Pith. sign in

REVIEW 3 major objections 5 minor 51 references

Flux-Sculptor: Text-Driven Rich-Attribute Portrait Editing through Decomposed Spatial Flow Control

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Text-driven portrait editing can be both flexible and conservative if the editor splits the denoising flow into a masked structure stage followed by a masked detail stage.

desk verdict Solid preservation results and a genuinely new two-stage control scheme, but the abstract's 'surpasses existing methods in rich-attribute editing' is contradicted by the paper's own numbers. read the letter →

arxiv 2507.03979 v1 pith:6UEZT2GK submitted 2025-07-05 cs.CV

Tianyao He , Runqi Wang , Yang Chen , Dejia Song , Nemo Chen , Xu Tang , Yao Hu This is my paper · ORCID
classification cs.CV
keywords text-drivenportraiteditingrectifiedflowmask-guidedimageattentionvaluefusionlatentPrompt-AlignedSpatialLocatoridentitypreservationfaceattribute
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that the central problem in text-driven portrait editing is a trade-off between following the prompt and leaving the rest of the face untouched. It proposes Flux-Sculptor, built on a rectified-flow image model, which first uses a Prompt-Aligned Spatial Locator to predict which facial region the target prompt refers to, then controls denoising in two stages: an early structuring stage that fuses source latents outside the mask, and a later detailing stage that fuses source attention values outside the mask. On a 2,400-sample CelebA-Edit benchmark, the authors report that this beats six prior editing methods on both editing alignment and preservation metrics, with the largest gains in identity similarity, PSNR, LPIPS and attribute preservation. If the claim holds, text-based portrait editing can be both flexible and conservative: it can change hair, makeup, expression or facial hair without making the person look like someone else.

What carries the argument

The load-bearing mechanism is the mask M produced by the Prompt-Aligned Spatial Locator, a sigmoid pixel-wise similarity between a CNN-extracted image feature and a projected AlphaCLIP text feature: M_i = $\sigma$(sum_c [f^img ∘ f^text]_i,c). The same mask gates both fusion operations of the Structure-to-Detail Edit Control: latent fusion during the first T denoising steps, replacing the masked region of the target latent with the source latent, and attention-value fusion during later steps, replacing the image-valued tokens of the masked region in the last m transformer blocks with the source's recorded values. The mask decides, per pixel or token, whether the source portrait's structure or texture is copied in, so everything hinges on whether it correctly labels the text-relevant region.

What would settle it

Measure PASL's predicted masks against human-annotated masks on a held-out set of open-set portrait prompts such as 'crow's feet at corners of eyes' and 'bright teeth', and report mIoU; if mIoU is low on these cases while editing quality stays high, the claimed role of precise localization is not doing the work, while high mIoU would confirm the central mechanism.

Watch

Extended reading notes

Core claim

Flux-Sculptor establishes that a two-stage, mask-gated control of the rectified-flow denoising process satisfies the two criteria the authors define for ideal portrait editing: precise localization and appropriate content modification. The Prompt-Aligned Spatial Locator computes an editing mask by taking pixel-wise products of a CNN image feature with a projected AlphaCLIP text feature. The Structure-to-Detail Edit Control then runs the first T denoising steps as a facial structuring stage, in which the reconstruction region outside the mask is taken from the source portrait's interpolated latent, and the remaining steps as a detailing stage, in which the image-valued tokens of the last m transformer blocks are fused with the source's recorded attention values in the same masked pattern. The authors claim this decomposition matches how the denoiser behaves—early steps set structure, later steps set detail—and that it outperforms both pure latent fusion and pure attention-value fusion on CLIP score and identity preservation.

Load-bearing premise

The whole system rests on the Prompt-Aligned Spatial Locator producing a mask that correctly marks the text-relevant region even for prompts never seen in training; if the mask is wrong, the edit either misses the target region or leaks into unrelated areas, and the paper reports no quantitative mask-accuracy numbers against human annotations.

Editorial extensions

If this is right

  • Edits such as adding a mustache or turning the eyes blue will change only the relevant region while the rest of the face, hair, background and identity remain nearly identical.
  • Because only the mask and target prompt are needed, the same framework handles many facial attributes without per-attribute retraining, relying on AlphaCLIP's open-set text understanding.
  • Users can supply their own mask instead of the automatically predicted one, enabling personalized edits and multi-attribute edits within a single denoising pass.
  • The two-stage schedule with T = 3 gives the best balance: switching earlier weakens edit strength, while switching later hurts preservation.
  • The reported benchmark results indicate that a flow-based editor can exceed instruction-tuned editing models on preservation metrics while remaining competitive on editing accuracy.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper, quantifying PASL's mask accuracy against human annotations on open-set prompts would separate mask errors from denoising robustness; the paper currently shows only qualitative masks and robustness to noisy masks.
  • Beyond the paper, the structure-versus-detail denoising split likely transfers to other flow- or diffusion-based editors, since low-frequency structure and high-frequency detail are general properties of iterative denoisers.
  • Beyond the paper, the decoupling of gender-biased attributes suggests that spatial mask gating could serve as a lightweight bias-control mechanism for generative editing.
  • Beyond the paper, the multi-attribute pipeline that composes masks from separate sub-prompts could scale to more than two attributes, and testing that would show how mask errors accumulate.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes Flux-Sculptor, a text-driven portrait editing framework built on the FLUX.1-dev rectified flow model. It has two main components: a Prompt-Aligned Spatial Locator (PASL) that predicts an editing-region mask from a text prompt and source image, and a Structure-to-Detail Edit Control (S2D-EC) strategy that decomposes the denoising process into a latent-fusion stage for facial structure and an attention-value-fusion stage for facial details, both guided by the predicted mask. The method is evaluated on a new CelebA-Edit benchmark of 2,400 samples spanning 24 attributes, with quantitative comparisons against StyleCLIP, MasaCtrl, InstructPix2Pix, MagicBrush, SmartEdit, and RF-Edit, plus qualitative comparisons, ablations, and supplementary experiments. The central claim is that Flux-Sculptor surpasses existing methods in both rich-attribute editing and facial information preservation.

Significance. If the claims were fully supported, the paper would make a solid contribution: it introduces a lightweight text-to-mask module specifically designed for fine-grained facial attributes, and it provides an interesting observation-based decomposition of spatial control in rectified-flow editing, with strong reported gains on preservation metrics (PSNR, LPIPS, SSIM, ID, AttrPreserve). The paper is also transparent in reporting ablations, overhead analysis, and robustness to mask noise. However, the headline claim of superiority in rich-attribute editing is not supported by the paper's own numbers, and the evaluation setup has a training/benchmark alignment that weakens the open-set generalizability claim. The preservation half of the contribution is credible, but the editing half needs either a softened claim or substantially stronger evidence.

major comments (3)
  1. [Abstract; §4.2, Table 1; Supplementary B.2] The abstract's claim that Flux-Sculptor 'surpasses existing methods in rich-attribute editing' is contradicted by the paper's own evaluation. In Table 1, Flux-Sculptor's AttrEdit is 0.7144, below InstructPix2Pix (0.7292) and SmartEdit (0.7260), and Supplementary B.2 explicitly states 'our Flux-Sculptor doesn't achieve the best editing performance.' The paper argues that competitors' higher AttrEdit stems from exaggerated edits, but that argument also invalidates AttrEdit as a positive evidence for Flux-Sculptor. The human evaluation (Table 2) covers only three gender-biased attributes and only four methods, not the full 24-attribute benchmark. Therefore the editing component of the central claim is unsupported and should be revised or replaced with evidence that covers the full attribute set.
  2. [§4.1; Supplementary A.1; Supplementary B] The evaluation setup creates a circularity risk: the Text2Mask training prompts are generated from the same CelebA-HQ attribute labels that define the CelebA-Edit benchmark, and the AttrEdit/AttrPreserve classifier is trained on CelebA. This means PASL is tuned to the exact attribute set and annotation scheme used for evaluation, which can inflate its reported scores and limit the support for open-set editing claims. The authors should demonstrate performance on held-out attribute sets (not derived from the training labels) or provide a comprehensive human evaluation across all 24 attributes to establish the claimed rich-attribute editing ability.
  3. [§4.4; Supplementary A.4; Eqs. (13) and (15)] The paper's core spatial-localization claim is not directly measured. S2D-EC's behavior hinges on the mask M: it decides which latent entries and attention values are taken from the source image. Yet §4.4 provides only qualitative mask visualizations, and Supplementary A.4 tests robustness to synthetic mask perturbations qualitatively. The paper should report a quantitative mask accuracy (e.g., IoU or Dice against human annotations) on a reasonable sample, for both in-distribution prompts and open-set prompts, to support the assertion that PASL provides precise localization and to enable assessment of the failure modes when M is inaccurate.
minor comments (5)
  1. [§4.1] The dataset size is reported as 1.25 million in §4.1 but as 1.35 million in Supplementary A.1; these numbers should be reconciled.
  2. [Supplementary B.1, Eq. (18)] The sentence following Eq. (18) says 'AttrEdit also ranges from 0 to 1' but the metric being described is AttrPreserve; the text should say 'AttrPreserve also ranges from 0 to 1.'
  3. [§3.2 and throughout] There are several typographical errors, including 'protrait' (should be 'portrait'), 'faicial' (should be 'facial'), 'evalation' (should be 'evaluation'), 'discreate' (should be 'discrete'), 'Begining' (should be 'Beginning'), and 'stretegy' (should be 'strategy').
  4. [Table 3] The header 'PSAL' in Table 3 should read 'PASL' for consistency with the module name.
  5. [Eq. (11)] In Eq. (11), the notation uses both 'Zti−1' and 'Zti' while the text refers to 'Zti' and 'Zti−1'; the subscript notation should be made consistent throughout §3.4 to avoid confusion between latent indices and timestep indices.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the mask-guided fusion pipeline and benchmark metrics are not derived from the claims being tested; the main issues are benchmark/training overlap and an under-supported headline claim, which are correctness concerns, not circular reasoning.

full rationale

The derivation chain is self-contained. PASL is trained with BCE and Dice losses against ground-truth masks obtained from face parsing and landmark detection (Eqs. 7-9), not from the evaluation labels or from the final editing claim. S2D-EC performs mask-guided latent fusion in Eq. 13 and mask-guided attention-value fusion in Eq. 15, injecting source information only in the reconstruction region; this is a construction of the method, not a prediction that reduces to its input. No fitted parameter is renamed as a predicted metric, no load-bearing result is imported from the authors' own prior work, and no equation is equivalent to its input by definition. The Text2Mask training prompts are built from CelebA-HQ attribute labels (Supplementary A.1) and the CelebA-Edit benchmark selects attributes from CelebA-HQ (Section 4.1), with the automated attribute classifier trained on CelebA (Supplementary B). This is a real train/evaluation alignment concern that can inflate reported scores, but it is a benchmark-overfitting issue rather than circular derivation: the mask predictions and edited images are still produced by a trained network and are not equal to the ground-truth attribute labels by construction. Likewise, the abstract's claim that Flux-Sculptor 'surpasses existing methods in rich-attribute editing' is undermined by the paper's own numbers (AttrEdit 0.7144 vs SmartEdit 0.7260 and InstructPix2Pix 0.7292 in Table 1) and by Supplementary B.2's admission that 'our Flux-Sculptor doesn't achieve the best editing performance,' but a self-contradicted or under-supported claim is an evidence/correctness problem, not a circularity problem. No specific step satisfies the required standard of exhibiting Eq. X = Eq. Y by construction or a fitted parameter renamed as a prediction.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical or architectural entities beyond the named modules (PASL, S2D-EC). The main free parameters are the stage-shifting timestep, the number of fusion blocks (unreported), and evaluation threshold choices. The axioms are standard assumptions for inversion-based editing pipelines, with the temporal decomposition and classifier validity being the most fragile.

free parameters (3)
  • Stage-shifting timestep T = 3 (selected via validation sweep over 1, 3, 5, 7, 9)
    Controls how many initial denoising steps use latent fusion before switching to attention value fusion. Chosen to maximize directional CLIP score on the validation set (Figure 8).
  • Number of attention fusion blocks m = Not reported in the paper
    The detailing stage fuses attention values in the last m Transformer blocks (Section 3.4), but m is never specified, an unstated design choice affecting the strength of detail preservation.
  • Mask classification threshold tau for AttrEdit = 0.1
    Used in Eq (17) to decide whether an attribute was successfully edited; this choice directly affects the reported AttrEdit scores and is not justified.
assumptions (5)
  • domain assumption PASL produces masks that accurately localize the text-relevant facial regions
    The whole spatial control depends on the mask locator's accuracy; the paper provides qualitative visualization but no quantitative mask accuracy against human-annotated masks (Supplementary A.4 uses one example).
  • domain assumption RF-Solver inversion accurately reconstructs the source latent so that the inverted noise is valid for editing
    Inversion-and-editing assumes the ODE integration (Eqs 10-12) yields a latent whose denoising reaches the source image; if inversion is lossy, reconstruction fidelity claims degrade.
  • domain assumption Early denoising steps establish structure; later steps refine details
    This temporal decomposition motivates S2D-EC, citing FreeU (Ref [39]) for general diffusion models; it is not directly verified for the Flux transformer on portraits.
  • domain assumption The attribute classifier trained on CelebA provides a valid measure of editing success
    AttrEdit and AttrPreserve rely on this classifier; the paper itself demonstrates its failure for gender-biased attributes, so its validity is attribute-dependent.
  • domain assumption The pretrained FLUX.1-dev model is capable of performing the required attribute edits without fine-tuning
    The framework freezes Flux weights and relies on prompt guidance plus spatial control; if the base model cannot render a desired attribute, the pipeline cannot add it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Flux-Sculptor: Text-Driven Rich-Attribute Portrait Editing through Decomposed Spatial Flow Control." pith.science (2026). https://pith.science/paper/6UEZT2GK

@misc{pith2026250703979,
  author       = {Pith},
  title        = {Pith review of: Flux-Sculptor: Text-Driven Rich-Attribute Portrait Editing through Decomposed Spatial Flow Control},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6UEZT2GK}},
  note         = {Machine review of arXiv:2507.03979}
}
read the original abstract

Text-driven portrait editing holds significant potential for various applications but also presents considerable challenges. An ideal text-driven portrait editing approach should achieve precise localization and appropriate content modification, yet existing methods struggle to balance reconstruction fidelity and editing flexibility. To address this issue, we propose Flux-Sculptor, a flux-based framework designed for precise text-driven portrait editing. Our framework introduces a Prompt-Aligned Spatial Locator (PASL) to accurately identify relevant editing regions and a Structure-to-Detail Edit Control (S2D-EC) strategy to spatially guide the denoising process through sequential mask-guided fusion of latent representations and attention values. Extensive experiments demonstrate that Flux-Sculptor surpasses existing methods in rich-attribute editing and facial information preservation, making it a strong candidate for practical portrait editing applications. Project page is available at https://flux-sculptor.github.io/.

Figures

Figures reproduced from arXiv: 2507.03979 by the authors.

Figure 1
Figure 1. Text-based portrait editing results of Flux-Sculptor. We use eight text prompts to edit the central portraits. Flux-Sculptor effectively follows the text to achieve diverse attribute modification with visual harmony, naturalness, and identity preservation. Abstract Text-driven portrait editing holds significant potential for various applications but also presents considerable chal￾lenges. An ideal text-driven portra… view at source ↗
Figure 2
Figure 2. The Flux-Sculptor pipeline. Flux-Sculptor first obtains the inverted latent through RF-Solver Inversion. Then, the PASL extracts the target text-related editing regions based on the source image and target prompt. Based on the editing regions, the framework designs the S2D-EC strategy to control facial structuring and detailing through region-guided latent fusion and attention-value fusion, respectively. source imag… view at source ↗
Figure 3
Figure 3. The Prompt-Aligned Spatial Locator. Given X0 and p˜, the PASL extracts visual and language features and generates the editing mask based on cross-modal similarities. modifications to be strictly confined to the text-related re￾gions, while the remaining parts should be reconstructed. Despite various open-set segmentation methods, existing segmenters fail to effectively locate facial attributes due to their complexit… view at source ↗
Figures from the paper (14 more)
Figure 4
Figure 4. Figure 4: Editing results under different mask guidance strate￾gies. We show the latent fusion, attention value fusion, and our S2D-EC’s results in each column. Finally, the mask loss Lmask can be written as: Lmask = LBCE + LDICE. (9) PASL can automatically generate precise faci…
Figure 5
Figure 5. Figure 5: Qualitative comparisons. We conduct portrait editing with eight different text prompts and visualize the results of our Flux￾Sculptor and all competitors (left column: MagicBrush, MasaCtrl, RF-Edit; right column: StyleCLIP, InstructPix2Pix, SmartEdit). RF-Edit Flux-Scu…
Figure 6
Figure 6. Figure 6: Gender-biased attribute editing. We edit male portraits with “wearing lipstick” and “with makeup” on RF-Edit and Flux￾Sculptor, and visualize classification scores S for comparison. cally proves that our method strikes a good balance between reconstruction fidelity and…
Figure 7
Figure 7. Figure 7: Text-driven editing region localization of PASL. Given the source portrait and seven different text prompts, our PASL can provide fine-grained and accurate spatial responses to various user prompts. 0.09 0.08 0.07 0.06 0.05 1 3 5 7 9 Stage-shifting timestep Directional…
Figure 8
Figure 8. Figure 8: Ablation study on stage-shifting timestep T. We re￾port S dir CLIP of the latent fusion, attention value fusion and our S2D￾EC strategies on T = 1, 3, 5, 7, 9. edited images. Furthermore, examples 2, 4, and 7 highlight our model’s capability to accurately edit complex …
Figure 9
Figure 9. Figure 9: Pre-defined and extracted face editing regions. A.1. Text2Mask Dataset Establishment To achieve the challenging fine-grained facial localization, we establish a Text2Mask dataset to train our Prompt￾Aligned Spatial Locator. We use 80,000 high-quality por￾trait images f…
Figure 10
Figure 10. Figure 10: Generated text prompts for each region. gions, and list two for each region in the table in [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]
Figure 11
Figure 11. Figure 11: Editing region localization results according to open-set text prompts. Layer In → Out Dim Stride Activation conv-1 3 → 64 1 LeakyReLU conv-2 64 → 128 2 LeakyReLU conv-3 128 → 256 2 LeakyReLU conv-4 256 → 512 2 LeakyReLU conv-5 512 → 512 1 LeakyReLU conv-6 512 → 512 1…
Figure 12
Figure 12. Figure 12: Edits over masks with different noises. B. Fine-Grained Portrait Editing Metrics In the manuscript’s Section 4.2, we design both coarse￾grained and fine-grained metrics to evaluate the editing and preservation abilities. These coarse-grained metrics have been widely a…
Figure 13
Figure 13. Figure 13: Attribute-wise AttrEdit and AttrPreserve accuracy. We exhibit the fine-grained editing and preservation performances of SmartEdit, InstructPix2Pix, MagicBrush, and our Flux-Sculptor across all 25 facial attributes. editing effects on many challenging attributes where …
Figure 14
Figure 14. Figure 14: Customized editing based on user input mask. C.2. Multi-Attribute Text-Driven Portrait Editing In the previous section, we primarily demonstrated single￾attribute portrait editing. In practical applications, users may want to simultaneously modify multiple facial at￾1…
Figure 15
Figure 15. Figure 15: Multi-attribute text-driven portrait editing. Through a combined mask, Flux-Sculptor can remove bangs, ap￾ply lipstick, and thicken eyebrows within a single denoising pro￾cess. tributes for efficiency. Our Flux-Sculptor can be easily adapted to this task. The multi-at…
Figure 16
Figure 16. Figure 16: More portrait editing visualization results. We show the text-driven portrait editing results of our Flux-sculptor, StyleCLIP, MagicBrush, RF-Edit, and SmartEdit. 16 [PITH_FULL_IMAGE:figures/full_fig_p016_16.png]
Figure 17
Figure 17. Figure 17: More portrait editing visualization results. We show the text-driven portrait editing results of our Flux-sculptor, StyleCLIP, MagicBrush, RF-Edit, and SmartEdit. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_17.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

51 extracted references · 27 canonical work pages

  1. [1]

    Lightroom

    Adobe. Lightroom. lightroom.adobe.com/, . 2025.3.5. 2

    work page 2025

  2. [2]

    Photoshop

    Adobe. Photoshop. www.adobe.com/products/ photo- shop.html, . 2025.3.5. 2

    work page 2025

  3. [3]

    In- structpix2pix: Learning to follow image editing instructions

    Tim Brooks, Aleksander Holynski, and Alexei A Efros. In- structpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 18392–18402, 2023. 2, 6

    work page 2023

  4. [4]

    Lan- guage models are few-shot learners

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Sub- biah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakan- tan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Lan- guage models are few-shot learners. Advances in neural in- formation processing systems, 33:1877–1901, 2020. 2

    1901

  5. [5]

    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. In Proceedings of the IEEE/CVF international con- ference on computer vision, pages 22560–22570, 2023. 3, 5, 6, 8

    work page 2023

  6. [6]

    Association between the use of social media and photograph editing applications, self-esteem, and cosmetic surgery acceptance

    Jonlin Chen, Masaru Ishii, Kristin L Bater, Halley Darrach, David Liao, Pauline P Huynh, Isabel P Reh, Jason C Nel- lis, Anisha R Kumar, and Lisa E Ishii. Association between the use of social media and photograph editing applications, self-esteem, and cosmetic surgery acceptance. JAMA facial plastic surgery, 21(5):361–367, 2019. 1

    work page 2019

  7. [7]

    Latentpaint: Image inpainting in latent space with diffusion models

    Ciprian Corneanu, Raghudeep Gadde, and Aleix M Mar- tinez. Latentpaint: Image inpainting in latent space with diffusion models. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 4334– 4343, 2024. 5, 8

    work page 2024

  8. [8]

    Diffedit: Diffusion-based seman- tic image editing with mask guidance

    Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based seman- tic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 8

    arXiv 2022

Show all 51 references
  1. [9]

    Arcface: Additive angular margin loss for deep face recognition

    Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pages 4690–4699, 2019. 6

    2019

  2. [10]

    High-fidelity and ar- bitrary face editing

    Yue Gao, Fangyun Wei, Jianmin Bao, Shuyang Gu, Dong Chen, Fang Wen, and Zhouhui Lian. High-fidelity and ar- bitrary face editing. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pages 16115–16124, 2021. 1

    2021

  3. [11]

    Mask-guided portrait editing with condi- tional gans

    Shuyang Gu, Jianmin Bao, Hao Yang, Dong Chen, Fang Wen, and Lu Yuan. Mask-guided portrait editing with condi- tional gans. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3436–3445,

  4. [12]

    Adaptively weighted multi-task deep network for person attribute classification

    Keke He, Zhanxiong Wang, Yanwei Fu, Rui Feng, Yu-Gang Jiang, and Xiangyang Xue. Adaptively weighted multi-task deep network for person attribute classification. In Proceed- ings of the 25th ACM international conference on Multime- dia, pages 1636–1644, 2017. 6

    2017

  5. [13]

    Prompt-to-prompt im- age editing with cross attention control

    Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt im- age editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2

    2022 arXiv

  6. [14]

    Denoising dif- fusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models. Advances in neural information processing systems, 33:6840–6851, 2020. 2

    2020

  7. [15]

    Smartedit: Exploring complex instruction-based image editing with multimodal large lan- guage models

    Yuzhou Huang, Liangbin Xie, Xintao Wang, Ziyang Yuan, Xiaodong Cun, Yixiao Ge, Jiantao Zhou, Chao Dong, Rui Huang, Ruimao Zhang, et al. Smartedit: Exploring complex instruction-based image editing with multimodal large lan- guage models. In Proceedings of the IEEE/CVF Conferen...

    2024

  8. [16]

    Scope of va- lidity of psnr in image/video quality assessment

    Quan Huynh-Thu and Mohammed Ghanbari. Scope of va- lidity of psnr in image/video quality assessment. Electronics letters, 44(13):800–801, 2008. 6

    2008

  9. [17]

    Hunting blemishes: Language- guided high-fidelity face retouching transformer with limited paired data

    Le Jiang, Yan Huang, Lianxin Xie, Wen Xue, Cheng Liu, Si Wu, and Hau-San Wong. Hunting blemishes: Language- guided high-fidelity face retouching transformer with limited paired data. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 5102–5111, 2024. 3

    2024

  10. [18]

    Talk-to-edit: Fine-grained facial editing via dialog

    Yuming Jiang, Ziqi Huang, Xingang Pan, Chen Change Loy, and Ziwei Liu. Talk-to-edit: Fine-grained facial editing via dialog. In Proceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 13799–13808, 2021. 2, 3

    2021

  11. [19]

    Toward tiny and high-quality facial makeup with data am- plify learning

    Qiaoqiao Jin, Xuanhong Chen, Meiguang Jin, Ying Chen, Rui Shi, Yucheng Zheng, Yupeng Zhu, and Bingbing Ni. Toward tiny and high-quality facial makeup with data am- plify learning. In European Conference on Computer Vision, pages 340–356. Springer, 2024. 3

    2024

  12. [20]

    Progressive growing of gans for improved quality, stability, and variation

    Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 6

    2017 arXiv

  13. [21]

    A style-based generator architecture for generative adversarial networks

    Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 4401–4410, 2019. 6

    2019

  14. [22]

    Imagic: Text-based real image editing with diffusion models

    Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. InPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6007–6017, 2023. 2

    2023

  15. [23]

    Diffu- sionclip: Text-guided diffusion models for robust image ma- nipulation

    Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffu- sionclip: Text-guided diffusion models for robust image ma- nipulation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2426–2435,

  16. [24]

    Auto-encoding vari- ational bayes, 2013

    Diederik P Kingma, Max Welling, et al. Auto-encoding vari- ational bayes, 2013. 3

    2013

  17. [25]

    Berg, Wan-Yen Lo, Piotr Doll ´ar, and Ross Girshick

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C. Berg, Wan-Yen Lo, Piotr Doll ´ar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 8

    2023 arXiv

  18. [26]

    Black Forest Labs. Flux. https://github.com/ black-forest-labs/flux, 2024. 6

    2024

  19. [27]

    Beautygan: Instance-level facial 9 makeup transfer with deep generative adversarial network

    Tingting Li, Ruihe Qian, Chao Dong, Si Liu, Qiong Yan, Wenwu Zhu, and Liang Lin. Beautygan: Instance-level facial 9 makeup transfer with deep generative adversarial network. In Proceedings of the 26th ACM international conference on Multimedia, pages 645–653, 2018. 3

    2018

  20. [28]

    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. arXiv preprint arXiv:2209.03003, 2022. 2, 3

    2022 arXiv

  21. [29]

    Fixing weight decay regularization in adam

    Ilya Loshchilov, Frank Hutter, et al. Fixing weight decay regularization in adam. arXiv preprint arXiv:1711.05101, 5: 5, 2017. 6

    2017 arXiv

  22. [30]

    Repaint: Inpainting using denoising diffusion probabilistic models

    Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11461–11471, 2022. 5, 8

    2022

  23. [31]

    Dpe: Dis- entanglement of pose and expression for general video por- trait editing

    Youxin Pang, Yong Zhang, Weize Quan, Yanbo Fan, Xi- aodong Cun, Ying Shan, and Dong-ming Yan. Dpe: Dis- entanglement of pose and expression for general video por- trait editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 427–436,

  24. [32]

    Styleclip: Text-driven manipulation of stylegan imagery

    Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF inter- national conference on computer vision , pages 2085–2094,

  25. [33]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF inter- national conference on computer vision , pages 4195–4205,

  26. [34]

    Diffusion autoen- coders: Toward a meaningful and decodable representation

    Konpat Preechakul, Nattanat Chatthee, Suttisak Wizad- wongsa, and Supasorn Suwajanakorn. Diffusion autoen- coders: Toward a meaningful and decodable representation. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 10619–10629, 2022. 3

    2022

  27. [35]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. In International conference on machine learning, ...

    2021

  28. [36]

    Sam 2: Segment anything in images and videos

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman R¨adle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junt- ing Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao- Yuan Wu, Ross Girshick, Piotr Doll´ar, and Christoph Feic...

    2024 arXiv

  29. [37]

    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. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 2

    2022

  30. [38]

    Semantic im- age inversion and editing using rectified stochastic differen- tial equations

    Litu Rout, Yujia Chen, Nataniel Ruiz, Constantine Carama- nis, Sanjay Shakkottai, and Wen-Sheng Chu. Semantic im- age inversion and editing using rectified stochastic differen- tial equations. arXiv preprint arXiv:2410.10792, 2024. 3

    arXiv

  31. [39]

    Freeu: Free lunch in diffusion u-net

    Chenyang Si, Ziqi Huang, Yuming Jiang, and Ziwei Liu. Freeu: Free lunch in diffusion u-net. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4733–4743, 2024. 5

    2024

  32. [40]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2

    2010 arXiv

  33. [41]

    Generalised dice overlap as a deep learning loss function for highly unbalanced segmen- tations

    Carole H Sudre, Wenqi Li, Tom Vercauteren, Sebastien Ourselin, and M Jorge Cardoso. Generalised dice overlap as a deep learning loss function for highly unbalanced segmen- tations. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support...

    2017

  34. [42]

    Alpha- clip: A clip model focusing on wherever you want

    Zeyi Sun, Ye Fang, Tong Wu, Pan Zhang, Yuhang Zang, Shu Kong, Yuanjun Xiong, Dahua Lin, and Jiaqi Wang. Alpha- clip: A clip model focusing on wherever you want. In Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13019–13029, 2024. 4, 6

    2024

  35. [43]

    Maniclip: Multi- attribute face manipulation from text

    Hao Wang, Guosheng Lin, Ana Garc ´ıa del Molino, An- ran Wang, Jiashi Feng, and Zhiqi Shen. Maniclip: Multi- attribute face manipulation from text. International Journal of Computer Vision, 132(10):4616–4632, 2024. 2, 3, 6

    2024

  36. [44]

    Tam- ing rectified flow for inversion and editing

    Jiangshan Wang, Junfu Pu, Zhongang Qi, Jiayi Guo, Yue Ma, Nisha Huang, Yuxin Chen, Xiu Li, and Ying Shan. Tam- ing rectified flow for inversion and editing. arXiv preprint arXiv:2411.04746, 2024. 3, 5, 6, 8

    2024 arXiv

  37. [45]

    Image quality assessment: from error visibility to structural similarity

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Si- moncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600–612, 2004. 6

    2004

  38. [46]

    Chatface: Chat-guided real face edit- ing via diffusion latent space manipulation

    Dongxu Yue, Qin Guo, Munan Ning, Jiaxi Cui, Yuesheng Zhu, and Li Yuan. Chatface: Chat-guided real face edit- ing via diffusion latent space manipulation. arXiv preprint arXiv:2305.14742, 2023. 2, 3

    2023 arXiv

  39. [47]

    Magicbrush: A manually annotated dataset for instruction- guided image editing

    Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction- guided image editing. Advances in Neural Information Pro- cessing Systems, 36:31428–31449, 2023. 2, 6

    2023

  40. [48]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shecht- man, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 586–595, 2018. 6

    2018

  41. [49]

    A person with bushy eyebrows

    Yuxuan Zhang, Qing Zhang, Yiren Song, Jichao Zhang, Hao Tang, and Jiaming Liu. Stable-hair: Real-world hair transfer via diffusion model. arXiv preprint arXiv:2407.14078, 2024. 3 10 Flux-Sculptor: Text-Driven Rich-Attribute Portrait Editing through Decomposed Spatial Flow Cont...

    2024 arXiv

  42. [50]

    forehead wrinkles

    …straight eyebrows; 2) …sparse eyebrows; 3) …brown eyebrows…Text promptsRegionsA person with a goatee beard.A person with a mustache…BeardA person with slim lower face.A person with a chubby face…face shapeA person with double chins.A person with a square chin…ChinA person wit...

  43. [51]

    a woman

    precise mask2. PASL output3. strong noise4. very coarse “a woman.” “a woman with pink eyeshadows.” Figure 12. Edits over masks with different noises. B. Fine-Grained Portrait Editing Metrics In the manuscript’s Section 4.2, we design both coarse- grained and fine-grained metr...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.