Pith. sign in

REVIEW 4 major objections 4 minor 50 references

Relighting improves when trained as illumination feature transport, with cross-identity light-matched supervision, rather than as generic image-to-image editing.

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 16:50 UTC pith:ROCA7G4D

load-bearing objection A useful training loss and dataset for flow-based relighting, but the key invariance assumption is unverified and the main evaluation is closed-loop. the 4 major comments →

arxiv 2607.17833 v1 pith:ROCA7G4D submitted 2026-07-20 cs.CV

Consistent Feature Transport for Image Relighting

Bohan Zhang , Huanwei Liang , Yuhan He , Hongteng Xu , Quxiao Chao , Luoqi Liu , Dixin Luo , Ting Liu This is my paper
classification cs.CV
keywords image relightingrectified flowfeature transportillumination consistencydiffusion-based editingportrait relighting datasetstyle transfercontent preservation
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 image relighting should be modeled explicitly as the transport of illumination features between a source and a target distribution, rather than as conditional generation or instance-level translation. It introduces Consistent Feature Transport (CFT), a training principle built on rectified flow that adds a direct source-to-target transport loss supervised by image pairs that share the same illumination transformation but differ in identity and scene content. This design encourages the model to isolate lighting-specific changes while preserving non-lighting content like identity and geometry. Experiments on a newly constructed portrait relighting dataset show consistent gains over several baseline methods on fidelity, perceptual, and lighting-specific metrics, and the same training principle transfers to style transfer. The paper also contributes a 34,695-pair portrait relighting dataset with complex and diverse illumination effects.

Core claim

The central claim is that a dual-transport training objective—combining noise-to-image generation (L1, L2) with a cross-instance illumination-consistent transport loss (L3)—improves relighting quality and content preservation compared with standard conditional-generation training. L3 uses the parallelogram law of rectified flow to approximate the direct source-to-target velocity and supervises it with a different image pair (x'_src, x'_tgt) that shares the same illumination transformation as the original pair but has different content. The authors report that this cross-instance supervision reduces the variance of the predicted transport field and yields consistent improvements in SSIM, PSNR

What carries the argument

The parallelogram-law velocity identity for rectified flow: along a linear interpolation path, the direct source-to-target direction is approximated as v_direct_t = v_theta(z_tgt_t, t, x_src, c_tgt) - v_theta(z_src_t, t, x_src, c_src). This identity lets the model compute an explicit illumination-transport velocity without inversion, and supervising it with cross-instance illumination-consistent pairs (same text condition, different identity) is what forces the model to separate illumination features from content.

Load-bearing premise

The loss L3 assumes that two different images generated from the same text description undergo the exact same illumination transformation, but the paper provides no measured or physical verification that the lighting change is actually equivalent across different identities and scenes.

What would settle it

Construct pairs where the same text prompt produces visibly different lighting effects (e.g., the same prompt but one image lit from the left and another from the right) and check whether CFT still improves relighting metrics; if it does, the gains are not attributable to shared illumination. Alternatively, on a real multi-illumination dataset with known light source positions, test whether L3 reduces irradiance error beyond using original or random pairs as supervision.

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

If this is right

  • CFT can be applied as a drop-in training loss for flow-based image editors, improving relighting fidelity without requiring extra control signals at inference.
  • The training principle generalizes to other editing tasks such as style transfer, improving structural preservation and perceptual fidelity at a modest cost in output diversity (slightly higher FID).
  • The constructed portrait relighting dataset, with 34,249 training and 446 test pairs across diverse complex lighting effects, provides a new resource for training and benchmarking relighting methods.
  • The explicitly modeled feature transport yields lower variance along the source-to-target trajectory, which the authors interpret as more consistent and coherent illumination changes.
  • CFT transfers to an unseen real-world multi-illumination dataset, suggesting that the learned illumination features generalize beyond the synthetic training distribution.

Where Pith is reading between the lines

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

  • If the shared-illumination assumption holds under a physical lighting model, the same consistent-transport supervision could be extended to other text-specified attribute edits where a transformation is content-invariant, such as material changes or seasonal shifts.
  • A stricter evaluation on multi-illumination captures with measured light directions could test whether CFT truly learns an illumination-invariant transport rather than just a prompt-conditioned content mapping.
  • The approach could be combined with physics-based conditions (e.g., environment maps or intrinsic decompositions as the conditioning signal c) to give the shared-illumination supervisory pairs a measurable grounding, potentially improving robustness on out-of-distribution lighting.
  • The reliance on text-to-image generation and LLM-based filtering for dataset construction means the real-world lighting physics may be underrepresented; a domain-randomization or physically simulated validation set would clarify how much of the gain survives in deployment.

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 / 4 minor

Summary. The paper proposes Consistent Feature Transport (CFT), a training principle for flow-based image editing models, applied to relighting. Built on rectified flow, CFT augments the standard noise-to-image losses L1 and L2 with a third loss L3 that supervises a 'direct' source-to-target velocity v_direct_t (Eq. 5) with the latent displacement z'_tgt - z'_src of a different image pair that is assumed to share the same illumination transformation as the original pair. The authors construct a synthetic portrait relighting dataset with text-based lighting descriptions, and report improvements on their synthetic test set, on transfer to the Multi-Illumination dataset, and for style transfer. The central claim is that L3 encourages illumination-consistent transport while preserving non-lighting content, and that CFT is a drop-in loss for flow-based editors.

Significance. If the central premise is validated, CFT would be a simple, model-agnostic training loss applicable to any rectified-flow editor, with plausible potential for relighting and other editing tasks. The synthetic dataset could also be useful for future research. The paper is clearly written and the experimental setup includes several baselines and ablations. However, the significance is conditional on the unverified assumption in Eq. (6) that identical text prompts induce identical, content-invariant illumination displacements in latent space. The evaluation is mostly closed-loop, and the only external transfer experiment is a single dataset with no statistical testing. These concerns are load-bearing for the claimed contribution.

major comments (4)
  1. [Sec. 3.1, Eq. (5)] The definition v_direct_t = v_theta(z_tgt_t, t, xsrc, ctgt) - v_theta(z_src_t, t, xsrc, csrc) is presented as 'the corresponding velocity along this path', but it is not the ODE velocity at the direct-path state z_direct_t; it is the difference of two model evaluations at different states and conditions. This is only an exact derivative of the parallelogram path if the model velocities are perfect. The paper does not analyze the approximation error or its influence during training. Please clarify the derivation and provide empirical evidence that this quantity behaves as intended, e.g., comparison against the true velocity on a known displacement.
  2. [Sec. 3.1, Eq. (6) and Sec. 3.2 dataset construction] The central premise that (x'_src, x'_tgt) 'shares the same illumination transformation' as (x_src, x_tgt) is asserted but never demonstrated. Same text prompt does not imply that the latent displacement z'_tgt - z'_src equals z_tgt - z_src; the generated relighting effect may depend on content, geometry, and albedo. The paper provides no quantitative evidence, such as clustering of latent displacements or comparison of estimated lighting maps for same-prompt pairs. Without this, L3 may simply teach a prompt-specific average displacement, and the improvements in Table 4 could reflect variance reduction rather than correct illumination transport.
  3. [Sec. 4.2 and Sec. 4.3] The evaluation is largely closed-loop: both training and test pairs come from the same generative pipeline with the same text-conditioned image generators and the same filtering criteria. The only external test is the Multi-Illumination transfer (Table 2), which is a single dataset, reports no error bars or significance tests, and does not isolate the effect of the cross-instance assumption. To support the claim that CFT learns a general illumination transform, the paper should include external evaluation on real paired relighting data with known illumination changes, multiple seeds, and statistical testing.
  4. [Sec. 4.3, Figure 5] The paper interprets the lower variance of v_direct_t under CFT as evidence of 'more coherent and consistent feature transport'. This is not convincing: reducing variance could also indicate collapse to a common displacement that ignores content-dependent lighting interactions. The paper should complement this with a metric that directly measures the correctness of the illumination change, e.g., estimated irradiance error or light-direction/color error for pairs whose illumination transformation is known. The current evidence does not distinguish consistency from collapse.
minor comments (4)
  1. [Table 5] Typo: 'QmniStyle' in the table header appears to be a typo for 'OmniStyle' used in the text. Please fix.
  2. [Table 3] Column header inconsistency: the text mentions 'Image Aesthetics (IA)' but the table uses 'PA'. Please align the abbreviation.
  3. [Sec. 3.1, Eq. (4)-(5)] The notation z_src_t and z_tgt_t is used without explicit definition of how they are sampled. Clarify that they are on the noise-to-source and noise-to-target trajectories sharing the same initial noise z0, and define c_src and c_tgt precisely.
  4. [Sec. 4.1] The hyperparameter alpha=0.1 is selected based on the synthetic test set. Since the test set is from the same generative pipeline as training, the selection may be overfit. Report the sensitivity on the external Multi-Illumination dataset as well.

Circularity Check

0 steps flagged

No significant circularity; CFT's training objective is an independent derivation, with one minor non-load-bearing self-citation.

full rationale

The paper's derivation chain is: L1/L2 are standard rectified-flow objectives (Eqs. 2-3), the direct-path velocity v_direct_t follows the parallelogram law from external inversion-free editing methods (Eqs. 4-5), and L3 (Eq. 6) supervises v_direct_t with the cross-instance displacement z'_tgt - z'_src. This is not circular: L3 is a constructed training signal, not a fitted parameter renamed as a prediction, and the model must still learn to produce the final relit image from noise/source under the prompt. The cross-instance supervision relies on the operational assumption that same-text-description pairs share the same illumination transformation (Sec. 3.1, Eq. 6; Sec. 3.2). That premise is unverified and physically nontrivial, but it is an assumption about data generation rather than an equivalence by construction: the paper does not define illumination consistency in terms of the model's output, and the Multi-Illumination transfer experiment (Table 2) provides an external out-of-distribution check. The only self-citation is [47] (first author Bohan Zhang) in related work on disentangled representations; it is not load-bearing. No uniqueness theorem, ansatz smuggling via self-citation, or renaming of a known result appears. The closed-loop synthetic train/test pipeline is a data-quality concern, not a logical circularity, especially since all compared methods share the same training data and the external transfer still favors CFT.

Axiom & Free-Parameter Ledger

1 free parameters · 5 axioms · 0 invented entities

The central method rests on a novel training loss whose supervision is defined by a data-generation assumption (same prompt = same illumination transformation) and on an unproven velocity decomposition (Eq. 5). The only fitted parameter is α, chosen via hyperparameter sweep. No new physical entities are introduced.

free parameters (1)
  • alpha (α) = 0.1
    Weight of the consistent feature transport loss L3 in Eq. (7); selected via hyperparameter sweep on {0.05, 0.1, 0.3, 0.6} reported in Fig. 6, with no separate validation set described.
axioms (5)
  • standard math Rectified flow / flow matching is a valid generative framework.
    The method is built on [27,30]; this is a standard framework in the field.
  • standard math Linear interpolation / parallelogram law gives the direct source-to-target path.
    For rectified flow, z_t = (1-t)z_0 + t z_1, so Eq. (4) is algebraically exact under linear interpolation.
  • ad hoc to paper Velocity along the direct path equals vθ(z_tgt_t) − vθ(z_src_t)
    Eq. (5) asserts this without proof; it is not the ODE velocity at z_direct_t, and no justification is given.
  • domain assumption Image pairs generated from the same text prompt share the same illumination transformation.
    The L3 ground truth in Eq. (6) relies on (x'_src, x'_tgt) being lit by the same transformation as (x_src, x_tgt), but this is never measured or verified.
  • domain assumption Pretrained generators (Nano Banana, Seedream, RoboNeo) and filters (QwenVL, EditScore, GPT-4o) produce reliable relighting ground truth.
    The entire dataset depends on these models' outputs and judgments; no independent physical ground truth is available.

pith-pipeline@v1.3.0-alltime-deepseek · 13055 in / 13269 out tokens · 129910 ms · 2026-08-01T16:50:54.281527+00:00 · methodology

0 comments
read the original abstract

Image relighting modifies illumination while preserving non-lighting content such as identity and geometry. Existing diffusion-based methods often suffer from unstable illumination changes or inconsistent content preservation under complex lighting, as they lack an explicit mechanism to learn feature transformations between images. We reformulate relighting as an illumination feature transport problem and introduce Consistent Feature Transport (CFT), a training principle that explicitly enforces illumination-consistent transport between source and target image distributions. Built upon rectified flow, CFT jointly models noise-to-image generation and illumination-consistent source-to-target transport through trajectory-level supervision. This dual-transport formulation encourages isolation of illumination-specific variations while preserving content-aligned features. To support complex lighting scenarios, we construct a large-scale portrait relighting dataset with diverse relighting effects. Experiments show consistent improvements over existing state-of-the-art relighting approaches and demonstrate that CFT can generalize to other editing tasks, including style transfer. Code is available at https://github.com/Dixin-Lab/CFT.

Figures

Figures reproduced from arXiv: 2607.17833 by Bohan Zhang, Dixin Luo, Hongteng Xu, Huanwei Liang, Luoqi Liu, Quxiao Chao, Ting Liu, Yuhan He.

Figure 1
Figure 1. Figure 1: Overall framework of the proposed method. condition. The training objective is defined as: \label {equ:problem_formulation} \mathcal {L}_{\mathrm {relight}}(\theta ) = \mathbb {E}_{z_0 \sim \pi _0,\; z_{\mathrm {tgt}},\; t \sim \mathcal {U}(0,1)} \left [ \left \| v_{\theta }(z_t, t, x_{\mathrm {src}}, c) - (z_{tgt} - z_0) \right \|^2 \right ], (1) where z0 denotes a sample drawn from the prior distribution… view at source ↗
Figure 2
Figure 2. Figure 2: Dataset Construction Pipeline. Text Source Image Target Image Generate neon light from the front￾right — medium-bright colorful ambient light (a mix of purple and red) evoking an urban nightscape with soft bokeh in the background. Try adjusting the light to a noon sunlight mode, shining brightly from above with neutral tones and strong lighting. Combine it with dappled light effects from sunlight filtering… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the constructed dataset. they lack diverse complex lighting effects and multiple light configurations, such as different colors, spatially varying illumination, and non-uniform shadows. This limitation reduces their effectiveness for training and evaluating models designed for realistic and controllable image relighting. To address this gap, we construct a portrait relighting dataset with diver… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison results. learns illumination feature transport from paired data, leading to a more ef￾ficient training process. Compared with decomposition-based editing methods (Intrinsic Edit and Latent Intrinsic), control-based editing approaches gener￾ally achieve stronger performance. This may be attributed to the inaccuracy in the intrinsic decomposition and lighting-related intrinsic editing,… view at source ↗
Figure 5
Figure 5. Figure 5: Variance of the predicted velocity. 0.05 0.1 0.3 0.6 0.895 0.900 0.905 0.910 0.915 0.920 S S I M (a) SSIM 0.05 0.1 0.3 0.6 21.75 22.00 22.25 22.50 22.75 23.00 23.25 23.50 P S N R (b) PSNR 0.05 0.1 0.3 0.6 0.096 0.098 0.100 0.102 0.104 L P I P S (c) LPIPS 0.05 0.1 0.3 0.6 19 20 21 22 23 F I D (d) FID [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: The impact of α. tions, which is less effective for learning meaningful feature transformations. Randomly sampled pairs (x rand src , xrand tgt ) fail to provide meaningful cues for the relighting transformation. In contrast, CFT leverages image pairs sharing the same illumination transformation as supervision, which significantly improves all evaluation metrics. This result validates the effectiveness of … view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparison results for the style transfer task. proposed consistent feature transport (CFT), the generated images exhibit a more balanced style transformation. Our method better preserves the structural content of the source image while transferring the key stylistic patterns from the reference style image, resulting in visually coherent and stylistically faithful outputs across different scene… 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

50 extracted references · 7 linked inside Pith

  1. [1]

    arXiv preprint arXiv:2511.21631 (2025)

    Bai, S., Cai, Y., Chen, R., Chen, K., Chen, X., Cheng, Z., Deng, L., Ding, W., Gao, C., Ge, C., et al.: Qwen3-vl technical report. arXiv preprint arXiv:2511.21631 (2025)

    Pith/arXiv arXiv 2025

  2. [2]

    In: Proceedings of the SIGGRAPH Asia 2025 Conference Papers

    Bharadwaj, S., Feng, H., Becherini, G., Fernandez Abrevaya, V., Black, M.J.: Gen- lit: Reformulating single-image relighting as video generation. In: Proceedings of the SIGGRAPH Asia 2025 Conference Papers. pp. 1–12 (2025)

    2025

  3. [3]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 18392–18402 (2023)

    2023

  4. [4]

    In: Pro- ceedings of the IEEE/CVF international conference on computer vision

    Cao, M., Wang, X., Qi, Z., Shan, Y., Qie, X., Zheng, Y.: Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In: Pro- ceedings of the IEEE/CVF international conference on computer vision. pp. 22560– 22570 (2023)

    2023

  5. [5]

    In: Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision

    Chadebec, C., Tasar, O., Sreetharan, S., Aubin, B.: Lbm: Latent bridge matching for fast image-to-image translation. In: Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision. pp. 29086–29098 (2025)

    2025

  6. [6]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Chaturvedi, S., Ren, M., Hold-Geoffroy, Y., Liu, J., Dorsey, J., Shu, Z.: Synthlight: Portrait relighting with diffusion model by learning to re-render synthetic faces. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 369–379 (2025)

    2025

  7. [7]

    Advances in Neural Information Processing Systems 38, 94919–94945 (2026)

    Chen, C., Guo, P., Song, L., Lu, J., Qian, R., Fu, T.J., Wang, X., Liu, W., Yang, Y., Schwing, A.: Car-flow: Condition-aware reparameterization aligns source and tar- get for better flow matching. Advances in Neural Information Processing Systems 38, 94919–94945 (2026)

    2026

  8. [8]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Choi, J.M., Wang, A., Peers, P., Bhattad, A., Sengupta, R.: Scribblelight: Single image indoor relighting with scribbles. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 5720–5731 (2025)

    2025

  9. [9]

    In: European Conference on Computer Vision

    Choi, Y., Kwak, S., Lee, K., Choi, H., Shin, J.: Improving diffusion models for authentic virtual try-on in the wild. In: European Conference on Computer Vision. pp. 206–235. Springer (2024)

    2024

  10. [10]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Dalva, Y., Venkatesh, K., Yanardag, P.: Fluxspace: Disentangled semantic edit- ing in rectified flow models. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 13083–13092 (2025)

    2025

  11. [11]

    In: Proceedings of the 27th annual conference on Computer graphics and interactive techniques

    Debevec, P., Hawkins, T., Tchou, C., Duiker, H.P., Sarokin, W., Sagar, M.: Ac- quiring the reflectance field of a human face. In: Proceedings of the 27th annual conference on Computer graphics and interactive techniques. pp. 145–156 (2000)

    2000

  12. [12]

    In: Forty-second International Conference on Machine Learning (2025)

    Deng, Y., He, X., Mei, C., Wang, P., Tang, F.: Fireflow: Fast inversion of recti- fied flow for image semantic editing. In: Forty-second International Conference on Machine Learning (2025)

    2025

  13. [13]

    ACM Transactions on Graphics (ToG)35(6), 1–12 (2016)

    Hasinoff, S.W., Sharlet, D., Geiss, R., Adams, A., Barron, J.T., Kainz, F., Chen, J., Levoy, M.: Burst photography for high dynamic range and low-light imaging on mobile cameras. ACM Transactions on Graphics (ToG)35(6), 1–12 (2016)

    2016

  14. [14]

    arXiv preprint arXiv:2005.05460 (2020)

    Helou, M.E., Zhou, R., Barthas, J., Süsstrunk, S.: Vidit: Virtual image dataset for illumination transfer. arXiv preprint arXiv:2005.05460 (2020)

    Pith/arXiv arXiv 2005

  15. [15]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Hold-Geoffroy, Y., Athawale, A., Lalonde, J.F.: Deep sky modeling for single im- age outdoor lighting estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6927–6935 (2019) Consistent Feature Transport for Image Relighting 17

    2019

  16. [16]

    arXiv preprint arXiv:2410.21276 (2024)

    Hurst, A., Lerer, A., Goucher, A.P., Perelman, A., Ramesh, A., Clark, A., Os- trow, A., Welihinda, A., Hayes, A., Radford, A., et al.: Gpt-4o system card. arXiv preprint arXiv:2410.21276 (2024)

    Pith/arXiv arXiv 2024

  17. [17]

    ACM Trans

    Kalantari, N.K., Ramamoorthi, R., et al.: Deep high dynamic range imaging of dynamic scenes. ACM Trans. Graph.36(4), 144–1 (2017)

    2017

  18. [18]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4401–4410 (2019)

    2019

  19. [19]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Kim, H., Jang, M., Yoon, W., Lee, J., Na, D., Woo, S.: Switchlight: Co-design of physics-driven architecture and pre-training framework for human portrait re- lighting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 25096–25106 (2024)

    2024

  20. [20]

    In: The Fourteenth International Conference on Learning Representations (2026)

    Kim, J., Hong, Y., Park, J., Ye, J.C.: Flowalign: Trajectory-regularized, inversion- free flow-based image editing. In: The Fourteenth International Conference on Learning Representations (2026)

    2026

  21. [21]

    Kocsis,P.,Philip,J.,Sunkavalli,K.,Nießner,M.,Hold-Geoffroy,Y.:Lightit:Illumi- nationmodelingandcontrolfordiffusionmodels.In:ProceedingsoftheIEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9359–9369 (2024)

    2024

  22. [22]

    In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition

    Kocsis, P., Sitzmann, V., Nießner, M.: Intrinsic image diffusion for indoor single- view material estimation. In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition. pp. 5198–5208 (2024)

    2024

  23. [23]

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

    Kulikov, V., Kleiner, M., Huberman-Spiegelglas, I., Michaeli, T.: Flowedit: Inversion-free text-based editing using pre-trained flow models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19721–19730 (2025)

    2025

  24. [24]

    1 kontext: Flow match- ing for in-context image generation and editing in latent space

    Labs, B.F., Batifol, S., Blattmann, A., Boesel, F., Consul, S., Diagne, C., Dock- horn, T., English, J., English, Z., Esser, P., et al.: Flux. 1 kontext: Flow match- ing for in-context image generation and editing in latent space. arXiv preprint arXiv:2506.15742 (2025)

    Pith/arXiv arXiv 2025

  25. [25]

    In: European Conference on Computer Vision

    Li, Z., Shi, J., Bi, S., Zhu, R., Sunkavalli, K., Hašan, M., Xu, Z., Ramamoorthi, R., Chandraker, M.: Physically-based editing of indoor scene lighting from a single image. In: European Conference on Computer Vision. pp. 555–572. Springer (2022)

    2022

  26. [26]

    arXiv preprint arXiv:2508.14814 (2025)

    Li, Z., Zhu, L., Shen, H., Ran, L., Liu, W., Wang, X.: Translight: Image- guided customized lighting control with generative decoupling. arXiv preprint arXiv:2508.14814 (2025)

    Pith/arXiv arXiv 2025

  27. [27]

    In: The Eleventh International Conference on Learning Representations (2023)

    Lipman, Y., Chen, R.T.Q., Ben-Hamu, H., Nickel, M., Le, M.: Flow matching for generative modeling. In: The Eleventh International Conference on Learning Representations (2023)

    2023

  28. [28]

    Advances in Neural Information Processing Systems36, 36951–36962 (2023)

    Liu, I., Chen, L., Fu, Z., Wu, L., Jin, H., Li, Z., Wong, C.M.R., Xu, Y., Ramamoor- thi, R., Xu, Z., et al.: Openillumination: A multi-illumination dataset for inverse rendering evaluation on real objects. Advances in Neural Information Processing Systems36, 36951–36962 (2023)

    2023

  29. [29]

    Advances in Neural Information Processing Systems38, 82052–82080 (2026)

    Liu, P., Yuan, H., Dong, B., Xing, J., Wang, J., Zhao, R., Chen, W., Wang, F.: Unilumos: Fast and unified image and video relighting with physics-plausible feed- back. Advances in Neural Information Processing Systems38, 82052–82080 (2026)

    2026

  30. [30]

    In: The Eleventh International Conference on Learning Representations (2023)

    Liu, X., Gong, C., qiang liu: Flow straight and fast: Learning to generate and transfer data with rectified flow. In: The Eleventh International Conference on Learning Representations (2023)

    2023

  31. [31]

    arXiv preprint arXiv:2506.14549 (2025) 18 B

    Liu, Y., Xiao, W., Wang, Q., Chen, J., Wang, S., Wang, Y., Wu, X., Tang, Y.: Dreamlight: Towards harmonious and consistent image relighting. arXiv preprint arXiv:2506.14549 (2025) 18 B. Zhang et al

    Pith/arXiv arXiv 2025

  32. [32]

    In: The Fourteenth International Conference on Learning Representations (2026)

    Luo, X., Wang, J., Wu, C., Xiao, S., Jiang, X., Lian, D., Zhang, J., Liu, D., Liu, Z.: Editscore: Unlocking online RL for image editing via high-fidelity reward modeling. In: The Fourteenth International Conference on Learning Representations (2026)

    2026

  33. [33]

    ACM Transactions on Graphics (TOG)44(4), 1–13 (2025)

    Lyu, L., Deschaintre, V., Hold-Geoffroy, Y., Hašan, M., Yoon, J.S., Leimkühler, T., Theobalt, C., Georgiev, I.: Intrinsicedit: Precise generative image manipulation in intrinsic space. ACM Transactions on Graphics (TOG)44(4), 1–13 (2025)

    2025

  34. [34]

    In: Proceed- ings of the Special Interest Group on Computer Graphics and Interactive Tech- niques Conference Conference Papers

    Magar, N., Hertz, A., Tabellion, E., Pritch, Y., Rav-Acha, A., Shamir, A., Hoshen, Y.: Lightlab: Controlling light sources in images with diffusion models. In: Proceed- ings of the Special Interest Group on Computer Graphics and Interactive Tech- niques Conference Conference Papers. pp. 1–11 (2025)

    2025

  35. [35]

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

    Murmann, L., Gharbi, M., Aittala, M., Durand, F.: A dataset of multi-illumination images in the wild. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4080–4089 (2019)

    2019

  36. [36]

    In: The Thirty-ninth Annual Conference on Neural Information Pro- cessing Systems (2025)

    Pan, Y., Bareinboim, E.: Counterfactual image editing with disentangled causal latent space. In: The Thirty-ninth Annual Conference on Neural Information Pro- cessing Systems (2025)

    2025

  37. [37]

    ACM Trans

    Pandey, R., Orts-Escolano, S., Legendre, C., Haene, C., Bouaziz, S., Rhemann, C., Debevec, P.E., Fanello, S.R.: Total relighting: learning to relight portraits for background replacement. ACM Trans. Graph.40(4), 43–1 (2021)

    2021

  38. [38]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition

    Ren, M., Xiong, W., Yoon, J.S., Shu, Z., Zhang, J., Jung, H., Gerig, G., Zhang, H.: Relightful harmonization: Lighting-aware portrait background replacement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition. pp. 6452–6462 (2024)

    2024

  39. [39]

    arXiv preprint arXiv:2511.23199 (2025)

    Tan, Z., Wang, Z., Yang, X., Liu, S., Wang, X.: Vision bridge transformer at scale. arXiv preprint arXiv:2511.23199 (2025)

    arXiv 2025

  40. [40]

    In: Forty-second International Conference on Machine Learning (2025)

    Wang, J., Pu, J., Qi, Z., Guo, J., Ma, Y., Huang, N., Chen, Y., Li, X., Shan, Y.: Taming rectified flow for inversion and editing. In: Forty-second International Conference on Machine Learning (2025)

    2025

  41. [41]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Wang, J., Liu, J., Sun, X., Singh, K.K., Shu, Z., Zhang, H., Yang, J., Zhao, N., Wang, T.Y., Chen, S.S., et al.: Comprehensive relighting: Generalizable and con- sistent monocular human relighting and harmonization. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 380–390 (2025)

    2025

  42. [42]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Wang, Y., Liu, R., Lin, J., Liu, F., Yi, Z., Wang, Y., Ma, R.: Omnistyle: Filtering high quality style transfer data at scale. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 7847–7856 (2025)

    2025

  43. [43]

    arXiv preprint arXiv:2508.02324 (2025)

    Wu, C., Li, J., Zhou, J., Lin, J., Gao, K., Yan, K., Yin, S.m., Bai, S., Xu, X., Chen, Y., et al.: Qwen-image technical report. arXiv preprint arXiv:2508.02324 (2025)

    Pith/arXiv arXiv 2025

  44. [44]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Xing, X., Groh, K., Karaoglu, S., Gevers, T., Bhattad, A.: Luminet: Latent in- trinsics meets diffusion models for indoor scene relighting. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 442–452 (2025)

    2025

  45. [45]

    Advances in Neural Information Processing Systems36, 48723–48743 (2023)

    Yang, Y., Peng, H., Shen, Y., Yang, Y., Hu, H., Qiu, L., Koike, H., et al.: Im- agebrush: Learning visual in-context instructions for exemplar-based image ma- nipulation. Advances in Neural Information Processing Systems36, 48723–48743 (2023)

    2023

  46. [46]

    In: European Conference on Computer Vision

    Yu, Y., Meka, A., Elgharib, M., Seidel, H.P., Theobalt, C., Smith, W.A.: Self- supervised outdoor scene relighting. In: European Conference on Computer Vision. pp. 84–101. Springer (2020)

    2020

  47. [47]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Zhang, B., Xie, H., Gao, Z., Wang, Y.: Choose what you need: Disentangled repre- sentation learning for scene text recognition removal and editing. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 28358–28368 (2024) Consistent Feature Transport for Image Relighting 19

    2024

  48. [48]

    Advances in Neural Information Pro- cessing Systems36, 31428–31449 (2023)

    Zhang, K., Mo, L., Chen, W., Sun, H., Su, Y.: Magicbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Pro- cessing Systems36, 31428–31449 (2023)

    2023

  49. [49]

    In: The Thirteenth International Conference on Learning Representations (2025)

    Zhang, L., Rao, A., Agrawala, M.: Scaling in-the-wild training for diffusion-based illumination harmonization and editing by imposing consistent light transport. In: The Thirteenth International Conference on Learning Representations (2025)

    2025

  50. [50]

    Advances in Neural Information Processing Systems37, 96775–96796 (2024)

    Zhang, X., Gao, W., Jain, S., Maire, M., Forsyth, D., Bhattad, A.: Latent intrin- sics emerge from training to relight. Advances in Neural Information Processing Systems37, 96775–96796 (2024)

    2024