Pith. sign in

REVIEW 3 major objections 6 minor 65 references

Aura generates multi-subject videos that keep each identity stable by aligning a vision-language model to a diffusion transformer.

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 · grok-4.5

2026-07-11 20:10 UTC pith:3XTUKCEQ

load-bearing objection Solid multi-subject video systems paper with real engineering depth; Total-score SOTA is real, but identity-centric metrics do not fully back the strongest consistency rhetoric. the 3 major comments →

arxiv 2607.04311 v2 pith:3XTUKCEQ submitted 2026-07-05 cs.CV

Aura: Consistent Multi-Subject Video Generation via VLM-Grounded Semantic Alignment

Zixiang Zhou , Zhentao Yu , Yifeng Ma , Hongmei Wang , Wenqing Yu , Cong Wang , Zilin Yang , Rui Chen
show 4 more authors
Jiarong Ou Yezhou Liu Yuan Zhou Qinglin Lu
This is my paper
classification cs.CV
keywords subject-to-videomulti-subject consistencydiffusion transformervision-language model alignmentRoPE shiftidentity preservationcontrollable video generation
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.

Aura claims that identity-consistent video generation with multiple heterogeneous subjects—people, objects, and scenes—is possible inside one diffusion transformer when structured director-style captions and VLM features are carefully fused with reference images. References are token-concatenated into the transformer; category-specific learnable tokens and a subject-aware rotary position shift keep humans, objects, scenes, and padding slots from colliding. A two-stage alignment places frozen VLM meta-queries onto the pretrained text-encoder manifold so both streams can share cross-attention without destroying prompt following. A four-stage curriculum and norm-only progressive guidance finish the system, trained on a large AIGC-curated multi-reference set. Readers who need controllable clips for pre-visualization, product ads, or personalized stories care because earlier methods either paste references rigidly or lose identity once several subjects interact.

Core claim

The paper establishes that a single diffusion transformer can jointly handle pure text-to-video and multi-reference subject-to-video once VLM meta-queries are mapped onto the text-encoder manifold by sentence-level asymmetric InfoNCE plus token-level Hungarian matching. That alignment makes parameter-free shared-KV cross-attention feasible, so multimodal semantics improve binding without overwriting the pretrained text prior, while subject-aware tokens and RoPE shifts stop cross-category identity leakage.

What carries the argument

Subject-aware RoPE-Shift plus dual-stream T5–VLM conditioning: per-category learnable tokens and hard-coded rotary offsets place human, object, scene, and memory references into mutually disjoint quadrants of the 3D rotary grid, while aligned VLM features supply fine-grained multimodal cues through shared cross-attention.

Load-bearing premise

The method assumes that aligning a frozen vision-language model to the text encoder after the diffusion backbone was already pretrained is enough to keep multi-subject binding and prompt following from fighting each other.

What would settle it

If ablating the Hungarian token-matching term or the subject-aware RoPE shift still yields equal or higher OpenS2V-Eval Total and NaturalScore on the multi-element test set under the same backbone and data, the claimed necessity of those two mechanisms would be falsified.

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

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 / 6 minor

Summary. Aura is a unified DiT-based framework for reference-conditioned single- and multi-subject video generation. It combines (i) token-concat reference injection with asymmetric clean-timestep embeddings, per-category learnable tokens, Memory Tokens, and a Subject-Aware RoPE-Shift; (ii) dual-stream conditioning that fuses frozen T5 with meta-query features from frozen Qwen2.5-VL via T5-teacher alignment (asymmetric InfoNCE + Hungarian matching) and parameter-free shared-KV cross-attention; (iii) a four-stage Coarse-Align → Fine-Align → Ref-Only → Joint-Mix curriculum; (iv) norm-only progressive dual-axis APG at inference; and (v) a large (~15M) AIGC grounding–augmenting–verification pipeline with MTSS director-style captions. On OpenS2V-Eval the method reports the best Total score (61.01) versus Wan2.7, HuMo, Kaleido, MAGREF, and RefAlign, with supporting ablations (Tables 2–3), a VLM-judge protocol, and a GSB user study.

Significance. If the multi-element identity and controllability claims hold under scrutiny, Aura is a substantial systems contribution: it unifies heterogeneous human/object/scene references in one DiT, supplies a concrete language-side alignment recipe that enables shared-KV fusion without new cross-attention weights, and documents a scalable AIGC multi-reference data pipeline. The ablations on dual-stream conditioning, curriculum stages, Hungarian matching, and norm-only APG (Tables 2–3; Appendix B–C), plus released project/code links, are genuine strengths. The work is of clear interest to the controllable video community even if the headline “identity-consistent SOTA” phrasing needs to be aligned more carefully with the identity-centric metrics.

major comments (3)
  1. Abstract and §4.4 claim state-of-the-art identity-consistent multi-element generation, with Total=61.01 as the primary quantitative support. Official OpenS2V-Eval weighting places 0.20 on FaceSim-Cur and 0.20 on NexusScore. In Table 1, Aura’s FaceSim-Cur is 38.50 (behind Wan 59.62, RefAlign 45.66, HuMo 40.18) and NexusScore is 71.30 (behind RefAlign 84.48, Kaleido 82.14, MAGREF 79.45, Wan 78.98). The Total lead is driven largely by NaturalScore (67.50), AES, and MotionAmp. The manuscript must either (a) reframe the central claim to match which axes actually improve, or (b) provide a clear analysis of why identity-centric metrics lag while VLM-judge Subject/HardCopy and the GSB study favor Aura (e.g., hard-copy inflation of FaceSim/Nexus for baselines, multi-subject Hungarian matching effects, or style/motion trade-offs). Without that reconciliation, the multi-element identity SOTA claim
  2. §3.3–3.4 and Limitations H rest on the assumption that post-hoc T5-teacher alignment of frozen Qwen2.5-VL meta-queries is sufficient for shared-KV fusion on a backbone never co-trained with a VLM stream. Table 3 shows Hungarian matching is critical in the Stage-1 plug-in protocol, and Table 2(a) shows removing the VLM branch hurts aesthetics/naturalness more than narrow text metrics. However, there is no multi-subject stress test that isolates residual attribute leakage / identity swap under long compositional MTSS prompts (the failure mode named in H). A load-bearing addition would be a controlled multi-entity binding evaluation (e.g., attribute-swap or role-assignment accuracy) comparing dual-stream vs. T5-only vs. VLM-only under matched reference counts, so that the claim that alignment improves binding rather than trading off prompt following is measured directly.
  3. §4.1’s 50 hand-crafted multi-category cases and the MTSS caption style used both for training (§3.6, Appendix A) and for VLM-judge evaluation (§E) create a mild distribution match that is not fully external. OpenS2V-Eval is the right primary benchmark, but the paper should report per-axis OpenS2V results stratified by single- vs multi-reference and by human/object/scene composition, and clarify how much of the VLM-judge and GSB advantage depends on MTSS-style prompts versus plain user prompts. Without stratification, the “more challenging multi-element scenarios” claim remains only partially evidenced.
minor comments (6)
  1. Table 1: MotionSmooth for Ours (88.21) is below Wan/HuMo/MAGREF; a short discussion of the motion-smoothness vs. MotionAmp trade-off would help readers interpret the Total score.
  2. §3.2 Eq. (2): the concrete numerical values of the per-category RoPE shifts Δ(c) are not stated; listing them (or the quadrant policy) would aid reproducibility.
  3. §3.3: N_q (number of meta-queries), Enc_φ depth, and λ_NCE / λ_Hun / τ are free parameters; please report the chosen values in §4.2.
  4. Figure 4 qualitative panel is dense; labeling which references are human/object/scene and which failure modes (swap, copy-paste, drift) each baseline exhibits would make the comparison easier to read.
  5. Related work §2.3 positions Aura against RefAlign’s vision-only alignment; a one-sentence note that language-side T5-teacher alignment is complementary (not a replacement) would reduce possible over-claim of novelty relative to concurrent VFM alignment work.
  6. Typos / consistency: “AI director-level” vs “AI director-style”; “Progressive-APG” vs “norm-only progressive APG”; arXiv id in the user message (2607.04311) should match the manuscript header when camera-ready.

Circularity Check

0 steps flagged

No circular derivation: Aura is an empirical systems paper whose SOTA claim rests on external OpenS2V-Eval and ablations, not on results forced by definition or self-cited uniqueness.

full rationale

Aura proposes architectural and training components (token-concat with subject-aware RoPE-Shift and category tokens, dual-stream T5–VLM conditioning with InfoNCE+Hungarian alignment, four-stage curriculum, norm-only progressive APG, AIGC data pipeline) and evaluates them empirically. There is no first-principles derivation chain in which a claimed prediction or uniqueness result reduces to its own inputs by construction. Primary quantitative claims use the external OpenS2V-Eval suite (Yuan et al., 2025a) with fixed official weights, plus comparisons to independent baselines under their recommended settings. Ablations (Tables 2–3) remove named modules and report metric changes rather than redefining the target. Progressive-APG is an empirical simplification of Sadat et al. (2024) justified by logged guidance-norm statistics, not a tautology. Self-citations (HunyuanCustom, PolyVivid, MTSS/Team 2026) describe related prior systems or caption format; none supply a load-bearing uniqueness theorem that forbids alternatives. Author-built 50-case / Gemma-4 protocol and MTSS captions used at train and eval time are evaluation-design choices, not circular reductions of a derived quantity to a fitted input. Metric tension (strong Total/NaturalScore vs. weaker FaceSim-Cur/NexusScore) is a claim–evidence gap, not circularity. Score 0 is therefore appropriate.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 5 invented entities

Aura’s claims rest on standard diffusion/DiT and VLM machinery plus many hand-chosen architectural and data hyperparameters. No new physical law is postulated; the invented entities are engineering modules whose value is only shown inside this training/eval loop. Free parameters (slot count, RoPE offsets, alignment weights, APG caps, filter thresholds) are load-bearing for the reported quality/identity trade-off.

free parameters (7)
  • Fixed reference slot count K = 6
    K=6 memory-padded slots keep sequence length fixed; chosen by design, not derived.
  • Subject-aware RoPE category offsets Δ(c) = hand-coded per-category (Δt,Δh,Δw)
    Hard-coded temporal/spatial quadrant shifts for human/object/scene/memory; geometry is stipulated, not learned from first principles.
  • Alignment loss weights λ_NCE, λ_Hun and temperature τ
    Balance sentence InfoNCE vs Hungarian token regression; values control whether shared-KV is feasible.
  • Norm-only progressive APG caps κ_s(t) and weights w_s = schedule-dependent; probe-tuned
    Linear schedule (probe-motivated ~50→15) and per-axis weights; Appendix B shows static caps fail; endpoints are tuned on a probe set.
  • Identity filter thresholds τ_face, τ_obj, τ_scn
    ArcFace/BLIP-2 cutoffs decide which edited references enter the 15M set; directly shape the identity–hard-copy trade-off.
  • AdamW learning rate and warmup = 1e-6, 500-step warmup
    1e-6 constant after 500-step warmup; standard but free training choice affecting final metrics.
  • Number of VLM meta-queries N_q and Enc_φ depth = Enc_φ = 8 layers; N_q unspecified numeric
    Fixes VLM interface length and refinement capacity; architectural free choice.
axioms (5)
  • domain assumption Flow-matching / diffusion DiT pretrained on large video data (Wan2.2) already encodes a usable T2V prior that can absorb reference tokens via self-attention without destroying motion.
    Entire token-concat design and Joint-Mix stage assume the backbone prior remains intact (§3.2, §3.4).
  • ad hoc to paper Frozen Qwen2.5-VL last-layer meta-query states, after MLP projection, can be aligned onto the T5 token manifold well enough for parameter-free shared KV cross-attention.
    Core of dual-stream design (§3.3); Limitations admit residual post-hoc misalignment.
  • ad hoc to paper Per-category feature tokens plus disjoint 3D RoPE quadrants are sufficient to prevent same- and cross-category identity collisions on the concatenated sequence.
    Stated as the dual identity signal; ablations claim residual cross-talk if either is removed (§3.2).
  • domain assumption VLM-guided I2I edits that pass ArcFace/BLIP-2 thresholds break hard-copy shortcuts without destroying identity supervision.
    Data pipeline premise (§3.6, Appendix D); Limitations note off-manifold edited references can still hurt stability.
  • domain assumption OpenS2V-Eval official weighted Total and the authors’ Gemma-4 rubrics are adequate proxies for multi-subject fidelity and naturalness.
    All SOTA claims route through these metrics (§4.3–4.4).
invented entities (5)
  • Subject-aware RoPE-Shift no independent evidence
    purpose: Place human/object/scene/memory reference tokens in disjoint rotary quadrants to reduce collisions and copy-paste.
    Architectural construct specific to this paper’s multi-category concat design; evidence is internal ablations only.
  • Memory Tokens no independent evidence
    purpose: Pad variable reference counts to fixed K so training signal and sequence budget stay balanced.
    Learnable padding category; no external falsifiable prediction beyond training convenience.
  • T5-teacher dual-stream shared-KV pathway with Hungarian+InfoNCE alignment no independent evidence
    purpose: Inject VLM multimodal semantics without adding new cross-attention weights or destroying T5 priors.
    Method invention; plug-in Table 3 supports Hungarian term inside this setup only.
  • Norm-only progressive dual-axis APG no independent evidence
    purpose: Stabilize high dual-CFG guidance by clipping guidance norms on a schedule without parallel/orthogonal split.
    Simplification of Sadat APG motivated by authors’ probe statistics; not independently validated outside Aura.
  • MTSS AI-director multi-stream captions no independent evidence
    purpose: Provide structured, relationally grounded scripts linking PERSON/OBJECT/SCENE ids across shots.
    Caption schema from related Team (2026) work used as control interface; value shown inside Aura training/eval.

pith-pipeline@v1.1.0-grok45 · 30091 in / 4384 out tokens · 49587 ms · 2026-07-11T20:10:06.924102+00:00 · methodology

0 comments
read the original abstract

Subject-driven and multi-element video generation are central to controllable video synthesis, but existing methods still struggle to preserve identity consistency and model complex relationships among multiple subjects. In this paper, we propose Aura, a unified framework for high-fidelity and identity-consistent video generation. To better capture scene dynamics and subject interactions, we introduce AI director-level captions that provide dense and structured descriptions of video content. We further leverage a vision-language model (VLM) with learnable queries to extract multimodal semantic features from textual and visual references, covering both global semantics and fine-grained visual cues. To bridge the representational gap between the VLM and the Diffusion Transformer (DiT), we design a two-stage alignment strategy that progressively maps VLM features into the DiT feature space. For visual conditioning, we adopt token concatenation to inject reference information directly into the generation process. To distinguish heterogeneous subject types and reduce common copy-paste artifacts, we develop a subject-aware RoPE-Shift mechanism. To further differentiate reference images of different categories, we introduce subject-aware learnable tokens. In addition, we introduce Memory Tokens to balance the training signal across examples with different numbers of reference subjects. During inference, Progressive-APG (Adaptive Prompt Guidance) further alleviates oversaturation and improves semantic alignment with user prompts. Finally, we build a high-quality video-subject image dataset through a dedicated data construction pipeline. Extensive experiments show that our method achieves state-of-the-art performance on both single-subject generation and more challenging multi-element scenarios.

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

65 extracted references · 23 linked inside Pith

  1. [1]

    arXiv preprint arXiv:2503.20314 , year=

    Wan: Open and advanced large-scale video generative models , author=. arXiv preprint arXiv:2503.20314 , year=

    Pith/arXiv arXiv

  2. [2]

    arXiv preprint arXiv:2412.03603 , year=

    Hunyuanvideo: A systematic framework for large video generative models , author=. arXiv preprint arXiv:2412.03603 , year=

    Pith/arXiv arXiv

  3. [3]

    arXiv preprint arXiv:2205.15868 , year=

    Cogvideo: Large-scale pretraining for text-to-video generation via transformers , author=. arXiv preprint arXiv:2205.15868 , year=

    Pith/arXiv arXiv

  4. [4]

    https://openai.com/sora , year=

    Sora: Creating video from text , author=. https://openai.com/sora , year=

  5. [5]

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

    Phantom: Subject-consistent video generation via cross-modal alignment , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  6. [6]

    arXiv preprint arXiv:2510.00438 , year=

    Bindweave: Subject-consistent video generation via cross-modal integration , author=. arXiv preprint arXiv:2510.00438 , year=

    arXiv

  7. [7]

    arXiv preprint arXiv:2603.25743 , year=

    RefAlign: Representation Alignment for Reference-to-Video Generation , author=. arXiv preprint arXiv:2603.25743 , year=

    Pith/arXiv arXiv

  8. [8]

    2026 , eprint=

    DreamID-Omni: Unified Framework for Controllable Human-Centric Audio-Video Generation , author=. 2026 , eprint=

    2026

  9. [9]

    arXiv preprint arXiv:2504.02436 , year=

    SkyReels-A2: Compose Anything in Video Diffusion Transformers , author=. arXiv preprint arXiv:2504.02436 , year=

    Pith/arXiv arXiv

  10. [10]

    arXiv preprint arXiv:2509.08519 , year=

    Humo: Human-centric video generation via collaborative multi-modal conditioning , author=. arXiv preprint arXiv:2509.08519 , year=

    Pith/arXiv arXiv

  11. [11]

    arXiv preprint arXiv:2505.04512 , year=

    Hunyuancustom: A multimodal-driven architecture for customized video generation , author=. arXiv preprint arXiv:2505.04512 , year=

    Pith/arXiv arXiv

  12. [12]

    arXiv preprint arXiv:2510.18573 , year=

    Kaleido: Open-Sourced Multi-Subject Reference Video Generation Model , author=. arXiv preprint arXiv:2510.18573 , year=

    arXiv

  13. [13]

    2025 , eprint=

    ID-Crafter: VLM-Grounded Online RL for Compositional Multi-Subject Video Generation , author=. 2025 , eprint=

    2025

  14. [14]

    arXiv preprint arXiv:2603.00586 , year=

    WildActor: Unconstrained Identity-Preserving Video Generation , author=. arXiv preprint arXiv:2603.00586 , year=

    arXiv

  15. [15]

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

    Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  16. [16]

    arXiv preprint arXiv:2208.01618 , year=

    An image is worth one word: Personalizing text-to-image generation using textual inversion , author=. arXiv preprint arXiv:2208.01618 , year=

    Pith/arXiv arXiv

  17. [17]

    arXiv preprint arXiv:2308.06721 , year=

    Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models , author=. arXiv preprint arXiv:2308.06721 , year=

    Pith/arXiv arXiv

  18. [18]

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

    Photomaker: Customizing realistic human photos via stacked id embedding , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  19. [19]

    arXiv preprint arXiv:2404.15275 , year=

    Id-animator: Zero-shot identity-preserving human video generation , author=. arXiv preprint arXiv:2404.15275 , year=

    Pith/arXiv arXiv

  20. [20]

    IEEE Transactions on Multimedia , year=

    Customvideo: Customizing text-to-video generation with multiple subjects , author=. IEEE Transactions on Multimedia , year=

  21. [21]

    Proceedings of the 32nd ACM International Conference on Multimedia , pages=

    Disenstudio: Customized multi-subject text-to-video generation with disentangled spatial control , author=. Proceedings of the 32nd ACM International Conference on Multimedia , pages=

  22. [22]

    arXiv preprint arXiv:2508.07901 , year=

    Stand-In: A Lightweight and Plug-and-Play Identity Control for Video Generation , author=. arXiv preprint arXiv:2508.07901 , year=

    arXiv

  23. [23]

    arXiv e-prints , pages=

    Motioncharacter: Identity-preserving and motion controllable human video generation , author=. arXiv e-prints , pages=

  24. [24]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Hallo3: Highly dynamic and realistic portrait image animation with video diffusion transformer , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  25. [25]

    Proceedings of the 33rd ACM International Conference on Multimedia , pages=

    Fantasytalking: Realistic talking portrait generation via coherent motion synthesis , author=. Proceedings of the 33rd ACM International Conference on Multimedia , pages=

  26. [26]

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

    Omnihuman-1: Rethinking the scaling-up of one-stage conditioned human animation models , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  27. [27]

    arXiv preprint arXiv:2506.09984 , year=

    Interacthuman: Multi-concept human animation with layout-aligned audio conditions , author=. arXiv preprint arXiv:2506.09984 , year=

    arXiv

  28. [28]

    arXiv preprint arXiv:2510.14179 , year=

    Virtually Being: Customizing Camera-Controllable Video Diffusion Models with Multi-View Performance Captures , author=. arXiv preprint arXiv:2510.14179 , year=

    arXiv

  29. [29]

    arXiv preprint arXiv:2505.20292 , year=

    Opens2v-nexus: A detailed benchmark and million-scale dataset for subject-to-video generation , author=. arXiv preprint arXiv:2505.20292 , year=

    Pith/arXiv arXiv

  30. [30]

    arXiv preprint arXiv:2501.04698 , year=

    Conceptmaster: Multi-concept video customization on diffusion transformer models without test-time tuning , author=. arXiv preprint arXiv:2501.04698 , year=

    Pith/arXiv arXiv

  31. [31]

    arXiv preprint arXiv:2505.23742 , year=

    MAGREF: Masked Guidance for Any-Reference Video Generation with Subject Disentanglement , author=. arXiv preprint arXiv:2505.23742 , year=

    arXiv

  32. [32]

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

    Vace: All-in-one video creation and editing , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  33. [33]

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

    Concat-ID: Towards Universal Identity-Preserving Video Synthesis , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  34. [34]

    arXiv preprint arXiv:2503.10391 , year=

    Cinema: Coherent multi-subject video generation via mllm-based guidance , author=. arXiv preprint arXiv:2503.10391 , year=

    Pith/arXiv arXiv

  35. [35]

    arXiv preprint arXiv:2507.12956 , year=

    Fantasyportrait: Enhancing multi-character portrait animation with expression-augmented diffusion transformers , author=. arXiv preprint arXiv:2507.12956 , year=

    Pith/arXiv arXiv

  36. [36]

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

    Videomage: Multi-subject and motion customization of text-to-video diffusion models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  37. [37]

    arXiv preprint arXiv:2506.07848 , year=

    PolyVivid: Vivid Multi-Subject Video Generation with Cross-Modal Interaction and Enhancement , author=. arXiv preprint arXiv:2506.07848 , year=

    Pith/arXiv arXiv

  38. [38]

    2025 , booktitle=

    Representation Alignment for Generation: Training Diffusion Transformers Is Easier Than You Think , author=. 2025 , booktitle=

    2025

  39. [39]

    2025 , journal=

    REPA-E: Unlocking VAE for End-to-End Tuning with Latent Diffusion Transformers , author=. 2025 , journal=

    2025

  40. [40]

    arXiv preprint arXiv:2504.05741 , year=

    DDT: Decoupled Diffusion Transformer , author=. arXiv preprint arXiv:2504.05741 , year=

    Pith/arXiv arXiv

  41. [41]

    arXiv preprint arXiv:2507.15897 , year=

    Redi: Rectified discrete flow , author=. arXiv preprint arXiv:2507.15897 , year=

    arXiv

  42. [42]

    generation: Taming optimization dilemma in latent diffusion models , author=

    Reconstruction vs. generation: Taming optimization dilemma in latent diffusion models , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  43. [43]

    arXiv preprint arXiv:2505.23656 , year=

    VideoREPA: Learning Physics for Video Generation through Relational Alignment with Foundation Models , author=. arXiv preprint arXiv:2505.23656 , year=

    Pith/arXiv arXiv

  44. [44]

    2025 , journal=

    SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features , author=. 2025 , journal=

    2025

  45. [45]

    arXiv preprint arXiv:2210.02747 , year=

    Flow matching for generative modeling , author=. arXiv preprint arXiv:2210.02747 , year=

    Pith/arXiv arXiv

  46. [46]

    The Eleventh International Conference on Learning Representations , year=

    Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow , author=. The Eleventh International Conference on Learning Representations , year=

  47. [47]

    Neurocomputing , volume=

    Roformer: Enhanced transformer with rotary position embedding , author=. Neurocomputing , volume=. 2024 , publisher=

    2024

  48. [48]

    arXiv preprint arXiv:2502.13923 , year=

    Qwen2.5-VL Technical Report , author=. arXiv preprint arXiv:2502.13923 , year=

    Pith/arXiv arXiv

  49. [49]

    The Thirteenth International Conference on Learning Representations , year=

    Eliminating oversaturation and artifacts of high guidance scales in diffusion models , author=. The Thirteenth International Conference on Learning Representations , year=

  50. [50]

    Improved aesthetic predictor , author=

  51. [51]

    2023 , booktitle=

    Exploring Video Quality Assessment on User Generated Contents from Aesthetic and Technical Perspectives , author=. 2023 , booktitle=

    2023

  52. [52]

    CVPR , pages=

    CurricularFace: Adaptive Curriculum Learning Loss for Deep Face Recognition , author=. CVPR , pages=

  53. [53]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Identity-preserving text-to-video generation by frequency decomposition , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  54. [54]

    2024 , eprint=

    GME: Improving Universal Multimodal Retrieval by Multimodal LLMs , author=. 2024 , eprint=

    2024

  55. [55]

    YOLO-World: Real-Time Open-Vocabulary Object Detection , author=. Proc. IEEE Conf. Computer Vision and Pattern Recognition (CVPR) , year=

  56. [56]

    arXiv preprint arXiv:2303.08774 , year=

    Gpt-4 technical report , author=. arXiv preprint arXiv:2303.08774 , year=

    Pith/arXiv arXiv

  57. [57]

    International conference on machine learning , pages=

    Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models , author=. International conference on machine learning , pages=. 2023 , organization=

    2023

  58. [58]

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

    Arcface: Additive angular margin loss for deep face recognition , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  59. [59]

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

    You only look once: Unified, real-time object detection , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  60. [60]

    arXiv preprint arXiv:2506.15742 , year=

    FLUX.1 Kontext: Flow Matching for In-Context Image Generation and Editing in Latent Space , author=. arXiv preprint arXiv:2506.15742 , year=

    Pith/arXiv arXiv

  61. [61]

    arXiv preprint arXiv:2509.22927 , year=

    HunyuanImage 3.0: A Unified Multimodal Model for Native Text-to-Image Generation , author=. arXiv preprint arXiv:2509.22927 , year=

    arXiv

  62. [62]

    arXiv preprint arXiv:2604.11244 , year=

    Script-a-Video: Deep Structured Audio-visual Captions via Factorized Streams and Relational Grounding , author=. arXiv preprint arXiv:2604.11244 , year=

    Pith/arXiv arXiv

  63. [63]

    2026 , howpublished =

    Wan2.7: Next-Generation AI Video Generation Platform , author =. 2026 , howpublished =

    2026

  64. [64]

    arXiv preprint arXiv:2207.12598 , year=

    Classifier-free diffusion guidance , author=. arXiv preprint arXiv:2207.12598 , year=

    Pith/arXiv arXiv

  65. [65]

    2025 , eprint=

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

    2025