Pith. sign in

REVIEW 2 major objections 4 minor 1 cited by

A fixed basis of 32 LoRA adapters, mixed per task by a learned router, generalizes to unseen visual analogy edits.

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 22:44 UTC pith:VLGJHTZB

load-bearing objection Solid empirical method with an oversold 'spanning' story; the bilinear composition and rank-4 bound need to be addressed, but the experiments justify a real review. the 2 major comments →

arxiv 2602.15727 v2 pith:VLGJHTZB submitted 2026-02-17 cs.CV cs.AIcs.GRcs.LGeess.IV

Spanning the Visual Analogy Space with a Weight Basis of LoRAs

Hila Manor , Rinon Gal , Haggai Maron , Tomer Michaeli , Gal Chechik This is my paper
classification cs.CV cs.AIcs.GRcs.LGeess.IV
keywords visual analogyLoRA basisadapter compositionimage editingdiffusion modelszero-shot generalizationdynamic routingflow matching
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.

The paper argues that visual analogy editing—turning an image b into b' the way a pair (a, a') transforms its subject—is better served by composing a set of learned transformation primitives than by fine-tuning a single adapter. It introduces LoRWeB, which jointly learns a fixed basis of 32 low-rank adapters and a lightweight encoder that, for each input analogy triplet, predicts softmax weights for combining those adapters on the fly. Across standard metrics, a user study, and a held-out set of 100 analogy tasks, this basis approach outperforms single-LoRA and prior analogy methods, especially on transformations not seen in training. The authors conclude that the bottleneck of previous approaches is the single fixed module, and that a learned basis decomposition is a flexible remedy.

Core claim

The central claim is that the space of visual analogies can be spanned by a fixed set of 32 rank-4 LoRA modules. Given an analogy pair (a, a') and a target image b, an encoder maps the triplet to a query vector; the softmax of the query's dot product with N learned keys yields coefficients e_i. The final adapter is the weighted sum A=Σe_i A_i, B=Σe_i B_i, injected into a frozen text-to-image flow model (Flux.1-Kontext) whose in-context attention receives the 2×2 composite of a, a', b, and the generated b'. Training these adapters, keys, and projection jointly on 208 analogy tasks enables composition of known primitives into novel edits. The paper reports state-of-the-art results on automated

What carries the argument

The key mechanism is a learnable basis of LoRA adapters with a learned key per adapter, and an encoder (frozen CLIP ViT plus a linear projection) that produces a query from the triplet {a, a', b}. Mixing coefficients are computed as a softmax over query-key similarities, so the composed LoRA is a convex combination of basis adapters. The same architecture is repeated for each weight matrix of the transformer, with separate keys and projections. This turns the problem of predicting an edit into the problem of locating a point in the adapter space.

Load-bearing premise

The assumption that every visual transformation required by an analogy is a non-negative linear combination of the 32 learned LoRA primitives; if a needed edit lies outside that convex hull, the router cannot produce it regardless of encoder quality.

What would settle it

Train LoRWeB and then test on an analogy pair whose correct edit is the negative or a sum greater than 1 of two basis edits; if coefficients are constrained to a simplex, the output should fail. Alternatively, compare coefficients predicted by the encoder with coefficients optimized directly on held-out pairs; if optimized coefficients lie far outside the convex hull, the basis does not span the space.

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

If this is right

  • If the 32-adapter basis truly spans the transformation space, single-LoRA capacity limits are real, and larger adapter counts should yield further gains.
  • The method enables per-task specialization at inference time with no test-time fine-tuning, since the router produces coefficients in one forward pass.
  • Because the router is conditioned on the triplet, the same basis can express very different edits—style transfer, object insertion, layout change—without retraining.
  • The learned basis and keys are per weight matrix, suggesting different layers can capture different granularities of the transformation.
  • The approach replaces a discrete set of trained LoRAs with a continuous composition, making interpolation between tasks explicit via the coefficient vector.

Where Pith is reading between the lines

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

  • If the convex-combination assumption holds only for seen task families, the method may still fail on truly novel compositions; a direct test would evaluate transformations that require extrapolated coefficients, such as weights summing to more than 1 or negative weights.
  • The small basis size (32) hints that the effective manifold of visual edits is low-dimensional, which could be probed by measuring reconstruction error on held-out analogies as N increases.
  • One could reuse the same basis for other generative backbones or image-to-image tasks, turning any fine-tuning setup into a routing problem; this is an untested extension consistent with the paper's framework.

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 introduces LoRWeB, a method for visual analogy completion built on Flux.1-Kontext. Instead of training a single LoRA for all analogy tasks, LoRWeB maintains N=32 rank-4 LoRA adapters and a lightweight CLIP-based encoder that predicts softmax coefficients from the analogy triplet {a,a',b}. The coefficients are used to combine the A and B factor matrices (Eq. 2), and the resulting 'mixed' adapter is injected into the flow model. The authors claim this spans the space of visual transformations and improves generalization to unseen tasks. They evaluate on a custom benchmark of 100 tasks (90 of which are author-constructed), using VLM scores, pairwise win rates, a user study, and ablations on basis size, rank, normalization, encoder input, and encoder backbone.

Significance. If the empirical results are taken at face value, the paper demonstrates a useful recipe: input-conditioned mixing of a small number of low-rank adapters outperforms a single larger LoRA in a same-parameter-budget comparison, across automated metrics and human preference. The experiments are extensive, and the VLM-human alignment check (66.7% agreement vs. 74.2% user-user agreement) is a positive sign for the evaluation protocol. However, the mathematical claim that the composition is a 'linear combination' or a 'basis' in weight space is incorrect: the actual weight update is bilinear in the coefficients and has rank at most r. This does not necessarily invalidate the empirical gains, but it requires a substantial reframing and additional analysis before the conceptual contribution is established.

major comments (2)
  1. [Sec. 3.2, Eq. (2)] The paper states that the coefficients e_i 'linearly combine' the basis LoRAs, but the injected update is ΔW = B A = (Σ_i e_i B_i)(Σ_j e_j A_j) = Σ_i e_i^2 B_i A_i + Σ_{i≠j} e_i e_j B_i A_j. This is not a linear (or convex) combination of the basis adapters ΔW_i = B_i A_i, and it is not the interpolation rule of Dravid et al. [12], which operates on the full weight matrices. Moreover, rank(ΔW) ≤ r = 4 regardless of N, so the construction does not 'span' a higher-dimensional space, and the comparison to a single r=128 LoRA is not a rank-matched comparison. Please either adopt a true weight-space combination (e.g., ΔW = Σ_i e_i B_i A_i) or revise the 'basis/span' claims and provide an ablation that disentangles factor-level mixing from weight-level mixing.
  2. [Sec. 4, Dataset] The key generalization claim rests on an evaluation set in which 90 of 100 tasks were created by the authors (LLM-generated prompts plus 18 community LoRAs, with manual filtering). This is not an established benchmark, and the filtering step ('we filter prompts where Flux.1-Kontext fails to produce a meaningful edit') may select for tasks that the base model can already handle. Please release the full evaluation data, report results on the 10 original Relation252k unseen tasks separately, and provide per-task breakdowns with confidence intervals. Without this, the claim of 'significantly improved generalization to unseen transformations' is difficult to verify independently.
minor comments (4)
  1. [Throughout] Numerous typos: 'Aadpter' in Fig. 4 and Fig. 7 captions, 'analoy' in Fig. S1, 'LoRBA' in Fig. S1 caption, 'https://https://huggingface.co/' in footnote 3, and 'Turn this into image' in Fig. 1.
  2. [Sec. 3.1] The scaling constant α in the LoRA definition (W = W0 + (α/r) BA) is not used anywhere in the method description or experiments. Please clarify whether α is set to 1 or absorbed into the learned factors.
  3. [Sec. 4.3, Table 1] The ablation narrative is not fully supported by the table: increasing rank from 4 to 16 at fixed N=32 lowers VLM Edit Accuracy from 5.94 to 4.92, and at r=16 increasing N from 8 to 32 lowers accuracy from 5.49 to 4.92. The text only says that reducing basis size or capacity hurts; it does not discuss the non-monotonic role of rank. Please reconcile these numbers or temper the 'large basis is important' claim.
  4. [Table 1 and Fig. 5] VLM scores (Preservation, Edit Accuracy) are reported as point estimates without variance or confidence intervals. Given that these are used for the main quantitative claim, please report standard errors or confidence intervals, as is done for the pairwise win rates.

Circularity Check

0 steps flagged

No derivational circularity: LoRWeB is an empirical basis-mixing method whose components and training set are stated; the rank/spanning critique is an expressivity objection, not a circularity.

full rationale

The paper does not derive its reported gains from its definitions. The mixed adapter is explicitly defined by Eqs. (2)-(4) (A=Sum e_i A_i, B=Sum e_i B_i, with softmax coefficients from a learned encoder), and the basis and router are trained jointly on Relation252k, not optimized to reproduce the evaluation outputs. The claim that a LoRA basis spans visual transformations is informal motivation, not a theorem derived from the equations; the skeptical observation that Delta-W = BA remains at most rank 4 is a real expressivity/correctness concern, but it is an unsupported modeling claim rather than a reduction of output to input. The motivating citation [12] is external work (Dravid et al.), not a self-citation, and no uniqueness or spanning theorem is imported. The custom evaluation set and ablations are author-designed and used to pick N, r, and the softmax, which is a validation-side risk, but all methods (including baselines) are evaluated on the same held-out tasks and the router is not fitted to those outputs. The Discussion's stated limitation that the model 'may still struggle with tasks that are significantly different from the training corpus' is an empirical caveat, not a circular step. No load-bearing step satisfies the quoted-equation test for circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The paper is an empirical ML method, so I do not enumerate the millions of trained network weights; the list above captures the task-level choices the central claim depends on, plus the assumptions that must hold for the method to work. The central capacity and routing assumptions are listed; the main one (linear span of 32 LoRAs) is plausible but unproven.

free parameters (4)
  • number of basis LoRAs N = 32
    Chosen by hand; ablation shows N=8/16/64 changes win rates, so the 'large basis' claim is tied to this value (Sec. 4.3, Tab. 1).
  • LoRA rank r = 4
    Chosen by hand; r=16 lowers edit accuracy, r=4,N=32 selected as best (Sec. 4.3, Tab. 1).
  • key/query dimension d = 128
    Hidden dimension of learned keys and projection; not ablated but sets routing capacity (Sec. 4).
  • normalization for mixing coefficients = softmax
    softmax chosen over tanh after ablations; tanh drastically underperforms (Sec. 4.3).
axioms (4)
  • domain assumption A linear combination of LoRA adapters (A=Sum(e_i A_i), B=Sum(e_i B_i)) is a valid adapter that can approximate a new transformation.
    Eq. (2), Sec. 3.2; load-bearing capacity assumption, motivated only by Dravid et al.'s interpolation results on personalization LoRAs.
  • domain assumption CLIP (or SigLIP) embeddings of the triplet contain enough information to select the correct basis combination.
    Eqs. (3)-(4), Sec. 3.2; the router quality depends on this; CLIP is frozen and coarse.
  • domain assumption Flux.1-Kontext's extended-attention mechanism can use a 2x2 composite image as context while preserving fine details.
    Sec. 3.2; the model's fine-detail conditioning comes from the composite grid, not from the encoder.
  • ad hoc to paper The author-built eval set (LLM-generated prompts + 18 community LoRAs) is a valid proxy for 'unseen transformations'.
    Sec. 4 Dataset; this assumption underlies all generalization claims and is not independently benchmarked.

pith-pipeline@v1.3.0-alltime-deepseek · 15832 in / 15298 out tokens · 132335 ms · 2026-08-02T22:44:58.987226+00:00 · methodology

0 comments
read the original abstract

Visual analogy learning enables image editing via demonstration rather than textual description, allowing users to specify complex transformations difficult to articulate in words. Given a triplet $\{\mathbf{a}$, $\mathbf{a}'$, $\mathbf{b}\}$, the goal is to generate $\mathbf{b}'$ such that $\mathbf{a} : \mathbf{a}' :: \mathbf{b} : \mathbf{b}'$. Recent methods adapt text-to-image models with a single Low-Rank Adaptation (LoRA) module, but they face a fundamental limitation: attempting to capture the diverse space of visual transformations within a fixed module constrains generalization. Inspired by recent work showing that LoRAs in constrained domains span meaningful, interpolatable semantic spaces, we propose LoRWeB, which specializes the model for each analogy task in a single inference pass. LoRWeB dynamically composes learned transformation primitives, informally, choosing a point in a "space of LoRAs". We introduce two key components: (1) a learnable basis of LoRAs to span the space of different visual transformations, and (2) a lightweight encoder that dynamically weighs these basis LoRAs given the input analogy pair. Comprehensive evaluations demonstrate state-of-the-art performance and significantly improved generalization to unseen transformations. Our findings suggest LoRA basis decompositions are a promising direction for flexible visual manipulation tasks. See https://research.nvidia.com/labs/par/lorweb for code.

Figures

Figures reproduced from arXiv: 2602.15727 by Gal Chechik, Haggai Maron, Hila Manor, Rinon Gal, Tomer Michaeli.

Figure 1
Figure 1. Figure 1: LoRWeB. We present a novel method for analogy-based editing, based on learnable mixing of low-rank adapters. Given a prompt and an image triplet {a, a ′ , b} that visually describe a desired transformation, LoRWeB dynamically constructs a single LoRA from a learnable basis of LoRA modules, and produces an editing result b ′ that applies the same analogy for the new image. Abstract Visual analogy learning e… view at source ↗
Figure 2
Figure 2. Figure 2: LoRWeB Overview. We first encode a and a ′ , that describe a visual transformation (e.g. adding a hat to the man), and b, which should be edited analogously (e.g. adding a hat to the woman) with CLIP [42], and a small learned projection module. The similarity between the encoded vector and a set of learned keys determines the linear coefficients for combining the learned LoRAs into a single, mixed LoRA. Th… view at source ↗
Figure 3
Figure 3. Figure 3: LoRWeB visual analogy results. Using a LoRA Basis allows LoRWeB to generalize to a wide variety of new analogy tasks, from adding objects to transferring specific styles or makeup or copying pose changes. Please zoom in for more details. categories. Including the unseen set of Relation252K, this gives 100 tasks across 840 analogy triplets. On all experi￾ments, we first aggregate the results per analogy tas… view at source ↗
Figure 4
Figure 4. Figure 4: Comparisons with baseline methods on unseen tasks. Our approach generalizes across more diverse tasks, and better maintains the visual details of both the subject and the analogy. User study. Beyond automated metrics, we also conduct a two-alternative forced choice user study. We show each user a reference pair (a, a ′ ), an input image b, and two results (one from our model and one of a random baseline), … view at source ↗
Figure 5
Figure 5. Figure 5: Quantitative comparisons. (left) Accuracy of the applied edit and preservation of b in b ′ using Gemma-3 [52]. Top right is better. (right) CLIP directional similarity and LPIPS between b ′ and b. Bottom-right is better. Our method pushes the Pareto front of edit accuracy-preservation, achieving higher edit accuracy while strongly preserving the input image. 0 50 100 Win Rate User Study Pairwise VLM 83.6 1… view at source ↗
Figure 6
Figure 6. Figure 6: Pairwise image comparisons. We compare LoRWeB to four baselines on overall edit quality preference via both a user study and using a VLM. LoRWeB produces edits that are favored by both. Error bars are the 68% Wilson score interval. Layout of encoder input. In our approach, we elected to separately encode each of the conditioning analogy images using CLIP, and concatenate their representations. Our intu￾iti… view at source ↗
Figure 7
Figure 7. Figure 7: Effect of different reference analogy pairs. LoRWeB directly leverages the analogy pair to understand the details of the proposed task, applying an edit that is beyond just text-based editing based on the given prompt. For example, when the prompt is “Give this creature a crown of crystals”, the analogy context passes information on the amount and color of the crystals. outperforms and generalizes better t… view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Delta-Adapter: Scalable Exemplar-Based Image Editing with Single-Pair Supervision

    cs.CV 2026-05 unverdicted novelty 7.0

    Delta-Adapter extracts a semantic delta from a single image pair via a pre-trained vision encoder and injects it through a Perceiver adapter to enable scalable single-pair supervised editing.

Reference graph

Works this paper leans on

70 extracted references · 9 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Cross-image attention for zero- shot appearance transfer

    Yuval Alaluf, Daniel Garibi, Or Patashnik, Hadar Averbuch- Elor, and Daniel Cohen-Or. Cross-image attention for zero- shot appearance transfer. InACM SIGGRAPH 2024 confer- ence papers, pages 1–12, 2024. 2

    2024

  2. [2]

    Build- ing normalizing flows with stochastic interpolants

    Michael Samuel Albergo and Eric Vanden-Eijnden. Build- ing normalizing flows with stochastic interpolants. InThe Eleventh International Conference on Learning Representa- tions, 2023. 3

    2023

  3. [3]

    Introducing claude 4

    Anthropic. Introducing claude 4. https : / / www . anthropic.com/news/claude-4, 2025. 12

    2025

  4. [4]

    Visual prompting via image inpaint- ing.Advances in Neural Information Processing Systems, 35: 25005–25017, 2022

    Amir Bar, Yossi Gandelsman, Trevor Darrell, Amir Glober- son, and Alexei Efros. Visual prompting via image inpaint- ing.Advances in Neural Information Processing Systems, 35: 25005–25017, 2022. 2

    2022

  5. [5]

    Black Forest Labs, Stephen Batifol, Andreas Blattmann, Fred- eric Boesel, Saksham Consul, Cyril Diagne, Tim Dockhorn, Jack English, Zion English, Patrick Esser, et al. FLUX. 1 kontext: Flow matching for in-context image generation and editing in latent space.arXiv preprint arXiv:2506.15742,

    Pith/arXiv arXiv

  6. [6]

    In- structpix2pix: Learning to follow image editing instructions

    Tim Brooks, Aleksander Holynski, and Alexei A Efros. In- structpix2pix: Learning to follow image editing instructions. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18392–18402, 2023. 1

    2023

  7. [7]

    Diffusion self- distillation for zero-shot customized image generation

    Shengqu Cai, Eric Ryan Chan, Yunzhi Zhang, Leonidas Guibas, Jiajun Wu, and Gordon Wetzstein. Diffusion self- distillation for zero-shot customized image generation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 18434–18443, 2025. 2

    2025

  8. [8]

    Masactrl: Tuning-free mu- tual self-attention control for consistent image synthesis and editing

    Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xi- aohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mu- tual self-attention control for consistent image synthesis and editing. InProceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 22560–22570, 2023. 2

    2023

  9. [9]

    Edit transfer: Learning image editing via vision in-context rela- tions.arXiv preprint arXiv:2503.13327, 2025

    Lan Chen, Qi Mao, Yuchao Gu, and Mike Zheng Shou. Edit transfer: Learning image editing via vision in-context rela- tions.arXiv preprint arXiv:2503.13327, 2025. 2, 4, 5

    Pith/arXiv arXiv 2025

  10. [10]

    Sci-lora: Mixture of scientific LoRAs for cross-domain lay paraphras- ing.arXiv preprint arXiv:2505.18867, 2025

    Ming Cheng, Jiaying Gong, and Hoda Eldardiry. Sci-lora: Mixture of scientific LoRAs for cross-domain lay paraphras- ing.arXiv preprint arXiv:2505.18867, 2025. 3

    Pith/arXiv arXiv 2025

  11. [11]

    Turboedit: Text-based image editing us- ing few-step diffusion models

    Gilad Deutch, Rinon Gal, Daniel Garibi, Or Patashnik, and Daniel Cohen-Or. Turboedit: Text-based image editing us- ing few-step diffusion models. InSIGGRAPH Asia 2024 Conference Papers, pages 1–12, 2024. 2

    2024

  12. [12]

    Interpreting the weight space of customized diffusion mod- els.Advances in Neural Information Processing Systems, 37: 137334–137371, 2024

    Amil Dravid, Yossi Gandelsman, Kuan-Chieh Wang, Rameen Abdal, Gordon Wetzstein, Alexei Efros, and Kfir Aberman. Interpreting the weight space of customized diffusion mod- els.Advances in Neural Information Processing Systems, 37: 137334–137371, 2024. 2, 3

    2024

  13. [13]

    Mixture-of-LoRAs: An efficient multitask tuning method for large language models

    Wenfeng Feng, Chuzhan Hao, Yuewei Zhang, Yu Han, and Hao Wang. Mixture-of-LoRAs: An efficient multitask tuning method for large language models. InProceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 11371–11380, 2024. 3

    2024

  14. [14]

    Dreamsim: Learning new dimensions of human visual similarity using synthetic data

    Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dreamsim: Learning new dimensions of human visual similarity using synthetic data. InAdvances in Neural Information Processing Systems, pages 50742–50768. Curran Associates, Inc., 2023. 14

    2023

  15. [15]

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

    Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-or. An image is worth one word: Personalizing text-to-image generation using textual inversion. InThe Eleventh Interna- tional Conference on Learning Representations, 2023. 2

    2023

  16. [16]

    LCM- lookahead for encoder-based text-to-image personalization

    Rinon Gal, Or Lichter, Elad Richardson, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. LCM- lookahead for encoder-based text-to-image personalization. InEuropean Conference on Computer Vision, pages 322–340. Springer, 2024. 2

    2024

  17. [17]

    Relationadapter: Learning and transferring visual re- lation with diffusion transformers

    Yan Gong, Yiren Song, Yicheng Li, Chenglin Li, and Yin Zhang. Relationadapter: Learning and transferring visual re- lation with diffusion transformers. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025. 2, 4, 12, 14

    2025

  18. [18]

    Analogist: Out-of-the-box visual in-context learning with image diffusion model.ACM Transactions on Graphics (TOG), 43(4):1–15, 2024

    Zheng Gu, Shiyuan Yang, Jing Liao, Jing Huo, and Yang Gao. Analogist: Out-of-the-box visual in-context learning with image diffusion model.ACM Transactions on Graphics (TOG), 43(4):1–15, 2024. 5

    2024

  19. [19]

    Discov- ering interpretable directions in the semantic latent space of diffusion models

    Ren´e Haas, Inbar Huberman-Spiegelglas, Rotem Mulayoff, Stella Graßhof, Sami S Brandt, and Tomer Michaeli. Discov- ering interpretable directions in the semantic latent space of diffusion models. In2024 IEEE 18th International Confer- ence on Automatic Face and Gesture Recognition (FG), pages 1–9. IEEE, 2024. 2

    2024

  20. [20]

    Delta de- noising score

    Amir Hertz, Kfir Aberman, and Daniel Cohen-Or. Delta de- noising score. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 2328–2337, 2023. 2

    2023

  21. [21]

    Prompt-to-prompt image editing with cross-attention control

    Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross-attention control. InThe Eleventh Interna- tional Conference on Learning Representations, 2023. 2

    2023

  22. [22]

    Style aligned image generation via shared at- tention

    Amir Hertz, Andrey V oynov, Shlomi Fruchter, and Daniel Cohen-Or. Style aligned image generation via shared at- tention. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4775–4785,

  23. [23]

    Jacobs, Nuria Oliver, Brian Curless, and David H

    Aaron Hertzmann, Charles E. Jacobs, Nuria Oliver, Brian Curless, and David H. Salesin. Image analogies. InProceed- ings of the 28th Annual Conference on Computer Graphics and Interactive Techniques, page 327–340, New York, NY , USA, 2001. Association for Computing Machinery. 1, 2, 3

    2001

  24. [24]

    LoRA: Low-rank adaptation of large language models

    Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 2, 3

    2022

  25. [25]

    Diffusion model-based image editing: A survey.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    Yi Huang, Jiancheng Huang, Yifan Liu, Mingfu Yan, Jiaxi Lv, Jianzhuang Liu, Wei Xiong, He Zhang, Liangliang Cao, and Shifeng Chen. Diffusion model-based image editing: A survey.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 5, 14 9

    2025

  26. [26]

    An edit friendly DDPM noise space: Inversion and manipulations

    Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly DDPM noise space: Inversion and manipulations. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12469– 12478, 2024. 2

    2024

  27. [27]

    Human preference-aligned concept customization benchmark via decomposed evaluation

    Reina Ishikawa, Ryo Fujii, Hideo Saito, and Ryo Hachiuma. Human preference-aligned concept customization benchmark via decomposed evaluation. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 7002– 7011, 2025. 14

    2025

  28. [28]

    Flowedit: Inversion-free text-based editing using pre-trained flow models

    Vladimir Kulikov, Matan Kleiner, Inbar Huberman- Spiegelglas, and Tomer Michaeli. Flowedit: Inversion-free text-based editing using pre-trained flow models. InProceed- ings of the IEEE/CVF International Conference on Computer Vision, pages 19721–19730, 2025. 2

    2025

  29. [29]

    Visualcloze: A universal image generation framework via visual in-context learning

    Zhong-Yu Li, Ruoyi Du, Juncheng Yan, Le Zhuo, Zhen Li, Peng Gao, Zhanyu Ma, and Ming-Ming Cheng. Visualcloze: A universal image generation framework via visual in-context learning. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 18969–18979,

  30. [30]

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maxim- ilian Nickel, and Matthew Le. Flow matching for genera- tive modeling. InThe Eleventh International Conference on Learning Representations, 2023. 3

    2023

  31. [31]

    Flow straight and fast: Learning to generate and transfer data with rectified flow

    Xingchao Liu, Chengyue Gong, and qiang liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. InThe Eleventh International Conference on Learning Representations, 2023. 3

    2023

  32. [32]

    Unifying image pro- cessing as visual prompting question answering

    Yihao Liu, Xiangyu Chen, Xianzheng Ma, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Unifying image pro- cessing as visual prompting question answering. InInterna- tional Conference on Machine Learning, pages 30873–30891. PMLR, 2024. 2

    2024

  33. [33]

    Decoupled weight de- cay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight de- cay regularization. InInternational Conference on Learning Representations, 2019. 12

    2019

  34. [34]

    Pairedit: Learning semantic variations for exemplar-based image editing

    Haoguang Lu, Jiacheng Chen, Zhenguo Yang, Aurele To- hokantche Gnanha, Fu Lee Wang, Li Qing, and Xudong Mao. Pairedit: Learning semantic variations for exemplar-based image editing. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025. 2

    2025

  35. [35]

    Zero-shot unsupervised and text-based audio editing using DDPM inversion

    Hila Manor and Tomer Michaeli. Zero-shot unsupervised and text-based audio editing using DDPM inversion. InPro- ceedings of the 41st International Conference on Machine Learning, pages 34603–34629. PMLR, 2024. 2

    2024

  36. [36]

    Omni-Effects: Unified and spatially-controllable visual effects generation.arXiv preprint arXiv:2508.07981, 2025

    Fangyuan Mao, Aiming Hao, Jintao Chen, Dongxia Liu, Xiaokun Feng, Jiashu Zhu, Meiqi Wu, Chubin Chen, Jia- hong Wu, and Xiangxiang Chu. Omni-Effects: Unified and spatially-controllable visual effects generation.arXiv preprint arXiv:2508.07981, 2025. 3

    arXiv 2025

  37. [37]

    SDEdit: Guided image synthesis and editing with stochastic differential equations

    Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2022. 2

    2022

  38. [38]

    Gpt-4o system card.arXiv preprint arXiv:2410.21276, 2024

    OpenAI. Gpt-4o system card.arXiv preprint arXiv:2410.21276, 2024. 12

    Pith/arXiv arXiv 2024

  39. [39]

    Magnitude invariant parametrizations improve hypernetwork learning

    Jose Javier Gonzalez Ortiz, John Guttag, and Adrian V Dalca. Magnitude invariant parametrizations improve hypernetwork learning. InThe Twelfth International Conference on Learn- ing Representations, 2024. 2

    2024

  40. [40]

    Zero-shot image-to-image translation

    Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. InACM SIGGRAPH 2023 conference proceed- ings, pages 1–11, 2023. 2

    2023

  41. [41]

    Dreambench++: A human-aligned bench- mark for personalized image generation

    Yuang Peng, Yuxin Cui, Haomiao Tang, Zekun Qi, Runpei Dong, Jing Bai, Chunrui Han, Zheng Ge, Xiangyu Zhang, and Shu-Tao Xia. Dreambench++: A human-aligned bench- mark for personalized image generation. InThe Thirteenth International Conference on Learning Representations, 2025. 14

    2025

  42. [42]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. InProceedings of the 38th International Conference on Machine Learning, pages 8748–8763. PMLR, 2021....

    2021

  43. [43]

    Hierarchical text-conditional image genera- tion with clip latents.arXiv preprint arXiv:2204.06125, 2022

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image genera- tion with clip latents.arXiv preprint arXiv:2204.06125, 2022. 2

    Pith/arXiv arXiv 2022

  44. [44]

    Deep visual analogy-making.Advances in neural information processing systems, 28, 2015

    Scott E Reed, Yi Zhang, Yuting Zhang, and Honglak Lee. Deep visual analogy-making.Advances in neural information processing systems, 28, 2015. 2

    2015

  45. [45]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. InIEEE/CVF Con- ference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 10674– 10685. IEEE, 2022. 2, 3

    2022

  46. [46]

    Dreambooth: Fine tuning text-to-image diffusion models for subject-driven gen- eration

    Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven gen- eration. InProceedings of the IEEE/CVF conference on com- puter vision and pattern recognition, pages 22500–22510,

  47. [47]

    Cloneofsimo/lora: Low-rank adaptation for fast text-to-image diffusion fine-tuning, 2023

    Simo Ryu. Cloneofsimo/lora: Low-rank adaptation for fast text-to-image diffusion fine-tuning, 2023. 3

    2023

  48. [48]

    ZipLoRA: Any subject in any style by effectively merging LoRAs

    Viraj Shah, Nataniel Ruiz, Forrester Cole, Erika Lu, Svet- lana Lazebnik, Yuanzhen Li, and Varun Jampani. ZipLoRA: Any subject in any style by effectively merging LoRAs. In European Conference on Computer Vision, pages 422–438. Springer, 2024. 3

    2024

  49. [49]

    Emu edit: Precise image editing via recognition and genera- tion tasks

    Shelly Sheynin, Adam Polyak, Uriel Singer, Yuval Kirstain, Amit Zohar, Oron Ashual, Devi Parikh, and Yaniv Taigman. Emu edit: Precise image editing via recognition and genera- tion tasks. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8871–8879,

  50. [50]

    LoRA of change: Learning to generate LoRA for the editing instruction from a single before-after image pair.arXiv preprint arXiv:2411.19156, 2024

    Xue Song, Jiequan Cui, Hanwang Zhang, Jiaxin Shi, Jingjing Chen, Chi Zhang, and Yu-Gang Jiang. LoRA of change: Learning to generate LoRA for the editing instruction from a single before-after image pair.arXiv preprint arXiv:2411.19156, 2024. 2, 5 10

    Pith/arXiv arXiv 2024

  51. [51]

    Ominicontrol: Minimal and univer- sal control for diffusion transformer

    Zhenxiong Tan, Songhua Liu, Xingyi Yang, Qiaochu Xue, and Xinchao Wang. Ominicontrol: Minimal and univer- sal control for diffusion transformer. InProceedings of the IEEE/CVF International Conference on Computer Vision,

  52. [52]

    Gemma 3 technical report.arXiv preprint arXiv:2503.19786,

    Gemma Team, Aishwarya Kamath, Johan Ferret, Shreya Pathak, Nino Vieillard, Ramona Merhej, Sarah Perrin, Ta- tiana Matejovicova, Alexandre Ram´e, Morgane Rivi`ere, et al. Gemma 3 technical report.arXiv preprint arXiv:2503.19786,

    Pith/arXiv arXiv

  53. [53]

    Training-free consistent text-to-image generation.ACM Transactions on Graphics (TOG), 43(4):1–18, 2024

    Yoad Tewel, Omri Kaduri, Rinon Gal, Yoni Kasten, Lior Wolf, Gal Chechik, and Yuval Atzmon. Training-free consistent text-to-image generation.ACM Transactions on Graphics (TOG), 43(4):1–18, 2024. 2

    2024

  54. [54]

    Add-it: Training-free object insertion in images with pretrained diffusion models

    Yoad Tewel, Rinon Gal, Dvir Samuel, Yuval Atzmon, Lior Wolf, and Gal Chechik. Add-it: Training-free object insertion in images with pretrained diffusion models. InThe Thirteenth International Conference on Learning Representations, 2025. 2

    2025

  55. [55]

    SigLIP 2: Multilingual vision-language en- coders with improved semantic understanding, localization, and dense features.arXiv preprint arXiv:2502.14786, 2025

    Michael Tschannen, Alexey Gritsenko, Xiao Wang, Muham- mad Ferjad Naeem, Ibrahim Alabdulmohsin, Nikhil Parthasarathy, Talfan Evans, Lucas Beyer, Ye Xia, Basil Mustafa, et al. SigLIP 2: Multilingual vision-language en- coders with improved semantic understanding, localization, and dense features.arXiv preprint arXiv:2502.14786, 2025. 7

    Pith/arXiv arXiv 2025

  56. [56]

    Plug-and-play diffusion features for text-driven image-to- image translation

    Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to- image translation. InProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, pages 1921–1930, 2023. 2

    1921

  57. [57]

    Images speak in images: A generalist painter for in-context visual learning

    Xinlong Wang, Wen Wang, Yue Cao, Chunhua Shen, and Tiejun Huang. Images speak in images: A generalist painter for in-context visual learning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6830–6839, 2023. 2

    2023

  58. [58]

    In- context learning unlocked for diffusion models.Advances in Neural Information Processing Systems, 36:8542–8562, 2023

    Zhendong Wang, Yifan Jiang, Yadong Lu, Pengcheng He, Weizhu Chen, Zhangyang Wang, Mingyuan Zhou, et al. In- context learning unlocked for diffusion models.Advances in Neural Information Processing Systems, 36:8542–8562, 2023. 2

    2023

  59. [59]

    Mixture of loRA experts

    Xun Wu, Shaohan Huang, and Furu Wei. Mixture of loRA experts. InThe Twelfth International Conference on Learning Representations, 2024. 3

    2024

  60. [60]

    Omnigen: Unified image genera- tion

    Shitao Xiao, Yueze Wang, Junjie Zhou, Huaying Yuan, Xin- grun Xing, Ruiran Yan, Chaofan Li, Shuting Wang, Tiejun Huang, and Zheng Liu. Omnigen: Unified image genera- tion. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 13294–13304, 2025. 1

    2025

  61. [61]

    Imagebrush: Learning visual in-context instructions for exemplar-based image manipula- tion.Advances in Neural Information Processing Systems, 36:48723–48743, 2023

    Yifan Yang, Houwen Peng, Yifei Shen, Yuqing Yang, Han Hu, Lili Qiu, Hideki Koike, et al. Imagebrush: Learning visual in-context instructions for exemplar-based image manipula- tion.Advances in Neural Information Processing Systems, 36:48723–48743, 2023. 2

    2023

  62. [62]

    Scaling vision transformers

    Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lu- cas Beyer. Scaling vision transformers. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12104–12113, 2022. 3

    2022

  63. [63]

    Sigmoid loss for language image pre-training

    Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. InProceedings of the IEEE/CVF international conference on computer vision, pages 11975–11986, 2023. 2

    2023

  64. [64]

    Subject or style: Adap- tive and training-free mixture of LoRAs.arXiv preprint arXiv:2508.02165, 2025

    Jia-Chen Zhang and Yu-Jie Xiong. Subject or style: Adap- tive and training-free mixture of LoRAs.arXiv preprint arXiv:2508.02165, 2025. 3

    Pith/arXiv arXiv 2025

  65. [65]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3836–3847, 2023. 2

    2023

  66. [66]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 586–595, 2018. 5

    2018

  67. [67]

    Enabling instructional image editing with in-context genera- tion in large scale diffusion transformer

    Zechuan Zhang, Ji Xie, Yu Lu, Zongxin Yang, and Yi Yang. Enabling instructional image editing with in-context genera- tion in large scale diffusion transformer. InThe Thirty-ninth Annual Conference on Neural Information Processing Sys- tems, 2025. 1 11 A. Experimental Details A.1. Implementation details In all our experiments, we train for 10K steps on 1 ...

    2025

  68. [68]

    metrics":{

    consistency: Asses how well the edited image B’ maintains the context of the original image B. Does it preserve the identity, objects, and layout in B that did not require a change, based on the infered transformation from A to A’? Consider in your evaluations other visual factors such as the localization of the edits, existence of redundant elements, sty...

  69. [69]

    Are there missing elements, are there redundant elements? Quantify the precision of the editing

    editing accuracy: Evaluate how closely B’ applies the transformation seen from A to A’. Are there missing elements, are there redundant elements? Quantify the precision of the editing

  70. [70]

    better":<1 or 2>,

    consistency: Asses how well the edited image B’ maintains the context of the original image B. Does it preserve the identity, objects, and layout in B that did not require a change, based on the inferred transformation from A to A’? Consider in your evaluations other visual factors such as the localization of the edits, existence of redundant elements, st...