arxiv: 2605.07846 · v2 · submitted 2026-05-08 · 💻 cs.CV

Recognition: no theorem link

BRIDGE: Background Routing and Isolated Discrete Gating for Coarse-Mask Local Editing

Peilin Xiong , Honghui Yuan , Junwen Chen , Keiji Yanai

Authors on Pith no claims yet

Pith reviewed 2026-05-12 03:47 UTC · model grok-4.3

classification 💻 cs.CV

keywords coarse-mask local editingmask-shape biasdiffusion transformerpositional embedding routingbackground preservationlocal image editinggenerative models

0 comments

The pith

BRIDGE uses background routing and a discrete geometric gate to stop coarse masks from dictating the shape of edited objects in image generation.

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

The paper establishes that rough user masks in local image editing often act as unintended shape priors, pulling generated content toward the mask boundary rather than following the edit instruction. It frames this as mask-shape bias under a Two-Zone Constraint that keeps the background stable while allowing flexible changes in the editable region. BRIDGE solves it by generating separate Main and Subject Paths outside the DiT backbone and routing positional embeddings through a learnable Discrete Geometric Gate that lets subject tokens borrow background coordinates near boundaries or retain subject-centric ones for geometric freedom. This produces measurable gains on BRIDGE-Bench and competitive zero-shot results on MagicBrush and ICE-Bench while adding only a small number of parameters. The approach matters because it turns coarse masks into flexible support rather than rigid templates, enabling more natural local edits without precise user masking or copied control networks.

Core claim

BRIDGE keeps masks outside the DiT backbone for support construction and blending, generates content via BridgePath with a Main Path that preserves background context from independent noise and a Subject Path for editable content, and introduces a learnable Discrete Geometric Gate that performs token-level positional-embedding routing so subject tokens can borrow background-anchored coordinates near fusion regions or keep subject-centric coordinates for geometric freedom.

What carries the argument

The Discrete Geometric Gate, which routes positional embeddings at the token level under the Two-Zone Constraint to control whether subject tokens inherit background coordinates or retain independent ones.

If this is right

Local SigLIP2-T rises from 0.262 with FLUX.1-Fill and 0.390 with ACE++ to 0.503 on BRIDGE-Bench, with parallel gains in local DINO and DreamSim.
Zero-shot transfer yields competitive alignment and source preservation on MagicBrush and ICE-Bench.
The routing module adds only 13.31M parameters instead of requiring copied ControlNet-style branches.
Masks remain external to the DiT backbone, avoiding internal mask injection while still supporting blending.

Where Pith is reading between the lines

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

The routing idea could extend to video or 3D editing tasks where coarse user annotations similarly risk imposing unwanted geometry.
If the gate generalizes across backbones, it might reduce reliance on precise mask drawing tools in consumer editing apps.
The separation of Main and Subject Paths suggests a broader pattern for preserving context in other conditional generation settings.

Load-bearing premise

The assumption that mask-shape bias from positional embeddings and attention connectivity is the dominant failure mode and that the discrete geometric gate will reliably prevent shape inheritance without introducing new artifacts or lowering overall image quality.

What would settle it

An ablation study on BRIDGE-Bench in which removing or disabling the Discrete Geometric Gate causes local SigLIP2-T, DINO, and DreamSim scores to fall back to the levels achieved by FLUX.1-Fill or ACE++ baselines.

Figures

Figures reproduced from arXiv: 2605.07846 by Honghui Yuan, Junwen Chen, Keiji Yanai, Peilin Xiong.

**Figure 1.** Figure 1: BRIDGE produces high-quality coarse-mask local edits while preserving seamless background fusion. All examples in this figure are generated by our final BRIDGE model. Across remove, add, and change edits, the generated results remain well integrated with the surrounding scene while allowing the edited subject to depart from the rough mask shape when needed. Together, these examples illustrate the central g… view at source ↗

**Figure 2.** Figure 2: Mask-shape bias in coarse-mask local editing. The left four examples are from the test split of our constructed dataset, and the right four examples are from the public ICE-Bench benchmark. On the left, ACE++ produces a jagged hat boundary that follows the rough mask, FLUX.1-Fill fails to complete the edit, and our method BRIDGE generates a natural round hat. On the right, BRIDGE generates a single cup as … view at source ↗

**Figure 3.** Figure 3: Diagnostic Qwen-Image [21] experiment. Positional embeddings determine which image context a token reuses. Sharing a PE across different visual tokens yields duplicated or highly similar content, while a cross-region attention mask suppresses this aliasing. This motivates BRIDGE’s token-level PE switch for controlling where subject tokens borrow context from. encoder and then provided to the DiT as latent … view at source ↗

**Figure 4.** Figure 4: Automated data pipeline for BRIDGE-Bench. BRIDGE-Bench is constructed to decouple object change from background drift. A VLM discovers local edit concepts, SAM proposes candidate masks, DreamSim filters for sufficient foreground change and limited background drift, and forced compositing restores the untouched scene. Mask perturbation further weakens contour dependence, producing training pairs where masks… view at source ↗

**Figure 5.** Figure 5: BRIDGE overview. BRIDGE separates edit support from subject geometry generation. The Main Path preserves background context, while the Subject Path generates the editable region from independent noise. A Discrete Geometric Gate routes subject-token positional embeddings between background-anchored coordinates for fusion and subject-centric coordinates for geometric freedom. Path is initialized with indepen… view at source ↗

**Figure 6.** Figure 6: Visual ablation of Subject Path generation. Discrete routing is necessary for coherent Subject Path generation. With BRIDGE, the Subject Path forms sharp and instruction-consistent objects; without adaptive PE routing, it often collapses into blurry copies or weak structure. This visual ablation explains the local-precision gains observed in [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗

**Figure 7.** Figure 7: BRIDGE improves prompt-consistent local generation under coarse masks. Left: our constructed test split. Right: ICE-Bench examples. Compared with Q-Control and ACE++, BRIDGE more often follows prompts, generates plausible local geometry, and blends edits cleanly into the scene. fusion. The remaining trade-off is support-level: bbox support gives more shape freedom, whereas stricter mask-based blending can … view at source ↗

**Figure 8.** Figure 8: Grid-search summary for the main setting. We compare 15 combinations of CFG and latent-blending strength α. The highlighted row marks the selected configuration used in the main experiments. • Background Preservation: BRIDGE still uses the user mask outside the transformer to derive the support used for latent blending [1]. Our claim in the main text is specifically comparative: unlike inpainting or in-con… view at source ↗

**Figure 9.** Figure 9: Selection rationale for CFG and blending alpha in the main setting. The left panel plots local DINO within the edit bounding box, and the right panel plots global DreamSim. The selected setting, CFG= 2 and α = 0.1, is highlighted in red because it provides the strongest joint trade-off between local edit fidelity and global perceptual consistency among the tested combinations [PITH_FULL_IMAGE:figures/full… view at source ↗

**Figure 10.** Figure 10: Selected processed examples from our internal dataset. These examples illustrate the instruction-guided local edits retained after semantic filtering, background auditing, and forced background compositing. alpha blending. This process weakens explicit shape boundaries, better simulates messy human user scribbles, and makes the retained background nearly identical at the pixel level (L1 ≈ 0). Curating BRI… view at source ↗

**Figure 11.** Figure 11: Additional results for Remove. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_11.png] view at source ↗

**Figure 12.** Figure 12: Additional results for Add. 21 [PITH_FULL_IMAGE:figures/full_fig_p021_12.png] view at source ↗

**Figure 13.** Figure 13: Additional results for Replace (1/2). renewcommand10.5 Replace the orange tabby cat with a fluffy, white Bichon Frise dog,... Add a large, realistic, vibrant green praying mantis perched on the railing,... Change the ram to a white goat with shaggy fur, standing next... noalign Change the lion to a friendly, plush grey and white wolf with... Change the rustic bread to light-brown, multi-grain crackers, ke… view at source ↗

**Figure 14.** Figure 14: Additional results for Replace (2/2). 22 [PITH_FULL_IMAGE:figures/full_fig_p022_14.png] view at source ↗

**Figure 15.** Figure 15: User-Controlled Fusion Boundary: Bounding Box Support and Mask-based Blending Comparison. (Top) The user provides an irregular freeform mask (red) to add a lake. BRIDGE converts this mask to its bounding box, giving the Subject Path a wider support region that can extend beyond the original scribble. (Bottom) Using mask-based background blending (blend α = 0.5) instead gives the user a stricter fusion bou… view at source ↗

read the original abstract

Coarse-mask local image editing asks a model to modify a user-indicated region while preserving the surrounding scene. In practice, however, rough masks often become unintended shape priors: instead of serving as flexible edit support, the mask can pull the generated object toward its accidental boundary. We study this failure as mask-shape bias and frame the task through a Two-Zone Constraint, where the background should remain stable while the editable region should follow the instruction without being forced to inherit the mask contour. BRIDGE addresses this setting by keeping masks outside the DiT backbone for support construction and blending, avoiding DiT-internal mask injection and copied control branches. It uses BridgePath generation, where a Main Path preserves background context and a Subject Path generates editable content from independent noise. Motivated by a diagnostic Qwen-Image experiment showing that positional embeddings and attention connectivity regulate which image context visual tokens reuse, BRIDGE introduces a learnable Discrete Geometric Gate for token-level positional-embedding routing. This gate lets subject tokens borrow background-anchored coordinates near fusion regions or keep subject-centric coordinates for geometric freedom. We evaluate BRIDGE on BRIDGE-Bench, MagicBrush, and ICE-Bench. On BRIDGE-Bench, BRIDGE improves Local SigLIP2-T from 0.262 with FLUX.1-Fill and 0.390 with ACE++ to 0.503, with parallel gains in local DINO and DreamSim. Zero-shot results on MagicBrush and ICE-Bench further indicate competitive alignment and source preservation beyond the curated benchmark, while the added routing module remains compact at 13.31M parameters compared with ControlNet-style copied branches.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

BRIDGE adds a compact positional routing gate and dual-noise paths to reduce mask-shape bias in DiT local editing, with clear benchmark gains, though the Qwen diagnostic may not map directly to the target backbone.

read the letter

The main takeaway is that BRIDGE keeps masks outside the DiT, runs a background-preserving main path alongside a subject path started from fresh noise, and adds a small learnable Discrete Geometric Gate that routes positional embeddings so subject tokens can either borrow background coordinates near the boundary or keep their own for more shape freedom. This is a targeted change rather than another copied control branch. The paper shows the gate is only 13M parameters and reports a lift on BRIDGE-Bench from 0.39 to 0.50 on Local SigLIP2-T, plus matching gains on local DINO and DreamSim, with competitive zero-shot numbers on MagicBrush and ICE-Bench. That is useful for anyone who has watched rough masks pull generated objects into unwanted contours. The architecture description is concrete and the parameter overhead is modest, which counts as a practical plus. The Two-Zone Constraint framing also gives a clean way to talk about background stability versus edit freedom. The soft spot is the transfer story. The motivating diagnostic is run on Qwen-Image, yet the comparisons are against FLUX.1-Fill, and the method deliberately avoids injecting masks into the DiT itself. Without ablations that isolate the gate from the BridgePath construction, it is hard to know whether the reported gains come from the routing or simply from giving the subject its own noise seed. The abstract also gives no detail on training controls, how the constraint is enforced, or statistical significance of the metric differences. A reader working on DiT editing tools would still find the design worth examining, especially the token-level routing idea. I would bring this to a reading group for the architecture discussion and send it to peer review; the core idea is focused enough that referees can check the missing controls and the Qwen-to-DiT gap.

Referee Report

2 major / 1 minor

Summary. The manuscript proposes BRIDGE to mitigate mask-shape bias in coarse-mask local image editing with DiT backbones. It frames the problem via a Two-Zone Constraint (stable background, instruction-following edit region free of mask contour inheritance), keeps masks outside the backbone, generates content via dual BridgePath (Main Path preserving background context, Subject Path from independent noise), and introduces a compact learnable Discrete Geometric Gate that routes positional embeddings at token level to control geometric freedom near fusion boundaries. Reported results show gains on BRIDGE-Bench (Local SigLIP2-T from 0.262/0.390 to 0.503 vs. FLUX.1-Fill/ACE++), with competitive zero-shot performance on MagicBrush and ICE-Bench and only 13.31M added parameters.

Significance. If the gains are robust and attributable to the routing mechanism rather than the dual-path construction alone, BRIDGE offers a lightweight alternative to copied ControlNet branches for practical local editing, with potential to reduce unintended shape inheritance while preserving source fidelity.

major comments (2)

[Motivation and Diagnostic Experiment] Motivation section (diagnostic Qwen-Image experiment): the Discrete Geometric Gate is motivated by positional-embedding and attention-connectivity effects observed in Qwen-Image, yet the target backbone is DiT (e.g., FLUX.1-Fill) with different attention patterns and noise schedules; no ablation or transfer experiment demonstrates that the same mask-shape bias mechanism dominates in the evaluated models or that the gate, rather than BridgePath itself, drives the reported metric lifts.
[Experiments and Results] Experimental evaluation: the abstract and results report benchmark improvements without detailing controls for the Two-Zone Constraint during training, statistical significance of the gains, or isolation of the gate's contribution via ablations against a BridgePath-only baseline; this leaves open whether the central claim holds or whether gains stem from unablated design choices.

minor comments (1)

[Abstract] Abstract: the phrase 'parallel gains in local DINO and DreamSim' is stated without the corresponding numerical values or exact baseline comparisons, reducing clarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We appreciate the referee's detailed feedback on our manuscript. We address the major comments point by point below and outline the revisions we plan to make.

read point-by-point responses

Referee: [Motivation and Diagnostic Experiment] Motivation section (diagnostic Qwen-Image experiment): the Discrete Geometric Gate is motivated by positional-embedding and attention-connectivity effects observed in Qwen-Image, yet the target backbone is DiT (e.g., FLUX.1-Fill) with different attention patterns and noise schedules; no ablation or transfer experiment demonstrates that the same mask-shape bias mechanism dominates in the evaluated models or that the gate, rather than BridgePath itself, drives the reported metric lifts.

Authors: We thank the referee for highlighting this important point regarding the transferability of our diagnostic insights. The diagnostic experiment on Qwen-Image was intended to illustrate the general role of positional embeddings in regulating context reuse via attention, which is a fundamental property of transformer-based models including DiTs. While attention patterns and noise schedules differ, the core mechanism of mask-shape bias through positional anchoring is expected to persist. Nevertheless, we agree that a direct diagnostic or ablation on the DiT backbone would provide stronger evidence. In the revised manuscript, we will include a transfer experiment or additional analysis on FLUX.1-Fill to confirm the mechanism and better isolate the gate's contribution from the BridgePath construction. revision: partial
Referee: [Experiments and Results] Experimental evaluation: the abstract and results report benchmark improvements without detailing controls for the Two-Zone Constraint during training, statistical significance of the gains, or isolation of the gate's contribution via ablations against a BridgePath-only baseline; this leaves open whether the central claim holds or whether gains stem from unablated design choices.

Authors: We acknowledge the need for greater rigor in our experimental reporting. Regarding controls for the Two-Zone Constraint, we will expand the methods section to detail how the constraint was enforced during training, including any specific loss terms or data sampling strategies. For statistical significance, we will report standard deviations across multiple runs or compute p-values for the observed improvements. Most importantly, to isolate the Discrete Geometric Gate's contribution, we will add an ablation study comparing the full BRIDGE model against a BridgePath-only variant without the gate. These additions will clarify that the reported gains are attributable to the proposed routing mechanism. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical method with independent benchmark results

full rationale

The paper introduces BRIDGE as a new architecture (BridgePath with Main/Subject paths plus learnable Discrete Geometric Gate) motivated by an internal diagnostic experiment on Qwen-Image. All reported outcomes are direct empirical scores on BRIDGE-Bench (Local SigLIP2-T rising from 0.262/0.390 baselines to 0.503), MagicBrush, and ICE-Bench. No equations, fitted parameters renamed as predictions, or self-citation chains appear that would make any claimed improvement equivalent to the method's own inputs by construction. The Two-Zone Constraint framing and gate design are presented as novel contributions evaluated against external baselines (FLUX.1-Fill, ACE++), keeping the derivation chain self-contained.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review provides no explicit free parameters, axioms, or invented entities. The learnable gate implies trainable weights, but no count or fitting procedure is stated. No background mathematical assumptions are listed.

pith-pipeline@v0.9.0 · 5610 in / 1280 out tokens · 56373 ms · 2026-05-12T03:47:42.388152+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Reference graph

Works this paper leans on

43 extracted references · 43 canonical work pages · 6 internal anchors

[1]

Avrahami, O

O. Avrahami, O. Fried, and D. Lischinski. Blended latent diffusion.ACM Transactions on Graphics (TOG), 42(4):1–11, 2023

work page 2023
[2]

S. Bai, Y . Cai, R. Chen, K. Chen, Z. Cheng, L. Deng, W. Ding, C. Gao, C. Ge, W. Ge, Z. Guo, J. Huang, F. Huang, B. Hui, S. Jiang, Z. Li, M. Li, M. Li, X. Chen, Q. Huang, K. Li, and Z. Lin. Qwen3-VL technical report.arXiv preprint arXiv:2511.21631, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025
[3]

FLUX.1 Kontext: Flow Matching for In-Context Image Generation and Editing in Latent Space

S. Batifol, A. Blattmann, F. Boesel, S. Consul, C. Diagne, T. Dockhorn, J. English, Z. English, P. Esser, S. Kulal, K. Lacey, Y . Levi, C. Li, D. Lorenz, J. Müller, D. Podell, R. Rombach, H. Saini, A. Sauer, and L. Smith. FLUX.1 Kontext: Flow Matching for in-context image generation and editing in latent space.arXiv preprint arXiv:2506.15742, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025
[4]

Estimating or Propagating Gradients Through Stochastic Neurons for Conditional Computation

Y . Bengio, N. Léonard, and A. Courville. Estimating or propagating gradients through stochastic neurons for conditional computation.arXiv preprint arXiv:1308.3432, 2013

work page internal anchor Pith review Pith/arXiv arXiv 2013
[5]

Brooks, A

T. Brooks, A. Holynski, and A. A. Efros. InstructPix2Pix: Learning to follow image editing instructions. InCVPR, 2023

work page 2023
[6]

M. Cao, Y . Wang, N. Sebe, and D. de Geus. MasaCtrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. InICCV, 2023

work page 2023
[7]

Carion, L

N. Carion, L. Gustafson, Y .-T. Hu, S. Debnath, R. Hu, D. Suris, C. Ryali, K. V . Alwala, H. Khedr, A. Huang, J. Lei, T. Ma, B. Guo, A. Kalla, M. Marks, J. Greer, M. Wang, P. Sun, R. Rädle, T. Afouras, E. Mavroudi, K. Xu, et al. SAM 3: Segment anything with concepts. InICLR, 2026

work page 2026
[8]

Couairon, T

G. Couairon, T. Lefort, A. Kadkhodamohammadi, O. Clercq, L. Sigal, and J.-F. Lalonde. DiffEdit: Diffusion-based semantic image editing with mask guidance. InICLR, 2023

work page 2023
[9]

Defazio, X

A. Defazio, X. A. Yang, H. Mehta, K. Mishchenko, A. Khaled, and A. Cutkosky. The Road Less Scheduled. InNeurIPS, 2024

work page 2024
[10]

Diffusion Templates: A Unified Plugin Framework for Controllable Diffusion

Zhongjie Duan, Hong Zhang, and Yingda Chen. Diffusion templates: A unified plugin frame- work for controllable diffusion.arXiv preprint arXiv:2604.24351, 2026

work page internal anchor Pith review Pith/arXiv arXiv 2026
[11]

Esser, S

P. Esser, S. Kulal, A. Blattmann, R. Entezari, J. Müller, H. Saini, Y . Levi, D. Lorenz, A. Sauer, F. Boesel, D. Podell, T. Dockhorn, K. Lacey, A. Goodwin, Y . Marek, and R. Rombach. Scaling Rectified Flow Transformers for high-resolution image synthesis. InICML, 2024

work page 2024
[12]

S. Fu, N. Tamir, S. Sundaram, L. Chai, R. Zhang, T. Dekel, and P. Isola. DreamSim: Learning new dimensions of human visual similarity using synthetic data. InNeurIPS, 2023

work page 2023
[13]

A. B. Gokmen, Y . Ekin, B. B. Bilecen, and A. Dundar. RoPECraft: Training-free motion transfer with trajectory-guided RoPE optimization on diffusion transformers. InNeurIPS, 2025

work page 2025
[14]

Z. Han, Z. Jiang, Y . Pan, J. Zhang, C. Mao, C. Xie, Y . Liu, and J. Zhou. ACE: All-round creator and editor following instructions via Diffusion Transformer. InICLR, 2025

work page 2025
[15]

Z. Han, Z. Jiang, Y . Pan, J. Zhang, C. Mao, C. Xie, Y . Liu, and J. Zhou. ACE++: Instruction- based image creation and editing via Context-Aware Content Filling. InProc. of IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2025

work page 2025
[16]

K. He, J. Sun, and X. Tang. Guided image filtering.IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(6):1397–1409, 2013

work page 2013
[17]

Hertz, R

A. Hertz, R. Mokady, J. Tenenbaum, K. Aberman, Y . Pritch, and D. Cohen-Or. Prompt-to- Prompt image editing with Cross-Attention Control. InICLR, 2023

work page 2023
[18]

Ho and T

J. Ho and T. Salimans. Classifier-free diffusion guidance. InNeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021

work page 2021
[19]

J. Ho, A. Jain, and P. Abbeel. Denoising diffusion probabilistic models. InNeurIPS, 2020

work page 2020
[20]

E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Elazar, and D. Chen. LoRA: Low-Rank Adaptation of large language models. InICLR, 2022. 10

work page 2022
[21]

Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion

Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. InECCV, pages 150–168. Springer, 2024

work page 2024
[22]

J. Ke, Q. Wang, Y . Wang, P. Milanfar, and F. Yang. MUSIQ: Multi-scale image quality transformer. InICCV, pages 5148–5157, 2021

work page 2021
[23]

S. Lin, B. Liu, J. Li, and X. Yang. Common diffusion noise schedules and sample steps are flawed. InProc. of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 5404–5413, 2024

work page 2024
[24]

Mishchenko and A

K. Mishchenko and A. Defazio. Prodigy: An expeditiously adaptive parameter-free learner. In ICML, 2024

work page 2024
[25]

Mokady, A

R. Mokady, A. Hertz, K. Aberman, Y . Pritch, and D. Cohen-Or. Null-text inversion for editing real images using guided diffusion models. InCVPR, 2023

work page 2023
[26]

Y . Pan, X. He, C. Mao, Z. Han, Z. Jiang, J. Zhang, and Y . Liu. ICE-Bench: A unified and comprehensive benchmark for image creating and editing. InICCV, pages 16586–16596, 2025

work page 2025
[27]

Peebles and S

W. Peebles and S. Xie. Scalable diffusion models with transformers. InCVPR, 2023

work page 2023
[28]

Y . Qian, L. Song, J. Tong, Y . Yang, J. Lu, W. Hu, and Z. Gan. Pico-Banana-400K: A large-scale dataset for text-guided image editing.arXiv preprint arXiv:2510.19808, 2025

work page arXiv 2025
[29]

Qwen-Image Technical Report

Qwen Team. Qwen-Image technical report.arXiv preprint arXiv:2508.02324, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025
[30]

Radford, J

A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever. Learning transferable visual models from natural language supervision. InICML, pages 8748–8763, 2021

work page 2021
[31]

Rombach, A

R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer. High-resolution image synthesis with latent diffusion models. InCVPR, 2022

work page 2022
[32]

Siméoni, H

O. Siméoni, H. V . V o, M. Seitzer, F. Baldassarre, M. Oquab, C. Jose, V . Khalidov, M. Szafraniec, S. Yi, M. Ramamonjisoa, X. Chen, L. Huang, Y . Liu, Y . Shen, D. Zhao, and H. Zhao. DINOv3. arXiv preprint arXiv:2508.13032, 2025

work page arXiv 2025
[33]

Resolution-robust large mask inpainting with fourier convolutions

Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 3172–3182, 2022

work page 2022
[34]

Ominicontrol: Minimal and universal control for diffusion transformer

Zhenxiong Tan, Songhua Liu, Xingyi Yang, Qiaochu Xue, and Xinchao Wang. Ominicontrol: Minimal and universal control for diffusion transformer. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 14940–14950, 2025

work page 2025
[35]

SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features

M. Tschannen, A. Gritsenko, X. Wang, M. F. Naeem, I. Alabdulmohsin, N. Parthasarathy, T. Evans, L. Beyer, Y . Xia, B. Mustafa, O. Hénaff, J. Harmsen, A. Steiner, and X. Zhai. SigLIP 2: Multilingual vision-language encoders with improved semantic understanding, localization, and dense features.arXiv preprint arXiv:2502.14786, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025
[36]

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 conference on computer vision and pattern recognition, pages 1921–1930, 2023

work page 1921
[37]

Boxdiff: Text-to-image synthesis with training-free box-constrained diffu- sion

Jinheng Xie, Yuexiang Li, Yawen Huang, Haozhe Liu, Wentian Zhang, Yefeng Zheng, and Mike Zheng Shou. Boxdiff: Text-to-image synthesis with training-free box-constrained diffu- sion. InICCV, pages 7452–7461, 2023

work page 2023
[38]

Xiong, J

P. Xiong, J. Chen, H. Yuan, and K. Yanai. PosBridge: Multi-view Positional Embedding transplant for identity-aware image editing. InProc. of the British Machine Vision Conference (BMVC), 2025. 11

work page 2025
[39]

Zhang, L

K. Zhang, L. Mo, W. Chen, H. Sun, and Y . Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. InNeurIPS, 2023

work page 2023
[40]

H. Zhao, X. Ma, L. Chen, S. Si, R. Wu, K. An, P. Yu, M. Zhang, Q. Li, and B. Chang. UltraEdit: Instruction-based fine-grained image editing at scale. InNeurIPS, 2024

work page 2024
[41]

Loco: Training-free layout-to-image synthesis with localized constraints

Peiang Zhao, Han Li, Ruiyang Jin, and S Kevin Zhou. Loco: Training-free layout-to-image synthesis with localized constraints. InProceedings of the 33rd ACM International Conference on Multimedia, pages 9481–9490, 2025

work page 2025
[42]

Lay- outdiffusion: Controllable diffusion model for layout-to-image generation

Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Lay- outdiffusion: Controllable diffusion model for layout-to-image generation. InCVPR, pages 22490–22499, 2023

work page 2023
[43]

Zoomed-in Analysis

Junhao Zhuang, Yanhong Zeng, Wenran Liu, Chun Yuan, and Kai Chen. A task is worth one word: Learning with task prompts for high-quality versatile image inpainting. InECCV, pages 195–211. Springer, 2024. 12 Table 5: Parameter breakdown of the final BRIDGE fine-tuning checkpoint, reported by checkpoint metadata. Component Parameters LoRA adapters 4,725,211,...

work page arXiv 2024