Pith. sign in

REVIEW 3 major objections 5 minor 63 references

MultiAnimate claims that multiple characters can be animated together in one shared scene, each driven by its own reference image and pose sequence, without losing identity or spatial relationships.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 05:13 UTC pith:ACHRIV5Y

load-bearing objection A useful system paper with a genuinely new input format, but its identity labels are assigned by initial position with no re-ID, which undermines the exact claim it makes about crossings. the 3 major comments →

arxiv 2607.13415 v1 pith:ACHRIV5Y submitted 2026-07-15 cs.CV

MultiAnimate: A Unified Framework for Controllable Multi-Character Animation

Zhongyi Zhang , Guangyuan Wang , Li Hu , Wenbo Zhou , Peng Zhang , Tianyi Wei , Weiming Zhang , Bang Zhang
show 1 more author
Nenghai Yu
This is my paper
classification cs.CV
keywords multi-character animationvideo diffusionidentity preservationpose-guided animationreference image animationspatial relationshipsinteraction generationcontrollable video synthesis
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

MultiAnimate is a framework for animating several characters at once. Its central claim is that each character can be controlled by a separate reference image and a separate pose sequence, rather than requiring all characters to appear in one reference image, which is the limitation of earlier character-animation systems. The paper argues that three design choices make this possible: per-reference position encodings that keep appearance features apart, a cross-attention step that binds each pose to the right identity, and an optional interaction guider that resolves occlusions during contact. Together these allow a shared background with multiple characters that swap positions and interact while each one stays recognizable. A sympathetic reader would care because this is the difference between animating a single person and animating a scene with several people who touch, occlude, and cross paths.

Core claim

In the paper's own terms, MultiAnimate demonstrates a way to extend character image animation from one reference to many. The identity-specific reference net adds a learnable position encoding to every reference feature before concatenation, so the denoising network can attribute each part of the generated frame to the correct character's appearance. The identity-aware pose encoder uses cross-attention between a pose feature and that same character's reference feature, giving each pose an explicit identity. The interaction guider renders occluded skeleton parts in white and drops low-confidence joints, teaching the model how characters should layer when they interact. The paper reports that

What carries the argument

The identity-specific reference net is the core addition: it adds a learned position encoding to each reference image's features before they are repeated, concatenated, and fed to cross-attention, so the network can tell which appearance belongs to which character. Alongside it, the identity-aware pose encoder binds a pose sequence to its character by querying that character's reference feature with the pose feature. The interaction guider, optional, uses per-character masks to mark occluded skeleton joints white before pose encoding, so the model learns occlusion and depth between characters. A background branch and a sliding-window latent fusion complete the pipeline for shared scenes and

Load-bearing premise

The training data labels each character's identity by its initial horizontal position; when characters cross or heavily occlude, that label can silently transfer to the wrong person, so the 'ground truth' for identity-pose binding may be corrupted and the model could learn the wrong mapping.

What would settle it

Take a held-out clip where two characters cross paths and swap horizontal positions mid-video, run the released pipeline, and measure whether the generated character with reference A's face is still following pose A after the swap; a visible identity swap at the moment the position-based labels would swap would refute the claim. A cheaper check would be to run an appearance-based identity tracker over the training videos and count how often the position-based labels disagree with it.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Each character can be animated from its own photo and motion sequence, so users no longer need to find or compose a single reference image containing everyone.
  • Identity stays attached to the correct character even when characters swap positions or occlude each other, because pose features carry an explicit identity binding.
  • The framework scales to more than two characters and still maintains per-character appearance, given corresponding references and poses.
  • Single-character animation remains possible and performs comparably to dedicated single-character methods.
  • Long videos can be generated by stitching overlapping segments in latent space, with temporal consistency maintained by fusion.

Where Pith is reading between the lines

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

  • If the position-based identity labels used in training are unreliable during crossings, the model may learn to swap identities in exactly the cases the paper highlights; using appearance-based tracking to label identities could make the framework safer.
  • The quality of the interaction guider is bounded by the segmentation model's ability to track occluded limbs; when the segmenter fails, the rendered skeleton states will be wrong, so interaction quality depends on upstream segmentation quality.
  • The recipe—per-reference position encodings plus pose-identity cross-attention—generalizes beyond this specific pipeline and could be applied to transformer-based video backbones, which the paper itself notes as future work.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. MultiAnimate proposes a diffusion-based framework for animating multiple characters from separate reference images and separate pose sequences within one shared video, while preserving each character's identity and inter-character spatial relations. The method consists of three modules: an identity-specific reference net that adds learnable position encodings to multiple reference features before concatenation; an identity-aware pose encoder that binds each pose feature to its reference image via cross-attention; and an optional interaction guider that encodes visible/masked/low-confidence skeleton states from SAM2 masks. Training is self-supervised on a newly collected multi-character video dataset with a position-based identity labeling protocol. The paper reports quantitative comparisons against adapted single-character animators and ablations supporting each module, plus a single-character sanity check on the TikTok dataset.

Significance. If the identity-labeling issue is resolved, the architecture is a genuine step beyond single-reference character animation: it lifts the 'one reference image containing all characters' restriction of prior multi-character methods and makes a concrete, falsifiable design proposal (per-reference positional encodings, cross-attention pose binding, mask-state occlusion guidance). The paper also ships useful engineering details and commits to releasing code and data. The external sanity check on TikTok and the use of established components (AnimateAnyone-style reference net, SAM2, DWPose) give the method credibility. However, the central multi-character claim is currently evaluated only on a small, self-collected, and withheld test set, and the labeling protocol that supervises the identity-aware pose encoder has a known failure mode under position interchange. The significance of the contribution therefore depends on resolving that issue and on obtaining external or released-data validation.

major comments (3)
  1. [Sec. 4.1; Sec. 3.4, Eq. (5)] The position-based labeling protocol assigns character indices from the sorted x-coordinates of bounding-box centers in the first frame and gives no rule for propagating identities to later frames. In a two-character crossing, the person initially at position 0 moves to position 1; under this protocol the label '0' would then name the other person. The identity-aware pose encoder of Sec. 3.4 (Eq. 5) is trained on these labels, so after a crossing the pose sequence associated with reference I_0 is actually the other character's pose. The same ambiguity contaminates the pose-interchange ablation in Sec. 4.6.2 and the test set (Sec. 4.1), so the metrics in Table 1 do not certify identity preservation through position interchange, which is precisely the scenario highlighted in Fig. 1, row 3 and Sec. 4.6.2. The 65% occlusion filter does not remove crossings. Please specify the tracking or re-
  2. [Sec. 4.2.2; Table 1] All quantitative baselines are single-character animators adapted by consolidating all target characters into one reference image and one pose map. The only published multi-character methods (Xue et al. [17], Wang et al. [18]) are excluded because their code is unavailable. This adapted comparison is not the native operating regime of those baselines and may disadvantage them. Consequently, the paper's claim in Sec. 1 that existing methods 'only work for some simple scenarios' is not substantiated by the quantitative evidence. Please either include results (even qualitative) for [17,18] where possible, or explicitly frame Table 1 as an upper-bound comparison against adapted single-character methods and soften the superiority claim accordingly.
  3. [Sec. 4.1; Sec. 4.2.2; Table 1] The multi-character evaluation uses only a 5% subset of the authors' own collected dataset, and the paper reports no standard errors, confidence intervals, or the number of test videos. The large reported metric gaps (e.g., MSE 1.58 vs. 4.60) could in principle be driven by a small or biased test sample. Please report the test-set size, per-video variance, and significance/confidence intervals. Ideally, evaluate on an external multi-character benchmark or a public multi-person pose/video dataset so the central claim is not self-referential.
minor comments (5)
  1. [Sec. 3.6, Algorithm 1] The sliding-window latent fusion appears asymmetric. For segment i, the 'previous' overlap update (line 9) uses z^{i-1}, but segment i-1 may already have been averaged with z^i in its own 'next' branch (line 11), so overlapping latents can be averaged more than once. Please clarify the intended averaging order or rewrite the algorithm as a single symmetric average over each overlapping pair.
  2. [Table 3; Sec. 4.6.1] The quantitative ablation of the identity-specific reference net reports only one-layer vs. all-layer position encodings; the 'no PE' and 'static PE' settings appear only qualitatively in Fig. 7. Add quantitative numbers for those baselines to make the module-level contribution measurable.
  3. [Sec. 4.5; Table 2] The user study reports very high preference rates (94.8-95.6%) but gives no number of participants, no confidence intervals, and no statistical test. Please report these details; the extreme skew relative to baselines warrants verification.
  4. [Sec. 4.7; Table 4] Several baseline numbers are taken from original papers and may use different evaluation protocols. Please state explicitly which values were recomputed on the TikTok split and which were copied, and confirm that the evaluation settings are identical.
  5. [Sec. 3.2; Fig. 2(a)] The background branch is mentioned as an integral part of the framework, but its architecture and training are not described and no ablation isolates its contribution. A short description or reference to the MIMO formulation would improve reproducibility.

Circularity Check

0 steps flagged

No significant circularity: the framework's modules are defined by explicit equations, trained on held-out data, and externally grounded by the TikTok single-character benchmark.

full rationale

The derivation chain is not circular. The main claimed results are the three architectural modules: the identity-specific reference net (Eq. 3-4), the identity-aware pose encoder (Eq. 5), and the interaction guider. Each is defined by explicit mechanisms and tested with ablations; none is a renamed output or a fitted parameter presented as a prediction. Training is self-supervised on an author-collected multi-character video dataset, and Table 1 is computed on a randomly selected held-out 5% subset, so the quantitative results are genuine predictions rather than fits to the test data. The position-based labeling protocol in Sec. 4.1 is a data-annotation choice, not a definition of the model output; even if it were imperfect under crossings, that would be a data-quality or robustness risk, not a circular reduction. The pose-interchange ablation in Sec. 4.6.2 directly tests identity-pose binding by swapping pose sequences across reference images, so the central claim is not equivalent to the labels by construction. Self-citations are present (AnimateAnyone [5], Movie Weaver [43]), but AnimateAnyone is an open, externally validated base architecture and Movie Weaver is cited only as partial inspiration; neither carries the paper's conclusion by itself. Single-character evaluation on the external TikTok benchmark (Table 4) provides independent grounding for the core pipeline. The dataset and code are not yet released, which limits reproducibility, but this is an availability limitation, not circularity.

Axiom & Free-Parameter Ledger

2 free parameters · 6 axioms · 0 invented entities

The paper is an empirical ML system; the central claim depends on a large set of learned parameters fitted to the authors' collected dataset, plus hand-chosen preprocessing thresholds and module designs. There are no invented physical entities. The background assumptions are standard diffusion models and off-the-shelf perception tools.

free parameters (2)
  • Learned network weights (identity-specific reference net, identity-aware pose encoder, interaction guider, denoising U-N = not reported
    The central claim is an empirical architecture; all performance is a function of these fitted weights, and no final parameter values or error bars are provided.
  • Dataset filtering thresholds: bounding box size, occlusion ratio = 10% of frame; 65% occlusion in >50% frames
    Hand-chosen in Section 4.1; they define the data distribution and thus the evaluation set, but are not justified or varied.
axioms (6)
  • standard math Latent diffusion training objective and denoising function (Eq. 1)
    The framework is built on the standard diffusion objective; the paper does not derive it but relies on it as background.
  • domain assumption Reference net architecture from AnimateAnyone [5] provides effective single-character appearance extraction
    The identity-specific reference net extends this architecture without re-validating the base component; it is taken as a known-good building block.
  • domain assumption DWPose and SAM2 provide reliable pose and mask estimates in multi-character scenes
    Both the pose encoder and interaction guider depend on the quality of these off-the-shelf tools; errors would propagate into training and evaluation.
  • ad hoc to paper Position-based labeling protocol preserves identity across inter-character position interchanges
    Section 4.1 assigns character IDs by initial horizontal order; the method assumes these IDs remain correct through swaps and occlusions, which is load-bearing for the identity-pose binding.
  • ad hoc to paper Interaction guider's three-state rendering (visible, white masked, omitted low-confidence) provides useful occlusion supervision
    Section 3.5 hand-designs these states; there is no quantitative comparison of alternative rendering schemes.
  • domain assumption SSIM/PSNR/LPIPS/MSE/FID/FVD adequately capture generation quality for this task
    The paper uses standard metrics but does not discuss their limitations for identity preservation or motion realism.

pith-pipeline@v1.3.0-alltime-deepseek · 14181 in / 11162 out tokens · 112176 ms · 2026-08-02T05:13:45.091661+00:00 · methodology

0 comments
read the original abstract

Recent advances in generative models and technological innovations have significantly addressed the fundamental challenges of character image animation. However, existing approaches predominantly focus on character animation from a single reference image, substantially limiting their applicability in scenarios such as multiple character interaction animation. To fill this gap, this paper introduces MultiAnimate, a comprehensive framework that enables concurrent animation of multiple characters within a shared environment while preserving both identity consistency and spatial relationships. The framework achieves these objectives through multiple well-designed mechanisms. First, we incorporate an identity-specific reference net that enables appearance extraction from multiple reference images, distinguishing MultiAnimate from existing approaches constrained to single reference inputs. Second, we implement an identity-aware pose encoder to address the character-pose binding challenge, wherein an attention mechanism enables the network to accurately differentiate and process multiple pose sequences during generation. Third, we introduce an interaction guider module that enhances the framework's capability to handle complex inter-character interactions by leveraging character-specific mask information, serving as an optional component that refines the pose sequences. Extensive experiments and ablation analyses demonstrate our framework's superiority in multiple character animation, particularly in scenarios involving complex motion sequences.

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

63 extracted references · 40 linked inside Pith

  1. [1]

    In: International Conference on Learning Representations (2022) 20

    Wang, Y., Yang, D., Bremond, F., Dantcheva, A.: Latent image animator: Learn- ing to animate images via latent space navigation. In: International Conference on Learning Representations (2022) 20

    2022

  2. [2]

    In: Conference on Neural Information Processing Systems (NeurIPS) (2019)

    Siarohin, A., Lathuili` ere, S., Tulyakov, S., Ricci, E., Sebe, N.: First order motion model for image animation. In: Conference on Neural Information Processing Systems (NeurIPS) (2019)

    2019

  3. [3]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp

    Gao, Y., Zhou, Y., Wang, J., Li, X., Ming, X., Lu, Y.: High-fidelity and freely controllable talking head video generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 5609–5619 (2023)

    2023

  4. [4]

    arXiv preprint arXiv:2311.12052 (2023)

    Chang, D., Shi, Y., Gao, Q., Fu, J., Xu, H., Song, G., Yan, Q., Yang, X., Soley- mani, M.: Magicdance: Realistic human dance video generation with motions & facial expressions transfer. arXiv preprint arXiv:2311.12052 (2023)

    Pith/arXiv arXiv 2023

  5. [5]

    arXiv preprint arXiv:2311.17117 (2023)

    Hu, L., Gao, X., Zhang, P., Sun, K., Zhang, B., Bo, L.: Animate anyone: Con- sistent and controllable image-to-video synthesis for character animation. arXiv preprint arXiv:2311.17117 (2023)

    Pith/arXiv arXiv 2023

  6. [6]

    arXiv preprint arXiv:2406.19680 (2024)

    Zhang, Y., Gu, J., Wang, L.-W., Wang, H., Cheng, J., Zhu, Y., Zou, F.: Mimic- motion: High-quality human motion video generation with confidence-aware pose guidance. arXiv preprint arXiv:2406.19680 (2024)

    Pith/arXiv arXiv 2024

  7. [7]

    arXiv preprint arXiv:2304.01186 (2023)

    Ma, Y., He, Y., Cun, X., Wang, X., Shan, Y., Li, X., Chen, Q.: Follow your pose: Pose-guided text-to-video generation using pose-free videos. arXiv preprint arXiv:2304.01186 (2023)

    Pith/arXiv arXiv 2023

  8. [8]

    In: arXiv (2023)

    Xu, Z., Zhang, J., Liew, J.H., Yan, H., Liu, J.-W., Zhang, C., Feng, J., Shou, M.Z.: Magicanimate: Temporally consistent human image animation using diffusion model. In: arXiv (2023)

    2023

  9. [9]

    arXiv preprint arXiv:2307.00040 (2023)

    Wang, T., Li, L., Lin, K., Zhai, Y., Lin, C.-C., Yang, Z., Zhang, H., Liu, Z., Wang, L.: Disco: Disentangled control for realistic human dance generation. arXiv preprint arXiv:2307.00040 (2023)

    Pith/arXiv arXiv 2023

  10. [10]

    In: European Conference on Computer Vision (ECCV) (2024)

    Zhu, S., Chen, J.L., Dai, Z., Xu, Y., Cao, X., Yao, Y., Zhu, H., Zhu, S.: Champ: Controllable and consistent human image animation with 3d parametric guidance. In: European Conference on Computer Vision (ECCV) (2024)

    2024

  11. [11]

    https://arxiv.org/ abs/2304.06025

    Karras, J., Holynski, A., Wang, T.-C., Kemelmacher-Shlizerman, I.: DreamPose: Fashion Image-to-Video Synthesis via Stable Diffusion (2023). https://arxiv.org/ abs/2304.06025

    Pith/arXiv arXiv 2023

  12. [12]

    https://arxiv.org/abs/2410

    Yoon, S., Koo, G., Lee, Y., Yoo, C.D.: TPC: Test-time Procrustes Calibration for Diffusion-based Human Image Animation (2025). https://arxiv.org/abs/2410. 24037

    2025

  13. [13]

    https:// 21 arxiv.org/abs/2407.09012

    Kim, J., Kim, M.-J., Lee, J., Choo, J.: TCAN: Animating Human Images with Temporally Consistent Pose Guidance using Diffusion Models (2024). https:// 21 arxiv.org/abs/2407.09012

    Pith/arXiv arXiv 2024

  14. [14]

    arXiv preprint arXiv:2406.01188 (2024)

    Wang, X., Zhang, S., Gao, C., Wang, J., Zhou, X., Zhang, Y., Yan, L., Sang, N.: Unianimate: Taming unified video diffusion models for consistent human image animation. arXiv preprint arXiv:2406.01188 (2024)

    Pith/arXiv arXiv 2024

  15. [15]

    https://arxiv.org/abs/2311.15127

    Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., Jampani, V., Rombach, R.: Stable Video Diffusion: Scaling Latent Video Diffusion Models to Large Datasets (2023). https://arxiv.org/abs/2311.15127

    Pith/arXiv arXiv 2023

  16. [16]

    International Conference on Learning Representations (2024)

    Guo, Y., Yang, C., Rao, A., Liang, Z., Wang, Y., Qiao, Y., Agrawala, M., Lin, D., Dai, B.: Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. International Conference on Learning Representations (2024)

    2024

  17. [17]

    https://arxiv

    Xue, J., Wang, H., Tian, Q., Ma, Y., Wang, A., Zhao, Z., Min, S., Zhao, W., Zhang, K., Shum, H.-Y., Liu, W., Liu, M., Luo, W.: Towards Multiple Character Image Animation Through Enhancing Implicit Decoupling (2025). https://arxiv. org/abs/2406.03035

    Pith/arXiv arXiv 2025

  18. [18]

    https://arxiv

    Wang, Z., Li, Y., Zeng, Y., Guo, Y., Lin, D., Xue, T., Dai, B.: Multi-identity Human Image Animation with Structural Video Diffusion (2025). https://arxiv. org/abs/2504.04126

    arXiv 2025

  19. [19]

    https://arxiv.org/abs/1808.07371

    Chan, C., Ginosar, S., Zhou, T., Efros, A.A.: Everybody Dance Now (2019). https://arxiv.org/abs/1808.07371

    Pith/arXiv arXiv 2019

  20. [20]

    https://arxiv.org/abs/2203.02910

    Zhang, P., Yang, L., Lai, J., Xie, X.: Exploring Dual-task Correlation for Pose Guided Person Image Generation (2022). https://arxiv.org/abs/2203.02910

    Pith/arXiv arXiv 2022

  21. [21]

    https://arxiv.org/abs/2203.14367

    Zhao, J., Zhang, H.: Thin-Plate Spline Motion Model for Image Animation (2022). https://arxiv.org/abs/2203.14367

    Pith/arXiv arXiv 2022

  22. [22]

    https://arxiv.org/abs/2104.11280

    Siarohin, A., Woodford, O.J., Ren, J., Chai, M., Tulyakov, S.: Motion Represen- tations for Articulated Animation (2021). https://arxiv.org/abs/2104.11280

    Pith/arXiv arXiv 2021

  23. [23]

    International Journal of Computer Vision128(5), 1398–1413 (2020)

    Vougioukas, K., Petridis, S., Pantic, M.: Realistic speech-driven facial animation with gans. International Journal of Computer Vision128(5), 1398–1413 (2020)

    2020

  24. [24]

    International Journal of Computer Vision128(3), 698–713 (2020)

    Pumarola, A., Agudo, A., Martinez, A.M., Sanfeliu, A., Moreno-Noguer, F.: Gan- imation: One-shot anatomically consistent facial animation. International Journal of Computer Vision128(3), 698–713 (2020)

    2020

  25. [25]

    https://arxiv.org/abs/ 2112.10752 22

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-Resolution Image Synthesis with Latent Diffusion Models (2022). https://arxiv.org/abs/ 2112.10752 22

    Pith/arXiv arXiv 2022

  26. [26]

    https://arxiv.org/abs/2410.10306

    Tan, S., Gong, B., Wang, X., Zhang, S., Zheng, D., Zheng, R., Zheng, K., Chen, J., Yang, M.: Animate-X: Universal Character Image Animation with Enhanced Motion Representation (2024). https://arxiv.org/abs/2410.10306

    Pith/arXiv arXiv 2024

  27. [27]

    International Journal of Computer Vision133(3), 1277–1289 (2025)

    Wang, Y., Ma, X., Chen, X., Chen, C., Dantcheva, A., Dai, B., Qiao, Y.: Leo: Gen- erative latent image animator for human video synthesis. International Journal of Computer Vision133(3), 1277–1289 (2025)

    2025

  28. [28]

    International Journal of Computer Vision, 1–17 (2025)

    Hu, Y., Chen, Z., Luo, C.: Lamd: Latent motion diffusion for image-conditional video generation. International Journal of Computer Vision, 1–17 (2025)

    2025

  29. [29]

    https://arxiv.org/abs/2312.00752

    Gu, A., Dao, T.: Mamba: Linear-Time Sequence Modeling with Selective State Spaces (2024). https://arxiv.org/abs/2312.00752

    Pith/arXiv arXiv 2024

  30. [30]

    In: Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pp

    Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Smpl: A skinned multi-person linear model. In: Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pp. 851–866 (2023)

    2023

  31. [31]

    https://arxiv.org/abs/2409.16160

    Men, Y., Yao, Y., Cui, M., Bo, L.: MIMO: Controllable Character Video Synthesis with Spatial Decomposed Modeling (2024). https://arxiv.org/abs/2409.16160

    Pith/arXiv arXiv 2024

  32. [32]

    https://arxiv

    Kwon, G., Jenni, S., Li, D., Lee, J.-Y., Ye, J.C., Heilbron, F.C.: Concept Weaver: Enabling Multi-Concept Fusion in Text-to-Image Models (2024). https://arxiv. org/abs/2404.03913

    Pith/arXiv arXiv 2024

  33. [33]

    https://arxiv.org/abs/2212

    Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.-Y.: Multi-Concept Customization of Text-to-Image Diffusion (2023). https://arxiv.org/abs/2212. 04488

    2023

  34. [34]

    https://arxiv.org/abs/2404

    Jang, S., Jo, J., Lee, K., Hwang, S.J.: Identity Decoupling for Multi-Subject Personalization of Text-to-Image Models (2024). https://arxiv.org/abs/2404. 04243

    2024

  35. [35]

    In: SIGGRAPH Asia 2023 Conference Papers

    Avrahami, O., Aberman, K., Fried, O., Cohen-Or, D., Lischinski, D.: Break-a- scene: Extracting multiple concepts from a single image. In: SIGGRAPH Asia 2023 Conference Papers. SA ’23, pp. 1–12. ACM, ??? (2023). https://doi.org/10. 1145/3610548.3618154 .http://dx.doi.org/10.1145/3610548.3618154

    arXiv 2023

  36. [36]

    https: //arxiv.org/abs/2305.10431

    Xiao, G., Yin, T., Freeman, W.T., Durand, F., Han, S.: FastComposer: Tuning- Free Multi-Subject Image Generation with Localized Attention (2023). https: //arxiv.org/abs/2305.10431

    Pith/arXiv arXiv 2023

  37. [37]

    https://arxiv.org/abs/2404.11565 23

    Wang, K.-C., Ostashev, D., Fang, Y., Tulyakov, S., Aberman, K.: MoA: Mixture-of-Attention for Subject-Context Disentanglement in Personalized Image Generation (2024). https://arxiv.org/abs/2404.11565 23

    Pith/arXiv arXiv 2024

  38. [38]

    https://arxiv.org/abs/ 2404.19427

    Kim, C., Lee, J., Joung, S., Kim, B., Baek, Y.-M.: InstantFamily: Masked Atten- tion for Zero-shot Multi-ID Image Generation (2024). https://arxiv.org/abs/ 2404.19427

    Pith/arXiv arXiv 2024

  39. [39]

    https://arxiv

    He, J., Geng, Y., Bo, L.: UniPortrait: A Unified Framework for Identity- Preserving Single- and Multi-Human Image Personalization (2024). https://arxiv. org/abs/2408.05939

    Pith/arXiv arXiv 2024

  40. [40]

    International Journal of Computer Vision133(3), 1175–1194 (2025)

    Xiao, G., Yin, T., Freeman, W.T., Durand, F., Han, S.: Fastcomposer: Tuning-free multi-subject image generation with localized attention. International Journal of Computer Vision133(3), 1175–1194 (2025)

    2025

  41. [41]

    zhao et al

    Zhao, C., Ding, G., Wang, W., Yang, Z., Liu, Z., Chen, H., Shen, C.: Freercustom: Training-free multi-concept customization for image and video generation: C. zhao et al. International Journal of Computer Vision134(1), 17 (2026)

    2026

  42. [42]

    https://arxiv.org/abs/2305.18292

    Gu, Y., Wang, X., Wu, J.Z., Shi, Y., Chen, Y., Fan, Z., Xiao, W., Zhao, R., Chang, S., Wu, W., Ge, Y., Shan, Y., Shou, M.Z.: Mix-of-Show: Decentralized Low- Rank Adaptation for Multi-Concept Customization of Diffusion Models (2023). https://arxiv.org/abs/2305.18292

    Pith/arXiv arXiv 2023

  43. [43]

    https:// arxiv.org/abs/2502.07802

    Liang, F., Ma, H., He, Z., Hou, T., Hou, J., Li, K., Dai, X., Juefei-Xu, F., Azadi, S., Sinha, A., Zhang, P., Vajda, P., Marculescu, D.: Movie Weaver: Tuning-Free Multi-Concept Video Personalization with Anchored Prompts (2025). https:// arxiv.org/abs/2502.07802

    Pith/arXiv arXiv 2025

  44. [44]

    https://arxiv.org/abs/1505.04597

    Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional Networks for Biomedical Image Segmentation (2015). https://arxiv.org/abs/1505.04597

    Pith/arXiv arXiv 2015

  45. [45]

    https:// arxiv.org/abs/1312.6114

    Kingma, D.P., Welling, M.: Auto-Encoding Variational Bayes (2022). https:// arxiv.org/abs/1312.6114

    Pith/arXiv arXiv 2022

  46. [46]

    https://arxiv.org/abs/1711.00937

    Oord, A., Vinyals, O., Kavukcuoglu, K.: Neural Discrete Representation Learning (2018). https://arxiv.org/abs/1711.00937

    Pith/arXiv arXiv 2018

  47. [47]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp

    Yang, Z., Zeng, A., Yuan, C., Li, Y.: Effective whole-body pose estimation with two-stages distillation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 4210–4220 (2023)

    2023

  48. [48]

    Natalia Neverova: Densepose: Dense human pose estimation in the wild

    Riza Alp G”uler, I.K. Natalia Neverova: Densepose: Dense human pose estimation in the wild. (2018)

    2018

  49. [49]

    https://arxiv

    Cao, Z., Hidalgo, G., Simon, T., Wei, S.-E., Sheikh, Y.: OpenPose: Realtime Multi-Person 2D Pose Estimation using Part Affinity Fields (2019). https://arxiv. org/abs/1812.08008

    Pith/arXiv arXiv 2019

  50. [50]

    arXiv preprint arXiv:2408.00714 (2024)

    Ravi, N., Gabeur, V., Hu, Y.-T., Hu, R., Ryali, C., Ma, T., Khedr, H., R¨ adle, 24 R., Rolland, C., Gustafson, L., Mintun, E., Pan, J., Alwala, K.V., Carion, N., Wu, C.-Y., Girshick, R., Doll´ ar, P., Feichtenhofer, C.: Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714 (2024)

    Pith/arXiv arXiv 2024

  51. [51]

    arXiv preprint arXiv:2107.08430 (2021)

    Ge, Z., Liu, S., Wang, F., Li, Z., Sun, J.: Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430 (2021)

    Pith/arXiv arXiv 2021

  52. [52]

    IEEE Transactions on Image Processing13(4), 600–612 (2004) https://doi.org/10.1109/TIP.2003.819861

    Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assess- ment: from error visibility to structural similarity. IEEE Transactions on Image Processing13(4), 600–612 (2004) https://doi.org/10.1109/TIP.2003.819861

    arXiv 2004

  53. [53]

    Hor´ e, A., Ziou, D.: Image quality metrics: Psnr vs. ssim. In: 2010 20th Inter- national Conference on Pattern Recognition, pp. 2366–2369 (2010). https://doi. org/10.1109/ICPR.2010.579

    doi:10.1109/icpr.2010.579 2010

  54. [54]

    https://arxiv.org/ abs/1801.03924

    Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The Unreasonable Effectiveness of Deep Features as a Perceptual Metric (2018). https://arxiv.org/ abs/1801.03924

    Pith/arXiv arXiv 2018

  55. [55]

    https://arxiv.org/abs/1706.08500

    Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium (2018). https://arxiv.org/abs/1706.08500

    Pith/arXiv arXiv 2018

  56. [56]

    https://arxiv.org/abs/1812.01717

    Unterthiner, T., Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., Gelly, S.: Towards Accurate Generative Models of Video: A New Metric & Challenges (2019). https://arxiv.org/abs/1812.01717

    Pith/arXiv arXiv 2019

  57. [57]

    https://arxiv.org/abs/2506.15742

    Labs, B.F., Batifol, S., Blattmann, A., Boesel, F., Consul, S., Diagne, C., Dock- horn, T., English, J., English, Z., Esser, P., Kulal, S., Lacey, K., Levi, Y., Li, C., Lorenz, D., M¨ uller, J., Podell, D., Rombach, R., Saini, H., Sauer, A., Smith, L.: FLUX.1 Kontext: Flow Matching for In-Context Image Generation and Editing in Latent Space (2025). https:...

    Pith/arXiv arXiv 2025

  58. [58]

    arXiv preprint arXiv:2503.20314 (2025)

    Wang, A., Ai, B., Wen, B., Mao, C., Xie, C.-W., Chen, D., Yu, F., Zhao, H., Yang, J., Zeng, J., Wang, J., Zhang, J., Zhou, J., Wang, J., Chen, J., Zhu, K., Zhao, K., Yan, K., Huang, L., Feng, M., Zhang, N., Li, P., Wu, P., Chu, R., Feng, R., Zhang, S., Sun, S., Fang, T., Wang, T., Gui, T., Weng, T., Shen, T., Lin, W., Wang, W., Wang, W., Zhou, W., Wang, W...

    Pith/arXiv arXiv 2025

  59. [59]

    https://arxiv.org/abs/2503.16153 25

    Wei, T., Zhou, Y., Chen, D., Pan, X.: FreeFlux: Understanding and Exploiting Layer-Specific Roles in RoPE-Based MMDiT for Versatile Image Editing (2025). https://arxiv.org/abs/2503.16153 25

    Pith/arXiv arXiv 2025

  60. [60]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp

    Jafarian, Y., Park, H.S.: Learning high fidelity depths of dressed humans by watching social media dance videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 12753–12762 (2021)

    2021

  61. [61]

    arXiv preprint arXiv:2412.03603 (2024)

    Kong, W., Tian, Q., Zhang, Z., Min, R., Dai, Z., Zhou, J., Xiong, J., Li, X., Wu, B., Zhang, J., et al.: Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603 (2024)

    Pith/arXiv arXiv 2024

  62. [62]

    https://arxiv.org/abs/2408.06072

    Yang, Z., Teng, J., Zheng, W., Ding, M., Huang, S., Xu, J., Yang, Y., Hong, W., Zhang, X., Feng, G., Yin, D., Zhang, Y., Wang, W., Cheng, Y., Xu, B., Gu, X., Dong, Y., Tang, J.: CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer (2025). https://arxiv.org/abs/2408.06072

    Pith/arXiv arXiv 2025

  63. [63]

    arXiv preprint arXiv:2212.09748 (2022) 26

    Peebles, W., Xie, S.: Scalable diffusion models with transformers. arXiv preprint arXiv:2212.09748 (2022) 26

    Pith/arXiv arXiv 2022