitations are of the method and which piece would be the biggest bottleneck for someone who wanted to integrate this work into their application.

Reviewer ooKH7/10 · confidence 4/52023-07-05

Summary

Authors proposed a solution for generating human heads based on text description. The technology is based on popularised concepts of control signal for pre-trained diffusion model and static 3D model to guarantee consistent head. Moreover, authors presents the concept of the “back” of the head with a textual inversion. The qualitative and quantitative evaluation shows the superiority of the method over existing competitors.

Strengths

- there are plenty of interesting ideas how to do specific solution including domain knowledge: 1) statistic 3d model (e.g. FLAME), 2) improved score distillation with additional special token, 3) editing directions - Authors demonstrate a wide set of experiments to evaluate their method including small use study

Weaknesses

There is a huge limitation of the method compared to the baselines - The user study can be biased - small set of people, without description of the task and based on the figure 4 it is extremely easy to select the introduced method. - Images are not photorealistic, furthermore the results have an disdvantage from the volumetric rendering (small noise is still visible on all counters) - Results are biased in many directions: huge emotion bias despite the neutral emotions (Figure 1), colour issue for the shadows and unrealistic light baked to the texture.

Questions

Most of the following questions can improve the manuscript: - Why do you use DMTet for discretization? The motivation is missing at line 144+ - In the paragraph near L155 you are elaborating the source of the existing issues. Do you think that main problem is an ambiguity and with rendering each representation for uniformly sample cameras will help to disambiguate the gradients during denoising? - How does the back-view concept is so important if you introduce the landmark based controlNet and the FLAME structure? Can you elaborate why it is not enough to use just the full head with landmarks? The Figure 6 slightly contradicts original motivation of the textual inversion concept for the “back”. - Could you explain in more details the limitation from InstructPix2Pix (L305)?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

- All faces only in a neutral position that is not always correlate with the textual description (e.g. neutral clown) - The overall image quality is biased to the unrealistic colours - Most of the ideas will work only for the presented setup of the heads

Reviewer ztub2023-08-17

Thank you for the detailed clarification and additional quantitative and qualitative evaluations of the method and baselines. This has certainly improved my opinion on the evaluation quality of the method. I do not have additional questions on this. Overall, I am now convinced the quality is an improvement over state-of-the-art, at the cost of limiting to a specific class of objects: heads. However, I do feel that the lack of being able to animate the result severely limits the applicability: a method specifically designed for heads should be able to include head-specific features such as expression editing. With both of these in mind, I still feel slightly positively about the paper in its current state.

Authorsrebuttal2023-08-18

Thank you for your reply - we are pleased our responses addressed your questions. We agree with you that animation is valuable for downstream tasks, but it should not compromise performance or generalization capabilities. At this stage, our framework supports expression editing, albeit not full animation. Enabling animation while retaining generalization ability remains an important challenge as discussed above. Exploring optimal solutions to unlock animation without sacrificing quality or generalization will be a key focus in our future work. Two potential solutions we aim to investigate are: 1) finding improved head representations that retain mesh structure to enable animation while maintaining generalization capacity; and 2) utilizing auto-animation tools in the off-the-shelf graphic pipeline as post-processing on current outputs to add motion. Your insightful comments will help strengthen our work. We are grateful for your time and input.

Reviewer thuE2023-08-19

Response to the Rebuttal and additional questions

Thanks for the rebuttal. After reading the other reviewers' comments, I realize we have similar concerns about the method pipeline. As for the method details, I have more questions, and I hope the authors can provide a thorough explanation. (1) How is the "Mixing" in Figure 2 implemented? The manuscript does not seem to provide corresponding formulas or descriptions for it. (2) In P5 L169, what are the criteria for selecting vertices? Is it random or algorithmic? (3) In P6 L186, where was the data collected from? Is it public or private? Was the selection criteria completely random? (4) In P6 L207, do I and C correspond to the two ControlNets in Figure 2(c) of the PSD? The authors claim they are identical, but based on what is mentioned in section 3.3 and the original ControlNets, they should represent different tasks. (5) In P6 L211, according to Figure 2, the reference image should be renderable from both DMTET (high-resolution) and Coarse Nerf. Why is it rendered only from Coarse Nerf in this case? I will raise my rating if all unclear aspects are adequately explained.

Authorsrebuttal2023-08-19

Thanks for your reply. We provide explanations with respect to these additional questions as follows: 1. Apologies for the confusion. The "Mixing" in Figure 2 refers to Equation 7, which is the weighted averaging of the two predicted noises for IESD. 2. These vertices are selected by a pre-defined index set, which includes the vertices that correspond to the contour of face, eyes, nose, and mouth. This selection process is used to process the projected dense landmarks in the same style as the sparse landmark maps used for training ControlNet. This procedure is the same in all experiments. 3. The images in the tiny dataset are manually downloaded from Google search. Please find more details and example images in the submitted supplementary material. 4. Yes, I and C correspond to the two ControlNets in Figure 2(c) of the PSD. We apologize for any misunderstanding - we did not claim I and C are identical. We would be grateful if reviewer `#thuE` could elaborate so we can address this point fully. 5. As shown in Figure 2, we use a frozen coarse NeRF to render reference images and a fine DMTET to learn the newly edited 3D representation. This means that frozen coarse NeRF-rendered images do not change, while fine DMTET-rendered images do. If we use the DMTET-rendered images as the reference images, the reference image at $i$ iteration will be the edited result from $i-1$ iteration, which will endlessly accumulate the editing appearance and thus leads to unsatisfactory results. Please let us know if these explanations address your questions. We remain open to further discussions and warmly welcome any additional feedback or inquiries you might have.

Reviewer thuE2023-08-21

Thanks for the answers, and I have raised my rating.

Reviewer ooKH2023-08-21

Reply and final note.

Thanks to the authors for their responses, I hope most of the common unclear moments will be added to the final text. After reading the rebuttal and comments from reviewers, I would keep my original ratings.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC