Pith. sign in

REVIEW 2 major objections 4 minor 22 references

Composing independently trained LoRAs as separate layers with one adapter active per pass removes multi-subject identity interference without weight merging.

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 13:17 UTC pith:QSTR7EZT

load-bearing objection Clean inference-time alternative to LoRA weight fusion: one adapter per pass under frozen conditioning, with real same-backbone evidence and the usual absolute-number caveats. the 2 major comments →

arxiv 2607.04801 v1 pith:QSTR7EZT submitted 2026-07-06 cs.CV

LILAC: Layer-Wise Independent LoRAs and Cascaded Conditioning for Multi-Concept Customization of Diffusion Models

Marian Lupascu , Sebastian Ripa , Mihai Trascau , Mariana-Iuliana Georgescu , Ionut Mironica This is my paper
classification cs.CV
keywords multi-concept personalizationdiffusion modelslow-rank adaptationlayered generationidentity preservationLoRA compositioncascaded conditioning
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.

Personalizing a diffusion model so that several specific people, characters, or objects appear together in one coherent image usually means merging their adapters into a shared set of weights. That merge causes identities to blend and styles to leak. This paper argues the problem is structural: once two adapters share active parameters, crosstalk is inevitable. LILAC instead treats multi-subject generation as layered synthesis. Each concept keeps its own independently trained low-rank adapter; at inference time subjects are placed one after another, each pass activating only that concept’s adapter and conditioning on the frozen composite of everything already drawn. Because the merged residual sum is never formed, parameter-level interference is eliminated by construction rather than minimized by training constraints. The method needs no joint retraining, scales linearly with the number of subjects, and works on both layered and ordinary image-editing backbones. Under a standard multi-concept protocol it records higher identity preservation than merge-based baselines while keeping comparable concept and text fidelity.

Core claim

Multi-subject personalization does not require concepts to share a weight space. By binding exactly one independently trained LoRA per generation pass and conditioning each pass on the frozen composite of previously placed subjects, LILAC removes cross-concept parameter interference by construction, so identities cannot blend at the weight level and no joint training or per-composition optimization is needed.

What carries the argument

Per-layer LoRA binding with cascaded frozen conditioning: at pass k the backbone runs with only adapter ΔW_k active and is conditioned on the clean latent of the alpha-composited image of subjects 1…k−1; the merged sum of residuals is never formed.

Load-bearing premise

That geometric and photometric coherence among subjects will reliably emerge from frozen clean-latent conditioning alone, and that sequential editing of a running composite will not accumulate identity-destroying drift on earlier subjects—especially when subjects must touch or mutually occlude.

What would settle it

Compose the same multi-subject groups with LILAC’s scaffold protocol and measure ArcFace detection rate as subject count grows and as mutual contact/occlusion increases; a sharp drop below merge-based baselines under contact-rich prompts, or progressive identity drift on earlier subjects after later passes, would falsify the claim that frozen conditioning plus single-adapter binding is sufficient.

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

2 major / 4 minor

Summary. The paper proposes LILAC, a multi-concept personalization method for text-to-image diffusion models that composes independently trained per-concept LoRA adapters at inference time rather than merging them in weight space. Subjects are generated sequentially as layers (or via scaffold/decomposition edits on non-layered backbones), with exactly one adapter active per pass and each generation conditioned on the frozen composite of previously placed subjects. This per-layer binding is claimed to eliminate parameter-level crosstalk by construction (Eq. 7), requiring no joint training or per-composition optimization, scaling linearly with the number of concepts, and remaining backbone-agnostic. Under the Orthogonal Adaptation concept bank and protocol the scaffold configuration on Qwen-Image-Edit reaches an ArcFace detection rate of 0.861 (vs. 0.745 reported for Orthogonal Adaptation on SDXL), with comparable CLIP-I and CLIP-T scores; a same-backbone re-run of Orthogonal Adaptation collapses to 0.000 identity while LILAC holds.

Significance. If the empirical claims hold, LILAC supplies a practical, training-free alternative to weight-space fusion methods (federated averaging, gradient fusion, orthogonality constraints) that currently dominate multi-concept customization. The isolation argument is definitional yet useful: by never forming the merged residual sum it removes a known failure mode (identity confusion, style bleeding) without rank-budget or joint-training costs. Strengths include public code, explicit same-backbone controls, single-concept screening documentation, ordering ablation, scalability curves, and a candid Limitations section. The result is of clear interest to the personalization community and to layered-generation work, even if absolute numbers partly reflect backbone strength.

major comments (2)
  1. Table 1 / Sec. 4.2: Absolute multi-subject ArcFace (0.861 scaffold) is compared with Orthogonal Adaptation’s 0.745 on SDXL, yet LILAC’s single-concept identity already sits at 0.961 versus 0.740 for the SDXL baselines, and Supplementary B screens out concepts that fail single-concept fidelity before multi-subject evaluation. The same-backbone OA collapse to 0.000 is convincing evidence that parameter-level interference is real, but the paper should more carefully disentangle composition benefit from backbone strength and pre-filtering—e.g., by reporting unscreened multi-subject numbers or matched single-concept baselines on the identical Qwen backbone—so that the headline claim is not overstated.
  2. Sec. 3.4–3.5 and Limitations: The central premise that geometric/photometric coherence emerges reliably from frozen clean-latent conditioning alone (without spatial losses or joint optimization) is acknowledged to break under close mutual occlusion and physical contact. These are common multi-subject scenarios; a quantitative failure-mode analysis (e.g., fraction of prompts that produce interpenetration artifacts, or a controlled occlusion subset of the Orthogonal Adaptation bank) would make the practical scope of the method clearer and would strengthen rather than weaken the contribution.
minor comments (4)
  1. Abstract: the sentence “while Orthogonal Adaptation reports 0.745 in its original setting. Adaptation reports 0.745 in its original setting.” is duplicated; remove the second occurrence.
  2. Fig. 7 caption and Sec. 3.5: the normalized overlap of 0.106 (5.3 imes chance level) is useful supporting evidence; consider adding the precise definition of the normalized Frobenius overlap so readers can reproduce the matrix.
  3. Sec. 4.4 / Fig. 8: ArcFace detection rate becomes noisy (std > 0.2) at N ≥ 3; a short note clarifying that this is attributed to face size/occlusion rather than identity blending would prevent misreading the scalability claim.
  4. Implementation details (Supp. B): rank-64 LoRA, 1 000 steps, blank-source training are stated; a one-line note on whether the same hyper-parameters were used for the Orthogonal Adaptation re-run on Qwen would improve reproducibility of the same-backbone comparison.

Circularity Check

0 steps flagged

No significant circularity; isolation is definitional of the non-merging protocol, metrics are external, and no fitted quantities are re-presented as predictions.

full rationale

LILAC is an empirical systems/methods paper whose central claim is a design choice (per-pass single-adapter binding + frozen clean-latent conditioning) rather than a first-principles derivation. Sec. 3.5 and Eq. 7 correctly observe that the merged residual sum of weight-space fusion is never formed, so parameter-level crosstalk is never instantiated; this is tautological of the procedure, not a circular reduction of a predicted quantity to its own inputs. Identity, concept, and text metrics (ArcFace detection rate, CLIP-I, CLIP-T) are pre-existing external evaluators taken from the Orthogonal Adaptation protocol; no free parameters are fitted on a subset and then called predictions. The LaDe backbone (overlapping authors) is used only as one of two validation platforms; the headline 0.861 number and the same-backbone Orthogonal Adaptation collapse (0.000) are obtained on independent Qwen-Image-Edit models. Concept screening (Supp. B) is an explicit pre-filter, not a hidden circular fit. No uniqueness theorems, ansatzes, or renamed known results appear. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 3 invented entities

Standard empirical CV methods paper. Load-bearing content is the cascade protocol itself plus ordinary diffusion/LoRA assumptions. Free parameters are the usual training and evaluation hyperparameters; no physical constants or fitted universal scales. Invented entities are the named protocol and its two non-native realizations.

free parameters (4)
  • LoRA rank r = 64
    Set to 64 on MMDiT attention layers; controls capacity of each concept adapter and is chosen by authors.
  • adapter training steps / learning rate = 1000 steps, 1e-4
    1000 steps at 1e-4; standard but free choices that affect single-concept fidelity before composition.
  • ArcFace identity threshold = 0.32 cosine
    Cosine > 0.32 (distance < 0.68) taken from Orthogonal Adaptation protocol; defines the binary detection rate that is the headline metric.
  • anchor selection rule = max ||ΔW||_F first
    Concept with largest adapter Frobenius norm is placed first; a design hyperparameter whose effect is ablated but still free.
axioms (4)
  • domain assumption A diffusion backbone that admits low-rank adaptation and can condition generation on a clean (t=0) latent or image edit channel exists and remains frozen.
    Stated as the abstract interface B=⟨θ,E,D⟩ in Sec. 3.1; required for both native layered and scaffold routes.
  • domain assumption An independently trained single-concept LoRA captures identity sufficiently that the subject can be regenerated when its adapter alone is active.
    Sec. 3.2 and the single-concept screening step in Supplementary B; multi-subject claims presuppose this.
  • domain assumption Alpha compositing (or image-edit replacement of placeholders) of sequentially generated layers yields a coherent final image without requiring joint optimization.
    Core of Sec. 3.4 and Algorithm 1; the method's claim that coherence emerges from frozen context alone.
  • standard math Standard linear algebra of residual adapters (ΔW = BA) and the definition of crosstalk χ_ij = ||ΔW_i^T ΔW_j||_F.
    Used in Sec. 3.5 to contrast merged vs per-layer binding.
invented entities (3)
  • LILAC cascaded inference protocol (per-layer LoRA binding + frozen composite conditioning) independent evidence
    purpose: Compose any number of independently trained concept adapters without forming a merged weight or joint training.
    The central methodological object introduced in Sec. 3; independent evidence is the empirical multi-subject metrics under a public protocol.
  • Scaffold configuration for non-layered editors independent evidence
    purpose: Realize the same cascade on ordinary image-editing models by first generating a placeholder scene then replacing one subject per pass.
    Sec. 3.4 / Fig. 2; engineering realization rather than a new physical entity.
  • Decomposition configuration via off-the-shelf layer split independent evidence
    purpose: Alternative route that splits a scene into RGBA layers then re-renders each person layer with its adapter.
    Sec. 3.4 / Fig. 3; shown to be weaker than scaffold, reported mainly for comparison.

pith-pipeline@v1.1.0-grok45 · 19120 in / 3467 out tokens · 32576 ms · 2026-07-11T13:17:04.916446+00:00 · methodology

0 comments
read the original abstract

Personalizing text-to-image diffusion models to render several specific subjects in a coherent image remains challenging: the model must preserve each subject's identity while keeping the scene spatially and visually coherent. Methods that fuse independently trained concept adapters in a shared weight space (via federated averaging, gradient fusion, or orthogonality constraints) suffer from identity confusion and style bleeding and require joint retraining. In this work, we show that composing concepts as separate image layers, instead of merging their adapters in a shared weight space, avoids parameter-level interference. We introduce LILAC, a framework that composes independently trained low-rank adapters at inference time: each subject is conditioned on the frozen composite of previously placed subjects, with exactly one adapter active at a time, therefore identities never interfere at the parameter level. LILAC composes the adapters without any joint training, scales linearly with the number of concepts, and is backbone-agnostic. Under the Orthogonal Adaptation protocol, LILAC applied on Qwen-Image-Edit reaches an ArcFace detection rate of 0.861, while Orthogonal Adaptation reports 0.745 in its original setting. Adaptation reports 0.745 in its original setting. Code is available at https://github.com/marianlupascu/LILAC.

Figures

Figures reproduced from arXiv: 2607.04801 by Ionut Mironica, Mariana-Iuliana Georgescu, Marian Lupascu, Mihai Trascau, Sebastian Ripa.

Figure 1
Figure 1. Figure 1: Overview of LILAC. Training (left): Each concept is trained in isolation into a single LoRA adapter, building a library of independently trained, per-concept adapters. LoRA Library (center): Concepts trained independently, one per subject. Inference (right): LILAC selects a subset and renders every subject in one coherent scene from a text prompt. Exactly one adapter is active per pass under frozen conditi… view at source ↗
Figure 2
Figure 2. Figure 2: Cascaded inference (scaffold configuration). Given an image (natural or gener￾ated), each concept is then added in its own pass with only its adapter active, condi￾tioned on the frozen image of the concepts placed so far. No two adapters are active together, and earlier concepts and the background are preserved as later ones are added; the final pass adds a non-person concept (sunglasses), showing that the… view at source ↗
Figure 3
Figure 3. Figure 3: Decomposition configuration on Qwen-Image-Edit [18]. Given an input image, Qwen-Image-Layered [19] then splits it into RGBA layers, here a background and two subject layers. Each subject is regenerated in its own pass: the frozen background is added as context, the placeholder subject is replaced with that concept’s adapter active, and the background is dropped to recover a transparent subject layer. The b… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative multi-subject comparison on the Orthogonal Adaptation concept bank [12]. Each pair of rows shows Orthogonal Adaptation [12] (top) and LILAC in its scaffold configuration (bottom) on the same prompt. The groups vary in size and combine a broad range of concepts, including real public figures, stylized fictional char￾acters, animals, and accessories, placed in diverse scene contexts. Implementati… view at source ↗
Figure 5
Figure 5. Figure 5: Additional multi-subject compositions produced by LILAC. Each panel shows one generated scene; the circular thumbnails above it are the reference identities of the composed concepts, one per subject. The examples span two to three subjects and mix real public figures, stylized fictional characters, and animals, placed across varied scenes, styles, and interactions (side by side, shaking hands, toasting, co… view at source ↗
Figure 6
Figure 6. Figure 6: Single-concept generations in the style of Cyberpunk 2077 for three concepts: input references (top) and a comparison of Orthogonal Adaptation [12] with LILAC. A Additional Qualitative Results Beyond the multi-subject compositions of [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Cross-concept overlap between the independently trained adapters. Each cell is the normalised overlap ovij ∈ [0, 1] between the LoRA residuals of concepts i and j (1 on the diagonal; 0 for orthogonal residuals). Off-diagonal values average 0.106, 5.3× the 0.020 chance level for random adapters of the same size: the residuals are not orthogonal. This is the crosstalk χij that weight-space fusion must suppre… view at source ↗
Figure 8
Figure 8. Figure 8: Per-subject quality versus the number of subjects N (scaffold configuration); N=1 is the single-concept reference, averaged over the concept bank. Concept fidelity (CLIP-I) and text alignment (CLIP-T) stay flat as N grows, while the ArcFace detec￾tion rate is lower and noisier at larger N, reflecting scene crowding and smaller faces rather than identity blending. Error bars are ±1 standard deviation, over … 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

22 extracted references · 3 canonical work pages

  1. [1]

    161:1–161:10 (2025).https://doi.org/10.1145/3757377.3763987

    Abdal,R.,Patashnik,O.,Deyneka,E.,Chen,H.,Siarohin,A.,Tulyakov,S.,Cohen- Or, D., Aberman, K.: Zero-shot dynamic concept personalization with grid-based lora pp. 161:1–161:10 (2025).https://doi.org/10.1145/3757377.3763987

    work page doi:10.1145/3757377.3763987 2025

  2. [2]

    In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019

    Deng, J., Guo, J., Xue, N., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019. pp. 4690–4699. Computer Vision Foundation / IEEE (2019).https://doi.org/10.1109/CVPR. 2019.00482

    doi:10.1109/cvpr 2019

  3. [3]

    In: Leonardis, A., Ricci, E., Roth, S., Russakovsky, O., Sattler, T., Varol, G

    Frenkel, Y., Vinker, Y., Shamir, A., Cohen-Or, D.: Implicit style-content sep- aration using b-lora. In: Leonardis, A., Ricci, E., Roth, S., Russakovsky, O., Sattler, T., Varol, G. (eds.) Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part X. Lecture Notes in Computer Science, vol. 15068, p...

    work page doi:10.1007/978-3-031-72684-2_11 2024

  4. [4]

    In: International Conference on Learning Representations (ICLR) (2023)

    Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion. In: International Conference on Learning Representations (ICLR) (2023)

    2023

  5. [5]

    In: Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S

    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. In: Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S. (eds.) Advances in Neural Information Processing Systems ...

    2023

  6. [6]

    In: The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29,

    Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. In: The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29,

    2022

  7. [7]

    OpenReview.net (2022)

    2022

  8. [8]

    Lecture Notes in Computer Science, vol

    Kong, Z., Zhang, Y., Yang, T., Wang, T., Zhang, K., Wu, B., Chen, G., Liu, W., Luo, W.: Omg: Occlusion-friendly personalized multi-concept generation in diffu- sionmodels.In:ComputerVision-ECCV2024-18thEuropeanConference,Milan, Italy, September 29-October 4, 2024, Proceedings, Part XXXI. Lecture Notes in Computer Science, vol. 15089, pp. 253–270. Springer (2024)

    2024

  9. [9]

    In: IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24,

    Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.: Multi-concept cus- tomization of text-to-image diffusion. In: IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24,

    2023

  10. [10]

    1931–1941

    pp. 1931–1941. IEEE (2023).https://doi.org/10.1109/CVPR52729.2023. 00192

    doi:10.1109/cvpr52729.2023 1931

  11. [11]

    In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17- 24, 2023

    Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., Li, C., Lee, Y.J.: GLIGEN: open- set grounded text-to-image generation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17- 24, 2023. pp. 22511–22521. IEEE (2023).https://doi.org/10.1109/CVPR52729. 2023.02156

    doi:10.1109/cvpr52729 2023

  12. [12]

    arXiv preprint arXiv:2603.17965 (2026) 16 M

    Lungu-Stan, V.C., Mironică, I., Georgescu, M.I.: Lade: Unified multi-layered graphic media generation and decomposition. arXiv preprint arXiv:2603.17965 (2026) 16 M. Lupaşcu et al

    arXiv 2026

  13. [13]

    In: IEEE/CVF Conference on Computer Vision and Pattern Recog- nition, CVPR 2025, Nashville, TN, USA, June 11-15, 2025

    Ouyang, Z., Li, Z., Hou, Q.: K-lora: Unlocking training-free fusion of any subject and style loras. In: IEEE/CVF Conference on Computer Vision and Pattern Recog- nition, CVPR 2025, Nashville, TN, USA, June 11-15, 2025. pp. 13041–13050. Com- puter Vision Foundation / IEEE (2025).https://doi.org/10.1109/CVPR52734. 2025.01217

    doi:10.1109/cvpr52734 2025

  14. [14]

    In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024

    Po, R., Yang, G., Aberman, K., Wetzstein, G.: Orthogonal adaptation for modular customization of diffusion models. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024. pp. 7964–7973. IEEE (2024).https://doi.org/10.1109/CVPR52733.2024.00761

    doi:10.1109/cvpr52733.2024.00761 2024

  15. [15]

    In: International Conference on Machine Learning (ICML) (2021)

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (ICML) (2021)

    2021

  16. [16]

    In: Meila, M., Zhang, T

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual E...

    2021

  17. [17]

    In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023

    Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dream- booth: Fine tuning text-to-image diffusion models for subject-driven generation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023. pp. 22500–22510. IEEE (2023). https://doi.org/10.1109/CVPR52729.2023.02155

    doi:10.1109/cvpr52729.2023.02155 2023

  18. [18]

    In: Leonardis, A., Ricci, E., Roth, S., Russakovsky, O., Sattler, T., Varol, G

    Shah, V., Ruiz, N., Cole, F., Lu, E., Lazebnik, S., Li, Y., Jampani, V.: Ziplora: Any subject in any style by effectively merging loras. In: Leonardis, A., Ricci, E., Roth, S., Russakovsky, O., Sattler, T., Varol, G. (eds.) Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part I. Lecture Note...

    work page doi:10.1007/978-3-031-73232-4_24 2024

  19. [19]

    CoRRabs/2303.09522(2023).https: //doi.org/10.48550/ARXIV.2303.09522

    Voynov, A., Chu, Q., Cohen-Or, D., Aberman, K.: P+: extended textual con- ditioning in text-to-image generation. CoRRabs/2303.09522(2023).https: //doi.org/10.48550/ARXIV.2303.09522

    Pith/arXiv arXiv doi:10.48550/arxiv.2303.09522 2023

  20. [20]

    Wu, C., Li, J., Zhou, J., Lin, J., Gao, K., Yan, K., Yin, S.m., Bai, S., Xu, X., et al.: Qwen-image technical report (2025)

    2025

  21. [21]

    arXiv preprint arXiv:2512.15603 (2025)

    Yin, S., Zhang, Z., Tang, Z., Gao, K., Xu, X., Yan, K., Li, J., Chen, Y., Chen, Y., Shum, H., Ni, L.M., Zhou, J., Lin, J., Wu, C.: Qwen-image-layered: Towards in- herent editability via layer decomposition. arXiv preprint arXiv:2512.15603 (2025)

    arXiv 2025

  22. [22]

    ACM Trans

    Zhang, L., Agrawala, M.: Transparent image layer diffusion using latent trans- parency. ACM Trans. Graph.43(4), 100:1–100:15 (2024).https://doi.org/10. 1145/3658150 LILAC: Layered Multi-Concept Customization of Diffusion Models 17 Orthogonal AdaptationOrthogonal AdaptationOrthogonal AdaptationLILAC (Ours)LILAC (Ours)LILAC (Ours) Input Images Fig. 6:Single...

    arXiv 2024