pith. sign in

arxiv: 2606.12562 · v1 · pith:5IYEN5ATnew · submitted 2026-06-10 · 💻 cs.CV · cs.GR

HairPort: In-context 3D-aware Hair Import and Transfer for Images

Alireza Heidari , Amirhossein Alimohammadi , Wallace Michel Pinto Lira , Adi Bar-Lev , Ali Mahdavi-Amiri This is my paper

Pith reviewed 2026-06-27 09:55 UTC · model grok-4.3

classification 💻 cs.CV cs.GR
keywords hairstyle transfer3D-aware synthesisbald converterimage compositingflow matchingvirtual try-onLoRA adaptationgeometric consistency
0
0 comments X

The pith

HairPort transfers hairstyles across large pose and scale gaps by first making the target face bald then re-rendering the source hair in 3D before synthesis.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents HairPort as a way to move hairstyles from one photo to another even when the people are viewed from very different angles or distances. It separates the problem into two steps: first turning the target person bald with a special converter, then using 3D reconstruction to place the reference hairstyle into the correct viewpoint before a generator combines everything. A new set of 6,000 paired bald and haired images trains the bald converter through in-context adaptation. This matters for virtual try-on tools, augmented reality, and film effects because earlier techniques only worked when poses were nearly the same and could not invent missing hair content reliably.

Core claim

HairPort separates hair removal from transfer by training a Bald Converter on the Baldy dataset of 6,000 paired images via LoRA-based in-context adaptation of FLUX.1 Kontext, then applies a 3D-Aware Transfer Pipeline that reconstructs and re-renders the reference hairstyle from the target viewpoint, and finally feeds the bald source plus geometry-aligned reference into a conditional flow-matching generator to synthesize the result, producing accurate, pose-consistent, and identity-preserving transfers that outperform prior methods under large viewpoint and scale differences.

What carries the argument

The 3D-Aware Transfer Pipeline, which reconstructs the reference hairstyle in 3D and re-renders it from the target viewpoint to enforce geometric consistency before compositing onto the bald source image.

If this is right

  • The method supports hairstyle transfer when source and target differ greatly in viewpoint or scale, cases where missing hair must be synthesized.
  • Results remain identity-preserving and pose-consistent across the transfer.
  • Performance exceeds existing methods in both visual quality and quantitative metrics.
  • The approach directly enables use cases in virtual try-on, augmented reality, and entertainment media.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The separation of bald conversion from geometry transfer could be applied to other facial attributes such as beards or accessories.
  • The Baldy dataset might serve as a starting point for training models that generate or remove other hair-related features.
  • Combining the pipeline with real-time 3D face tracking could support live hairstyle preview in video applications.

Load-bearing premise

The Bald Converter produces realistic bald faces and the 3D reconstruction accurately captures hairstyle geometry so it can be re-rendered from any viewpoint.

What would settle it

A collection of source-target image pairs with extreme pose and scale differences where the output hair either fails to match the expected 3D geometry when viewed from a new angle or looks inconsistent with the bald base would show the method does not achieve the claimed accuracy and consistency.

Figures

Figures reproduced from arXiv: 2606.12562 by Adi Bar-Lev, Ali Mahdavi-Amiri, Alireza Heidari, Amirhossein Alimohammadi, Wallace Michel Pinto Lira.

Figure 1
Figure 1. Figure 1: Given a source portrait and a reference hairstyle, HairPort transfers the reference hair while preserving source identity and background. Explicit 3D [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: HairPort pipeline. From a source image and a reference hairstyle, HairPort first removes the source hair, then reconstructs and aligns the reference hair in 3D to the source viewpoint. A flow-matching synthesizer transfers the aligned hairstyle onto the bald source while preserving identity and background. unstable boundaries for subsequent transfer. Our Bald Converter instead uses geometry-derived segment… view at source ↗
Figure 3
Figure 3. Figure 3: Effect of segmentation guidance on bald reconstruction. (a) Without guidance, the model incorrectly expands the scalp beyond the original hair boundary, distorting head geometry. (b) With our segmentation guidance, the reconstruction remains confined within the correct region, preserving head shape and identity. To generate the dataset, we set SMPL-X in different body poses and facial expressions, and add … view at source ↗
Figure 4
Figure 4. Figure 4: Bald Converter training. (a) We render 3D assets and provide depth, Canny, and segmentation conditions to ControlNet++ together with background cues. SDXL produces bald images matched to the rendered geometry (blue dashed box), and SDXL-Inpaint generates corresponding hair images (orange dashed box), yielding paired hair–bald samples. (b) Each pair and its segmentation maps form a 2 × 2 composite for LoRA … view at source ↗
Figure 5
Figure 5. Figure 5: Ablation. We analyze the impact of key components, including the 3D-aware signal, flow-matching-based hair synthesis, and the balding step. Removing the 3D signal degrades hair geometry preservation, omitting flow matching leads to poor blending and unnatural placement, and removing the balding step results only in color changes without properly removing existing hair. Face Hair 3D Recon [PITH_FULL_IMAGE:… view at source ↗
Figure 6
Figure 6. Figure 6: Failure case. When the 3D reconstruction recovers a hair color that does not match the reference, synthesis is conditioned on this inaccurate signal and the transferred result inherits an inconsistent hair color (here, the reference’s dark olive-green hair is rendered closer to brown). editing. Code, trained models, and the Baldy dataset are publicly available at https://github.com/deepmancer/HairPort/. Ha… view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparisons on face-aligned portraits. HairPort more faithfully matches reference hairstyles while preserving source identity and background. Hair Face AnyDoor HairFusion Ours * MimicBrush * FLUX.2 * ein [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative comparisons on full-frame images. HairPort preserves reference-hair structure and placement under challenging source–reference pose differences. SIGGRAPH Conference Papers ’26, July 19–23, 2026, Los Angeles, CA, USA [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Additional qualitative results. HairPort handles diverse hairstyles, poses, identities, and selected cross-domain source–reference pairs. SIGGRAPH Conference Papers ’26, July 19–23, 2026, Los Angeles, CA, USA [PITH_FULL_IMAGE:figures/full_fig_p010_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: FLUX.2 [klein] 9B integration. We align reference hair to the source view, add the estimated source pose, and extract a hair insertion mask (top). In parallel, our Bald Converter generates the bald source; in rare scale-mismatch cases, soft outpainting expands the source context (bottom). FLUX.2 [klein] 9B then synthesizes the transferred hairstyle inside the mask. ethnicity, facial expression, facial hai… view at source ↗
Figure 11
Figure 11. Figure 11: Baldy dataset samples. Each pair shows a bald image and its corresponding hair version for the same subject. The dataset covers diverse hairstyles (color, length, and texture), viewpoints and head poses (frontal to profile), and a wide range of scenes, including indoor and outdoor backgrounds with varying lighting and camera framing. • Studio / controlled: white, gray, or black seamless backdrop; textured… view at source ↗
Figure 12
Figure 12. Figure 12: Full-frame qualitative comparisons. Hair transfer results on uncropped full-resolution images comparing HairFusion [Chung et al. 2024], AnyDoor [Chen et al. 2024b], MimicBrush [Chen et al. 2024a], and our method (HairPort). Our approach preserves the source identity and background more faithfully while producing geometrically consistent hair placement under large pose differences. SIGGRAPH Conference Pape… view at source ↗
Figure 13
Figure 13. Figure 13: Bald Converter results on in-the-wild–style portraits. For each pair, we show the input portrait (left) and the corresponding bald output generated by our Bald Converter (right). The model removes hair while preserving facial identity and non-hair regions such as skin tone, accessories (e.g., glasses), clothing, and background content, across diverse hairstyles, lighting conditions, and camera viewpoints.… view at source ↗
Figure 14
Figure 14. Figure 14: Bald conversion on non-photorealistic imagery. Examples on anime and cartoon portraits show outputs that retain salient style, facial-proportion, and color-palette cues from each input. SIGGRAPH Conference Papers ’26, July 19–23, 2026, Los Angeles, CA, USA [PITH_FULL_IMAGE:figures/full_fig_p022_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Visual comparison of academic bald-conversion baselines. Face-cropped results for HairCLIPv2, HairMapper, Stable-Hair, and our Bald Converter. SIGGRAPH Conference Papers ’26, July 19–23, 2026, Los Angeles, CA, USA [PITH_FULL_IMAGE:figures/full_fig_p023_15.png] view at source ↗
read the original abstract

Transferring hairstyles between images is an important but challenging task in computer graphics, computer vision, and visual effects. It enables users to explore new looks without physically altering their hair, with applications in virtual try-on systems, augmented reality, and entertainment. Most prior works operate best under small pose gaps, and they fall short under large viewpoint and scale differences, where missing hair content must be synthesized rather than transferred. We propose HairPort, a 3D-aware hairstyle transfer framework that attempts to solve these issues by explicitly separating hair removal from transfer and enforcing geometric consistency before synthesis. We introduce a Bald Converter, which produces realistic bald versions of faces through LoRA-based in-context adaptation of FLUX.1 Kontext. To train our Bald Converter, we introduce a new dataset, Baldy, containing 6,000 paired bald and original images across diverse identities and conditions. We also use a 3D-Aware Transfer Pipeline that reconstructs and re-renders the reference hairstyle from the target viewpoint before compositing it onto the source image. Being 3D aware, our method supports large pose and scale discrepancies between the source and target. Finally, a conditional flow-matching generator synthesizes the transferred result from the bald source and geometry-aligned reference guidance. Together, our method enables accurate, pose-consistent, and identity-preserving hairstyle transfer, outperforming existing methods both qualitatively and quantitatively.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper presents HairPort, a 3D-aware hairstyle transfer framework for images that handles large pose and scale differences. It introduces a Bald Converter (LoRA-based in-context adaptation of FLUX.1 Kontext) trained on the new Baldy dataset of 6,000 paired bald/original images, a 3D-Aware Transfer Pipeline that reconstructs and re-renders reference hairstyle geometry from the target viewpoint before compositing, and a conditional flow-matching generator for final synthesis. The central claim is that this enables accurate, pose-consistent, and identity-preserving transfers that outperform prior methods both qualitatively and quantitatively.

Significance. If the 3D reconstruction and transfer steps hold, the work would advance virtual try-on and AR applications by addressing a key limitation of prior 2D methods under large viewpoint changes. The explicit separation of hair removal from transfer, the new Baldy dataset, and the use of modern components (LoRA adaptation, flow-matching) are concrete strengths. The approach provides a falsifiable pipeline whose components can be ablated.

major comments (3)
  1. [§3] §3 (3D-Aware Transfer Pipeline): The central claim of pose-consistent transfer under large discrepancies rests on the assumption that the 3D reconstruction 'accurately captures hairstyle geometry for re-rendering from arbitrary viewpoints.' No quantitative metrics (e.g., strand-level reconstruction error, novel-view PSNR/SSIM on held-out poses, or ablation removing the 3D step) are referenced, which directly risks the downstream conditional flow-matching generator failing to deliver the claimed consistency.
  2. [§4] §4 (Quantitative Evaluation): The abstract asserts outperformance 'both qualitatively and quantitatively,' yet no tables, metrics (identity cosine similarity, pose error, FID under large pose gaps), or cross-method comparisons with error bars are cited. This is load-bearing for the superiority claim and cannot be verified from the stated results.
  3. [Bald Converter] Bald Converter description: The pipeline assumes the converter produces 'realistic bald versions' without hair-region artifacts that would break identity preservation. No ablation on converter quality (e.g., perceptual metrics on bald outputs or effect on final transfer identity scores) is provided, leaving the identity-preserving claim unsupported.
minor comments (2)
  1. [Related Work] The paper would benefit from explicit comparison to recent 3D hair modeling baselines (e.g., strand-based or Gaussian splatting methods) in the related work section.
  2. Notation for the flow-matching conditioning (bald source + geometry-aligned reference) could be formalized with an equation for clarity.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback highlighting gaps in quantitative validation. We address each major comment below and will perform a major revision to incorporate the requested metrics, ablations, and clarifications.

read point-by-point responses

  1. Referee: [§3] §3 (3D-Aware Transfer Pipeline): The central claim of pose-consistent transfer under large discrepancies rests on the assumption that the 3D reconstruction 'accurately captures hairstyle geometry for re-rendering from arbitrary viewpoints.' No quantitative metrics (e.g., strand-level reconstruction error, novel-view PSNR/SSIM on held-out poses, or ablation removing the 3D step) are referenced, which directly risks the downstream conditional flow-matching generator failing to deliver the claimed consistency.

    Authors: We agree that the 3D reconstruction requires explicit quantitative validation. The submitted manuscript relies primarily on qualitative results for this component. In the revised version we will add novel-view PSNR/SSIM metrics on held-out poses, strand-level error where measurable, and an ablation that removes the 3D re-rendering step to quantify its contribution to final pose consistency. These additions will appear in an expanded §3. revision: yes

  2. Referee: [§4] §4 (Quantitative Evaluation): The abstract asserts outperformance 'both qualitatively and quantitatively,' yet no tables, metrics (identity cosine similarity, pose error, FID under large pose gaps), or cross-method comparisons with error bars are cited. This is load-bearing for the superiority claim and cannot be verified from the stated results.

    Authors: We acknowledge that the submitted manuscript does not present the detailed quantitative tables or error bars referenced in the abstract. In revision we will expand §4 with tables reporting identity cosine similarity, pose error, and FID under large pose gaps, including cross-method comparisons and error bars. The abstract will be updated to explicitly cite these metrics and link to the new tables. revision: yes

  3. Referee: [Bald Converter] Bald Converter description: The pipeline assumes the converter produces 'realistic bald versions' without hair-region artifacts that would break identity preservation. No ablation on converter quality (e.g., perceptual metrics on bald outputs or effect on final transfer identity scores) is provided, leaving the identity-preserving claim unsupported.

    Authors: We agree that an ablation study on Bald Converter quality is needed. The revised manuscript will include perceptual metrics (e.g., LPIPS) on bald outputs versus ground truth and will measure the converter's effect on downstream identity preservation scores in the transfer pipeline. This will be added as a dedicated ablation subsection. revision: yes

Circularity Check

0 steps flagged

No circularity: method components are independently described and trained

full rationale

The paper introduces a Bald Converter via LoRA adaptation on a newly collected Baldy dataset, a 3D reconstruction/re-rendering pipeline, and a conditional flow-matching generator. These are presented as engineering steps with external training data and standard techniques; no equations, fitted parameters, or predictions are shown reducing to self-defined inputs. No self-citation chains or uniqueness theorems are invoked as load-bearing. The central claim is an empirical performance statement, not a derivation that collapses by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 1 invented entities

Abstract-only view supplies no explicit free parameters, axioms, or invented entities beyond the high-level description of the Bald Converter and 3D pipeline.

invented entities (1)
  • Bald Converter no independent evidence
    purpose: Produce realistic bald versions of input faces via LoRA-based in-context adaptation of FLUX.1 Kontext
    Introduced as a core component whose output quality is required for the downstream transfer step.

pith-pipeline@v0.9.1-grok · 5805 in / 1131 out tokens · 14114 ms · 2026-06-27T09:55:04.206351+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

62 extracted references · 2 canonical work pages · 1 internal anchor

  1. [1]

    2022 , eprint=

    Style Your Hair: Latent Optimization for Pose-Invariant Hairstyle Transfer via Local-Style-Aware Hair Alignment , author=. 2022 , eprint=

    2022

  2. [2]

    2022 , eprint=

    HairFIT: Pose-Invariant Hairstyle Transfer via Flow-based Hair Alignment and Semantic-Region-Aware Inpainting , author=. 2022 , eprint=

    2022

  3. [3]

    2020 , eprint=

    MichiGAN: Multi-Input-Conditioned Hair Image Generation for Portrait Editing , author=. 2020 , eprint=

    2020

  4. [4]

    Barbershop: GAN-based image compositing using segmentation masks , volume=

    Zhu, Peihao and Abdal, Rameen and Femiani, John and Wonka, Peter , year=. Barbershop: GAN-based image compositing using segmentation masks , volume=. ACM Transactions on Graphics , publisher=. doi:10.1145/3478513.3480537 , number=

    work page doi:10.1145/3478513.3480537

  5. [5]

    2021 , eprint=

    LOHO: Latent Optimization of Hairstyles via Orthogonalization , author=. 2021 , eprint=

    2021

  6. [6]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Wu, Yiqian and Yang, Yong-Liang and Jin, Xiaogang , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2022 , pages =

    2022

  7. [7]

    2022 , eprint=

    HairCLIP: Design Your Hair by Text and Reference Image , author=. 2022 , eprint=

    2022

  8. [8]

    2023 , eprint=

    HairCLIPv2: Unifying Hair Editing via Proxy Feature Blending , author=. 2023 , eprint=

    2023

  9. [9]

    2019 , eprint=

    A Style-Based Generator Architecture for Generative Adversarial Networks , author=. 2019 , eprint=

    2019

  10. [10]

    2024 , eprint=

    Stable-Hair: Real-World Hair Transfer via Diffusion Model , author=. 2024 , eprint=

    2024

  11. [11]

    2025 , eprint=

    Stable-Hair v2: Real-World Hair Transfer via Multiple-View Diffusion Model , author=. 2025 , eprint=

    2025

  12. [12]

    2025 , eprint=

    In-Context Edit: Enabling Instructional Image Editing with In-Context Generation in Large Scale Diffusion Transformer , author=. 2025 , eprint=

    2025

  13. [13]

    2025 , eprint=

    Edit Transfer: Learning Image Editing via Vision In-Context Relations , author=. 2025 , eprint=

    2025

  14. [14]

    2024 , eprint=

    HairFastGAN: Realistic and Robust Hair Transfer with a Fast Encoder-Based Approach , author=. 2024 , eprint=

    2024

  15. [15]

    2024 , eprint=

    What to Preserve and What to Transfer: Faithful, Identity-Preserving Diffusion-based Hairstyle Transfer , author=. 2024 , eprint=

    2024

  16. [16]

    Hair-GANs: Recovering 3D Hair Structure from a Single Image

    Zhang, Meng and Zheng, Youyi , title = ". doi:10.48550/arXiv.1811.06229 , archivePrefix =. 1811.06229 , primaryClass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1811.06229

  17. [17]

    Advances in Intelligent Systems and Computing , publisher =

    GAN with Multivariate Disentangling for Controllable Hair Editing , author =. Advances in Intelligent Systems and Computing , publisher =

  18. [18]

    Computer Vision -- ECCV 2022 , series=

    HairNet: Hairstyle Transfer with Pose Changes , author=. Computer Vision -- ECCV 2022 , series=. 2022 , doi=

    2022

  19. [19]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Hairnerf: Geometry-aware image synthesis for hairstyle transfer , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=. 2023 , publisher=

    2023

  20. [20]

    IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    Khwanmuang, Sasikarn and Phongthawee, Pakkapon and Sangkloy, Patsorn and Suwajanakorn, Supasorn , title =. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages =. 2023 , publisher =

    2023

  21. [21]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Few-shot head swapping in the wild , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=. 2022 , publisher=

    2022

  22. [22]

    2022 , eprint=

    Hierarchical Text-Conditional Image Generation with CLIP Latents , author=. 2022 , eprint=

    2022

  23. [23]

    2022 , eprint=

    Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding , author=. 2022 , eprint=

    2022

  24. [24]

    SKED: Sketch-guided Text-based 3D Editing , booktitle =

    Aryan Mikaeili and Or Perel and Mehdi Safaee and Daniel Cohen-Or and Ali Mahdavi-Amiri , pages=. SKED: Sketch-guided Text-based 3D Editing , booktitle =

  25. [25]

    2024 , howpublished=

    Black Forest Labs , title=. 2024 , howpublished=

    2024

  26. [26]

    2025 , eprint =

    ASIA: Adaptive 3D Segmentation Using Few Image Annotations , author =. 2025 , eprint =

    2025

  27. [27]

    2023 , eprint=

    Sdxl: Improving latent diffusion models for high-resolution image synthesis , author=. 2023 , eprint=

    2023

  28. [28]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    High-resolution image synthesis with latent diffusion models , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=. 2022 , publisher=

    2022

  29. [29]

    2024 , eprint=

    SLiMe: Segment Like Me , author=. 2024 , eprint=

    2024

  30. [30]

    2024 , eprint=

    EmerDiff: Emerging Pixel-level Semantic Knowledge in Diffusion Models , author=. 2024 , eprint=

    2024

  31. [31]

    2023 , eprint=

    MasaCtrl: Tuning-Free Mutual Self-Attention Control for Consistent Image Synthesis and Editing , author=. 2023 , eprint=

    2023

  32. [32]

    2025 , eprint=

    h-Edit: Effective and Flexible Diffusion-Based Editing via Doob's h-Transform , author=. 2025 , eprint=

    2025

  33. [33]

    2024 , eprint=

    Zero-Shot Video Semantic Segmentation based on Pre-Trained Diffusion Models , author=. 2024 , eprint=

    2024

  34. [34]

    2025 , eprint=

    FLUX.1 Kontext: Flow Matching for In-Context Image Generation and Editing in Latent Space , author=. 2025 , eprint=

    2025

  35. [35]

    2019 , eprint=

    Expressive Body Capture: 3D Hands, Face, and Body from a Single Image , author=. 2019 , eprint=

    2019

  36. [36]

    and Li, Hao and Romero, Javier , journal=

    Li, Tianye and Bolkart, Timo and Black, Michael J. and Li, Hao and Romero, Javier , journal=. Learning a Model of Facial Shape and Expression from. 2017 , month=nov, doi=

    2017

  37. [37]

    2025 , eprint=

    DiffLocks: Generating 3D Hair from a Single Image using Diffusion Models , author=. 2025 , eprint=

    2025

  38. [38]

    ACM Transactions on Graphics (Proceedings of SIGGRAPH) , volume =

    Liwen Hu and Chongyang Ma and Linjie Luo and Hao Li , title =. ACM Transactions on Graphics (Proceedings of SIGGRAPH) , volume =. 2015 , url =

    2015

  39. [39]

    2024 , howpublished =

    Hair20K: A Large 3D Hairstyle Database for Hair Modeling , author =. 2024 , howpublished =

    2024

  40. [40]

    Computer Graphics Forum , volume =

    Matt Jen-Yuan Chiang and Benedikt Bitterli and Chuck Tappan and Brent Burley , title =. Computer Graphics Forum , volume =. 2016 , doi =

    2016

  41. [41]

    2023 , eprint=

    BEDLAM: A Synthetic Dataset of Bodies Exhibiting Detailed Lifelike Animated Motion , author=. 2023 , eprint=

    2023

  42. [42]

    2024 , note =

    ControlNet++: All-in-one ControlNet for Image Generation and Editing , author =. 2024 , note =

    2024

  43. [43]

    Bovik and Hamid R

    Zhou Wang and Alan C. Bovik and Hamid R. Sheikh and Eero P. Simoncelli , title =. IEEE Transactions on Image Processing , volume =

  44. [44]

    Jiankang Deng and Jia Guo and Niannan Xue and Stefanos Zafeiriou , title =. Proc. CVPR , pages =. 2019 , publisher =

    2019

  45. [45]

    Jia Guo and Jiankang Deng and others , title =

  46. [46]

    Alec Radford and Jong Wook Kim and Chris Hallacy and Aditya Ramesh and Gabriel Goh and Sandhini Agarwal and Girish Sastry and Amanda Askell and Pamela Mishkin and Jack Clark and Gretchen Krueger and Ilya Sutskever , title =. Proc. ICML , series =. 2021 , publisher =

    2021

  47. [47]

    2025 , eprint=

    DINOv3 , author=. 2025 , eprint=

    2025

  48. [48]

    2025 , eprint=

    Ultra3D: Efficient and High-Fidelity 3D Generation with Part Attention , author=. 2025 , eprint=

    2025

  49. [49]

    2025 , eprint=

    Hi3DGen: High-fidelity 3D Geometry Generation from Images via Normal Bridging , author=. 2025 , eprint=

    2025

  50. [50]

    2025 , eprint=

    Insert Anything: Image Insertion via In-Context Editing in DiT , author=. 2025 , eprint=

    2025

  51. [51]

    2024 , eprint=

    AnyDoor: Zero-shot Object-level Image Customization , author=. 2024 , eprint=

    2024

  52. [52]

    2026 , howpublished=

    Black Forest Labs , title=. 2026 , howpublished=

    2026

  53. [53]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Cao, Zhe and Simon, Tomas and Wei, Shih-En and Sheikh, Yaser , title=. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=. 2017 , publisher=

    2017

  54. [54]

    2024 , howpublished=

    2024

  55. [55]

    2025 , month=nov, howpublished=

    Introducing. 2025 , month=nov, howpublished=

    2025

  56. [56]

    2018 , eprint=

    Progressive Growing of GANs for Improved Quality, Stability, and Variation , author=. 2018 , eprint=

    2018

  57. [57]

    2024 , eprint=

    Zero-shot Image Editing with Reference Imitation , author=. 2024 , eprint=

    2024

  58. [58]

    2018 , eprint=

    GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium , author=. 2018 , eprint=

    2018

  59. [59]

    2025 , eprint=

    SAM 3: Segment Anything with Concepts , author=. 2025 , eprint=

    2025

  60. [60]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Mv-adapter: Multi-view consistent image generation made easy , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=. 2025 , publisher=

    2025

  61. [61]

    2025 , eprint=

    SHeaP: Self-Supervised Head Geometry Predictor Learned via 2D Gaussians , author=. 2025 , eprint=

    2025

  62. [62]

    2025 , eprint=

    Pixel3DMM: Versatile Screen-Space Priors for Single-Image 3D Face Reconstruction , author=. 2025 , eprint=

    2025