Pith. sign in

REVIEW 4 major objections 5 minor 204 references

Persistent world-state tokens make multi-agent video generation more logically consistent

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

2026-08-01 06:57 UTC pith:P2RWSREA

load-bearing objection Persistent world-state registers are a real contribution, but the core consistency numbers rest on an unnamed evaluation VLM that may overlap with the supervision. the 4 major comments →

arxiv 2607.21594 v1 pith:P2RWSREA submitted 2026-07-23 cs.CV

Streaming Multi-Agent Autoregressive Diffusion Model with World State Registers

Sicheng Mo , Yuheng Li , Ziyang Leng , Krishna Kumar Singh , Bolei Zhou This is my paper
classification cs.CV
keywords world state registersmulti-agent video generationstreaming autoregressive diffusionlogical consistencyscene text supervisionbird's-eye viewMixture-of-TransformersMinecraft world model
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.

This paper tries to establish that a video world model can keep a shared world state across multiple agents by carrying a small set of learnable register tokens through the rollout, instead of re-inferring the world from each agent's frame history. The registers are updated after each generated chunk, shared across agents, and supervised with agent positions, overhead bird's-eye views, and scene text. In two-agent Minecraft experiments, the approach improves both visual quality and logical consistency, raising an aggregate world score from 81.0 for a retrained prior baseline to 105.1. The largest VLM-accuracy gains are on state-sensitive categories: grounding, building, and consistency. If correct, the work gives streaming multi-agent world models a concrete mechanism for consistency that is inspectable and does not require unlimited frame memory.

Core claim

WorldWeaver's central claim is that explicit world-state modeling, not just frame-level conditioning, is what lets multi-agent video generation stay logically coherent across views and rollout steps. A group of learnable world-state register tokens is interleaved with frame tokens in a causal transformer, so each committed register precedes and conditions the next frame. The registers are grounded with three supervision signals: per-agent position/velocity/orientation, a shared overhead bird's-eye view compared through a frozen visual encoder, and scene text decoded by a frozen language-model scorer. At inference the prediction heads are discarded, so the extra state costs nothing at rollout

What carries the argument

The world state register (WSR): a fixed set of learnable tokens that persist across agents and rollout steps and are re-committed after each generated chunk. In the interleaved causal layout [frame0, register0, frame1, register1, ...], the latest register acts as a compressed state summary conditioning the next frame, while a sliding window keeps only the recent frame KV cache. A Mixture-of-Transformers gives register tokens their own parameter branch while keeping joint attention with frame tokens. Three training-only decoder heads supervise the register with agent statistics, bird's-eye-view features, and scene text; a self-forcing stage exposes the register to its own rollout drift.

Load-bearing premise

The load-bearing measurement assumption is that VLM accuracy, as reported, is a valid and independent measure of world-state consistency: the scene-text supervision is generated from the same overhead/egocentric views and ground-truth controller inputs that define the spatial relations the evaluation VLM is asked about, and the paper never identifies that evaluator or shows its queries are disjoint.

What would settle it

Run the same evaluation pipeline but replace the VLM judge with a model from a different family than the caption generator, using queries that do not overlap the caption template, and compare against a baseline that conditions on the same scene text without a persistent register. If the advantage over that baseline shrinks or disappears, the gains are better explained by exposure to the supervision distribution than by persistent world-state modeling.

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

If this is right

  • Two or more agents can share one compact state pathway instead of each re-inferring the world from its own partial view, so cross-agent consistency is enforced at the state level.
  • State-sensitive semantic categories — grounding, building, and consistency — improve sharply, suggesting the registers carry verifiable information rather than only improving pixel fidelity.
  • The register supervision heads are dropped at inference, so better consistency does not add per-step rollout cost.
  • Semi-supervised training shows that a small labeled subset can anchor the register semantics while larger unlabeled video still improves rollout quality.
  • Separating the state branch from the frame branch becomes valuable specifically when the register is given rich supervised semantics.

Where Pith is reading between the lines

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

  • The register design could scale to more agents without new pairwise-attention machinery, since the state is shared rather than per-agent-pair.
  • Because the scene-text supervision and the evaluation VLM queries are both built from the same view geometry and controller summaries, the reported consistency gains may in part reflect caption-distribution familiarity; swapping the evaluator to a different family with disjoint queries would test whether the register itself carries the state.
  • The same register mechanism could be applied to any simulator with logged state, or to real-world data where approximate state is available, by replacing the three supervision heads with whatever state annotation exists.
  • A testable extension is to track register prediction accuracy (e.g., agent-coordinate decoding) over a long rollout; if it degrades, the state pathway is not truly persistent.

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

4 major / 5 minor

Summary. WorldWeaver (W^2) proposes a streaming multi-agent autoregressive video diffusion model augmented with world state registers (WSR): a fixed bank of learnable tokens that is shared across agents, updated after each generated chunk, and supervised with agent statistics, bird's-eye-view features, and scene text. A Mixture-of-Transformers (MoT) backbone uses separate parameter branches for register and frame tokens. Training is a three-stage curriculum: bidirectional teacher, causal student with register supervision, and self-forcing with frame and register rollout. On a two-agent Minecraft benchmark, the authors report that W^2 improves an aggregate 'world score' from 81.0 (Solaris baseline) to 105.1, with large VLM-accuracy gains on Grounding, Building, and Consistency. Ablations study the supervision signals, the MoT architecture, and a semi-supervised extension. The central claim is that explicit world-state modeling improves logical consistency and generation quality in multi-agent video generation.

Significance. If the central claim holds, the paper makes a useful architectural contribution: persistent, supervised register tokens provide an interpretable and controllable state pathway for multi-agent streaming world models, going beyond KV-cache history. The three-stage training recipe is concrete, the appendix is unusually detailed about data collection, decoders, and hyperparameters, and the comparison to a same-data retrained Solaris baseline is a fairer design than many prior works. The registers-only row of Table 2 also suggests a benefit that is not entirely dependent on the scene-text supervision. However, the quantitative case is tightly coupled to an un-validated VLM metric, and several ablation comparisons are confounded or lack statistical support. The significance is therefore conditional on resolving the measurement-validity concern.

major comments (4)
  1. [§4.1, §A.4] The primary evidence for 'logical consistency' is VLM accuracy, but the evaluation VLM is never named and the queries are not specified. The scene-text supervision used to train the registers is generated by Qwen2.5-VL-72B-Instruct from the same overhead/egocentric views and ground-truth controller inputs (including relative position, facing, and change) that the eval queries ask about. If the evaluator is from the same model family, or even a different VLM, the reported Grounding and Consistency gains may reflect the generator learning to match a caption distribution the evaluator is familiar with rather than a persistent register state. Please disclose the evaluator and prompts, and report agreement with an independent VLM or human judgment on the same test clips. Without this, the central quantitative claim is not independently verified.
  2. [§4.2, Tables 1–4] No confidence intervals, significance tests, or test-set sizes are reported. The VLM accuracy values (e.g., 93.8, 76.6, 28.1) suggest a small number of clips per category; a gain of 9–19 points may correspond to only a few test items. The statement that WorldWeaver 'surpasses the previous best by a large margin' is therefore not statistically supported. Please report the number of evaluation clips, bootstrapped confidence intervals, and pairwise tests for the main comparisons.
  3. [§B.1, Table 5] The supervision ablation in Table 2 is not a controlled comparison. Single-signal ablations use the full 256-token register bank, whereas the combined setting splits the bank into disjoint 64/64/128 slices for the three heads. Thus the improvement from 'Registers only' to '+All' conflates added supervision with changes in register capacity and head grouping. This confounds the additive narrative in §4.3. Please run single-signal ablations on the same per-slice capacity, or otherwise match the register allocation across conditions.
  4. [§4.4, Table 3] The MoT versus dense comparison is only presented with 'no supervision' and 'scene text supervision', but the main system uses all three supervision signals. Dense is never evaluated with the combined supervision, so the architectural conclusion for the final model is incomplete. Add a dense + all-signals condition to Table 3, or qualify the MoT claim as applying only to the scene-text case.
minor comments (5)
  1. [Algorithm 2, line 17] The register update is written as 'Set ri ← Gθ(0, KV, ci)', which is inconsistent with Algorithm 1's 'Gθ(KVR, x̂i0, ci)'. Presumably the generated latent should be passed; please fix the pseudocode.
  2. [Eq. (3)] K is used both for the number of register tokens and for the number of categories in the world score. Rename one of them to avoid confusion.
  3. [Figure 4] Axis labels are missing, and the procedure for measuring 'VLM accuracy over training' is not described (evaluator, number of clips, prompts). Please specify.
  4. [Table 4] The text says 'generally consistent improvement', but several cells are non-monotonic (e.g., Grounding drops from 75.0 to 56.3 when adding 2.5K unlabeled clips; Movement drops from 78.1 to 68.8). A more cautious reading is needed.
  5. [§3.1 and references] Typo: 'Self-F orcing' should be 'Self-Forcing'. The 'Multiverse' reference is cited as a blog without a URL; please add a full citation.

Circularity Check

0 steps flagged

No significant circularity: the central result is an empirical architecture comparison, and the only self-citation is motivational rather than load-bearing.

full rationale

The paper's central claim is an empirical comparison, not a derivation. Conditional flow matching (Eq. 1), the Self-Forcing objective (Eq. 2), and register supervision are all defined independently of the VLM-accuracy metric; the reported gains in Table 1 come from the Solaris test split and include a registers-only ablation (world score 93.8 vs 81.0) that does not use scene-text supervision. The main concern is measurement validity rather than circularity: Appendix A.4 generates scene-text training targets with Qwen2.5-VL-72B-Instruct from the same overhead/egocentric views and ground-truth controller inputs that encode relative position and consistency, while Section 4.1 defines VLM accuracy as checking 'relative object position and cross-player consistency' without naming the evaluation VLM. If the evaluation VLM were the same model family, some supervised gains could reflect distribution familiarity; however, the paper never states that the evaluator is Qwen2.5-VL, so this is a confound, not an exhibited reduction. Self-citations to X-Fusion (Mo et al., 2025) appear only as motivation ('suggest that understanding tasks can improve generation') and are not used to prove the central result; the MoT design is credited to external work (Liang et al., 2024) and is itself ablated in Table 3. The paper's own limitation statement ('Our key improvement comes from the additional state supervision') concedes dependence on labeled state, but that is a practical limitation, not a definitional circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 7 axioms · 1 invented entities

The paper's contribution rests on a small set of hand-chosen hyperparameters (register count, window size, loss weights) and on domain assumptions about simulator ground truth, caption quality, and the validity of VLM-based evaluation. The main invented entity—the world state register—has multiple supervision handles, which lowers the graviton problem, but its capacity and drift properties are not analyzed.

free parameters (5)
  • λ_state (global register-loss weight) = 1.0 in Stage 2 and Stage 3
    Hand-set weight on the sum of register supervision losses; not swept. Affects how much the registers are pushed toward the auxiliary state targets.
  • Per-head loss weights λ_m = agent stats 0.02/0.2, bird's-eye view 2.0, scene text 0.1/0.5 (Stage 2 / Stage 3)
    Table 5; manually chosen relative weights for the three supervision signals. Stage 3 raises agent-stats and scene-text weights to stabilize self-forcing.
  • K, number of register tokens = 256 (split 64/64/128 in the combined setting)
    Capacity of the world-state pathway; chosen by hand, with no capacity scaling study.
  • W, sliding window size = 6 latent frames
    Local frame context window in Stages 2 and 3; affects how much history the register can consult.
  • Stage-3 training duration and freeze schedule = 3–4K steps; visual transformer frozen after step 1500
    Empirically chosen from Figure 4 to stabilize register training; a training-protocol parameter, not part of the method's derivation.
axioms (7)
  • standard math Flow-matching objective (Eq. 1) with Diffusion Forcing noise schedules is a valid training framework for autoregressive video diffusion.
    Adopted from cited prior work (Chen et al., 2024; Huang et al., 2025); no re-derivation in this paper.
  • standard math DMD-style distribution matching (Eq. 2), with sreal from the Stage-1 teacher and sfake as a trainable critic, improves rollout stability.
    Relied on for Stage 3 self-forcing; the paper provides no convergence analysis of its own.
  • domain assumption Simulator ground truth—agent positions/velocities/orientations at 20 Hz, shared overhead camera, and controller inputs—is an accurate world-state annotation.
    Used to define supervision targets in §3.3 and Appendix A; if these signals are noisy or misaligned with the egocentric views, the register targets are noisy.
  • domain assumption Qwen2.5-VL-72B scene-text captions, generated from ground-truth controller inputs, are semantically valid state descriptions.
    §A.4; caption quality is not separately evaluated, yet captions are training targets for the scene-text head.
  • domain assumption The unnamed evaluation VLM gives a valid and independent measure of logical consistency on generated videos.
    §4.1; the paper does not identify the eval VLM or demonstrate independence from the captioning VLM used for supervision. This is the paper's most load-bearing measurement assumption.
  • ad hoc to paper A fixed set of 256 shared tokens can represent the persistent world state across two agents and rollout steps.
    Core design choice in §3.3; no capacity analysis, scaling experiments, or guarantees that the register does not collapse or drift.
  • ad hoc to paper Separating register and frame weights in the Mixture-of-Transformers reduces objective conflict without hurting frame generation.
    §4.4; supported only by the dense-vs-MoT ablation under scene-text supervision, not by a general principle or analysis.
invented entities (1)
  • World state registers (WSR): a bank of 256 persistent cross-agent learnable tokens independent evidence
    purpose: Carry shared world information across agents and rollout chunks; updated by the generator after each generated frame and conditioned on by the next frame.
    Unlike a pure free parameter, the registers have explicit train-time decoders for agent statistics, bird's-eye-view features, and scene text (§3.3), and can be decoded qualitatively at inference (Fig. 3). These provide falsifiable checks beyond the frame-generation loss, though the checks share the same supervision pipeline.

pith-pipeline@v1.3.0-alltime-deepseek · 17154 in / 15626 out tokens · 149552 ms · 2026-08-01T06:57:46.857179+00:00 · methodology

0 comments
read the original abstract

Multi-agent interactive world models should not only generate consistent observations, but also maintain world states that persist across agents and evolve across views. Existing autoregressive video diffusion pipelines carry forward observation history as conditioning context, which makes shared state difficult to maintain in multi-agent and multi-view settings. We present WorldWeaver (W^2), a streaming multi-agent video diffusion model that augments rollout with cross-agent world state registers: learnable tokens that store shared world information, track individual agent status, and are dynamically updated after each generated chunk. We ground these registers with supervision signals spanning individual agent status, global state views including bird's-eye views, and scene text. We further improve the architecture with a Mixture-of-Transformers design that uses separate weights for world state modeling and visual frame modeling. Extensive experiments in two-agent Minecraft video generation show that explicit world-state modeling improves logical consistency and generation quality.

Figures

Figures reproduced from arXiv: 2607.21594 by Bolei Zhou, Krishna Kumar Singh, Sicheng Mo, Yuheng Li, Ziyang Leng.

Figure 1
Figure 1. Figure 1: Our model enables interactive agents acting in a shared world. During rollout, a set of [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Pipeline overview. (A) A standard streaming autoregressive diffusion model denoises each [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative visualization. We show generated multi-agent rollouts with grounded world states: agent coordinates, bird’s-eye-view maps, and scene text. These signals make the persistent world state inspectable and help verify cross-player consistency over time. Please refer to the project page for more detailed video demo. 4.3 ANALYSIS OF SUPERVISION SIGNALS A core problem is not only how to supervise the w… view at source ↗
Figure 4
Figure 4. Figure 4: VLM accuracy over Stage 3 self-forcing. Additionally, we investigate the training dynamics of the MoT model during Stage 3 training. Separate weight parameters also allow a more controlled learning process. We empirically find that the register loss in Stage 3 training converges more slowly than the DMD distillation loss, while the register loss mainly updates the register transformer and the DMD distillat… view at source ↗
Figure 5
Figure 5. Figure 5: Visualization of the data collection and annotation streams used in our multi-agent Minecraft [PITH_FULL_IMAGE:figures/full_fig_p016_5.png] view at source ↗

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

204 extracted references · 49 linked inside Pith

  1. [1]

    2020 , issue_date =

    Hanocka, Rana and Metzer, Gal and Giryes, Raja and Cohen-Or, Daniel , title =. 2020 , issue_date =

    2020

  2. [2]

    2017 , volume=

    Zhou, Bolei and Zhao, Hang and Puig, Xavier and Fidler, Sanja and Barriuso, Adela and Torralba, Antonio , booktitle=CVPR, title=. 2017 , volume=

    2017

  3. [3]

    Scene Parsing through ADE20K Dataset , author=

  4. [4]

    and Pons-Moll, Gerard and Black, Michael J

    Mahmood, Naureen and Ghorbani, Nima and Troje, Nikolaus F. and Pons-Moll, Gerard and Black, Michael J. , booktitle = ICCV, pages =. 2019 , month_numeric =

    2019

  5. [5]

    Jay , booktitle=ICIP, title=

    Wang, Haiqiang and Gan, Weihao and Hu, Sudeng and Lin, Joe Yuchieh and Jin, Lina and Song, Longguang and Wang, Ping and Katsavounidis, Ioannis and Aaron, Anne and Kuo, C.-C. Jay , booktitle=ICIP, title=. 2016 , volume=

    2016

  6. [6]

    RAVE: Randomized Noise Shuffling for Fast and Consistent Video Editing with Diffusion Models , author=

  7. [7]

    Deep unsupervised learning using nonequilibrium thermodynamics , author=

  8. [8]

    Denoising Diffusion Probabilistic Models , volume =

    Ho, Jonathan and Jain, Ajay and Abbeel, Pieter , booktitle = NIPS, editor =. Denoising Diffusion Probabilistic Models , volume =

  9. [9]

    Score-based generative modeling through stochastic differential equations , author=

  10. [10]

    Cascaded diffusion models for high fidelity image generation , author=

  11. [11]

    GitHub repository , howpublished =

    Patrick von Platen and Suraj Patil and Anton Lozhkov and Pedro Cuenca and Nathan Lambert and Kashif Rasul and Mishig Davaadorj and Dhruv Nair and Sayak Paul and William Berman and Yiyi Xu and Steven Liu and Thomas Wolf , title =. GitHub repository , howpublished =. 2022 , publisher =

    2022

  12. [12]

    Exploring the limits of transfer learning with a unified text-to-text transformer , author=

  13. [13]

    Learning transferable visual models from natural language supervision , author=

  14. [14]

    Midjourney , howpublished=

  15. [15]

    arXiv preprint arXiv:2204.06125 , year=

    Hierarchical text-conditional image generation with clip latents , author=. arXiv preprint arXiv:2204.06125 , year=

    Pith/arXiv arXiv

  16. [16]

    Glide: Towards photorealistic image generation and editing with text-guided diffusion models , author=

  17. [17]

    Photorealistic text-to-image diffusion models with deep language understanding , author=

  18. [18]

    Vector quantized diffusion model for text-to-image synthesis , author=

  19. [19]

    High-resolution image synthesis with latent diffusion models , author=

  20. [20]

    arXiv preprint arXiv:2211.01324 , year=

    ediffi: Text-to-image diffusion models with an ensemble of expert denoisers , author=. arXiv preprint arXiv:2211.01324 , year=

    Pith/arXiv arXiv

  21. [21]

    Podell, Dustin and English, Zion and Lacey, Kyle and Blattmann, Andreas and Dockhorn, Tim and M

  22. [22]

    2024 , archivePrefix=

    AnimateDiff: Animate Your Personalized Text-to-Image Diffusion Models without Specific Tuning , author=. 2024 , archivePrefix=

    2024

  23. [23]

    arXiv preprint arXiv:2309.15103 , year=

    LAVIE: High-Quality Video Generation with Cascaded Latent Diffusion Models , author=. arXiv preprint arXiv:2309.15103 , year=

    Pith/arXiv arXiv

  24. [24]

    Stylegan-t: Unlocking the power of gans for fast large-scale text-to-image synthesis , author=

  25. [25]

    Scaling up gans for text-to-image synthesis , author=

  26. [26]

    Zero-shot text-to-image generation , author=

  27. [27]

    Cogview: Mastering text-to-image generation via transformers , author=

  28. [28]

    Make-a-scene: Scene-based text-to-image generation with human priors , author=

  29. [29]

    Scaling autoregressive models for content-rich text-to-image generation , author=

  30. [30]

    Muse: Text-to-image generation via masked generative transformers , author=

  31. [31]

    Arbitrary Style Transfer in Real-time with Adaptive Instance Normalization , author=

  32. [32]

    arXiv preprint arXiv:2401.15636 , year=

    Freestyle: Free lunch for text-guided style transfer using diffusion models , author=. arXiv preprint arXiv:2401.15636 , year=

    Pith/arXiv arXiv

  33. [33]

    2021 , pages =

    Liu, Songhua and Lin, Tianwei and He, Dongliang and Li, Fu and Wang, Meiling and Li, Xin and Sun, Zhengxing and Li, Qian and Ding, Errui , title =. 2021 , pages =

    2021

  34. [34]

    Adding conditional control to text-to-image diffusion models , author=

  35. [35]

    Gligen: Open-set grounded text-to-image generation , author=

  36. [36]

    Sketch-guided text-to-image diffusion models , author=

  37. [37]

    T2I-Adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models , author=

  38. [38]

    Uni-ControlNet: All-in-One Control to Text-to-Image Diffusion Models , author=

  39. [39]

    Reco: Region-controlled text-to-image generation , author=

  40. [40]

    Spatext: Spatio-textual representation for controllable image generation , author=

  41. [41]

    Universal Guidance for Diffusion Models , author=

  42. [42]

    Ju, Xuan and Zeng, Ailing and Zhao, Chenchen and Wang, Jianan and Zhang, Lei and Xu, Qiang , booktitle=CVPR, year=. Human

  43. [43]

    Diffusion self-guidance for controllable image generation , author=

  44. [44]

    Localizing Object-level Shape Variations with Text-to-Image Diffusion Models , author=

  45. [45]

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

  46. [46]

    Boxdiff: Text-to-image synthesis with training-free box-constrained diffusion , author=

  47. [47]

    Expressive Text-to-Image Generation with Rich Text , author=

  48. [48]

    Training-free layout control with cross-attention guidance , author=

  49. [49]

    arXiv preprint arXiv:2306.05427 , year=

    Grounded Text-to-Image Synthesis with Attention Refocusing , author=. arXiv preprint arXiv:2306.05427 , year=

    Pith/arXiv arXiv

  50. [50]

    FreeControl: Training-Free Spatial Control of Any Text-to-Image Diffusion Model with Any Condition , author=

  51. [51]

    2023 , pages =

    Kim, Yunji and Lee, Jiyoung and Kim, Jin-Hwa and Ha, Jung-Woo and Zhu, Jun-Yan , title =. 2023 , pages =

    2023

  52. [52]

    Inversion-Free Image Editing with Natural Language , author=

  53. [53]

    arXiv preprint arXiv:2304.06140 , year =

    An Edit Friendly DDPM Noise Space: Inversion and Manipulations , author =. arXiv preprint arXiv:2304.06140 , year =

    Pith/arXiv arXiv

  54. [54]

    R&B: Region and Boundary Aware Zero-shot Grounded Text-to-image Generation , author=

  55. [55]

    Dense text-to-image generation with attention modulation , author=

  56. [56]

    Image-to-image translation with conditional adversarial networks , author=

  57. [57]

    Semantic image synthesis with spatially-adaptive normalization , author=

  58. [58]

    Palette: Image-to-image diffusion models , author=

  59. [59]

    Instructpix2pix: Learning to follow image editing instructions , author=

  60. [60]

    Imagic: Text-based real image editing with diffusion models , author=

  61. [61]

    Sine: Single image editing with text-to-image diffusion models , author=

  62. [62]

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

    Layoutdiffusion: Controllable diffusion model for layout-to-image generation , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  63. [63]

    2024 , eprint=

    InstanceDiffusion: Instance-level Control for Image Generation , author=. 2024 , eprint=

    2024

  64. [64]

    2024 , eprint=

    MIGC: Multi-Instance Generation Controller for Text-to-Image Synthesis , author=. 2024 , eprint=

    2024

  65. [65]

    Sdedit: Guided image synthesis and editing with stochastic differential equations , author=

  66. [66]

    Prompt-to-prompt image editing with cross attention control , author=

  67. [67]

    2023 , pages =

    Tumanyan, Narek and Geyer, Michal and Bagon, Shai and Dekel, Tali , title =. 2023 , pages =

    2023

  68. [68]

    Zero-shot image-to-image translation , author=

  69. [69]

    Cross-Image Attention for Zero-Shot Appearance Transfer , author=

  70. [70]

    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

  71. [71]

    Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation , author=

  72. [72]

    An image is worth one word: Personalizing text-to-image generation using textual inversion , author=

  73. [73]

    Multi-concept customization of text-to-image diffusion , author=

  74. [74]

    Break-A-Scene: Extracting Multiple Concepts from a Single Image , author=

  75. [75]

    2023 , eprint=

    Orthogonal Adaptation for Modular Customization of Diffusion Models , author=. 2023 , eprint=

    2023

  76. [76]

    2023 , eprint=

    HyperDreamBooth: HyperNetworks for Fast Personalization of Text-to-Image Models , author=. 2023 , eprint=

    2023

  77. [77]

    arXiv preprint arXiv:2303.09522 , year=

    P+ : Extended Textual Conditioning in Text-to-Image Generation , author=. arXiv preprint arXiv:2303.09522 , year=

    Pith/arXiv arXiv

  78. [78]

    A Neural Space-Time Representation for Text-to-Image Personalization , author=

  79. [79]

    Encoder-based domain tuning for fast personalization of text-to-image models , author=

  80. [80]

    Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation , author=

Showing first 80 references.