Pith. sign in

REVIEW 3 major objections 6 minor 47 references

Controllable image generation works better as a closed feedback loop than as open-loop feed-forward control.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-11 15:04 UTC pith:SVLIL4EO

load-bearing objection Clean training-free closed-loop wrapper that actually beats compute-matched best-of-N on ID/pose/depth; the modified PID residual is the real idea, soft spots are the usual manual gains and linear cost. the 3 major comments →

arxiv 2607.04691 v1 pith:SVLIL4EO submitted 2026-07-06 cs.CV

From Open Loop to Closed Loop: A Test-Time Iterative Optimization Framework for Reference-Consistent Image Generation

Baixuan Zhao , Xinyu Zhang , Huayu Zheng , Shuaicheng Liu , Xiongkuo Min , Guangtao Zhai , Xiaohong Liu This is my paper
classification cs.CV
keywords closed-loop generationPID controlreference-consistent generationtest-time optimizationdiffusion modelsidentity preservationpose controldepth control
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.

Current methods that condition diffusion models on a face, pose, or depth map simply inject that condition once and hope the output stays faithful. The authors argue this is an open-loop system that cannot correct its own drift. They reframe the problem as classical tracking control: a frozen generative model is the plant, a pre-trained sensor measures how far the current image is from the reference, and a modified PID controller updates the latent condition each iteration. Because the loop runs only at test time, no weights are changed and the same wrapper works for identity, pose, and depth tasks. Under equal compute budgets the closed loop raises facial similarity by as much as 25 percent and cuts pose and depth error by roughly 28 percent relative to best-of-N open-loop sampling. The practical claim is that reference consistency can be improved by treating generation itself as a feedback-control problem rather than by training larger adapters.

Core claim

Reference-consistent image generation can be cast as a discrete-time closed-loop tracking problem: a sensor measures the discrepancy between the generated image and a target reference state, a modified PID law converts that error into an updated latent control signal, and the frozen generative model is driven iteratively toward asymptotic alignment without any retraining.

What carries the argument

The modified PID update (Eq. 3) that keeps the target state as a persistent base signal and adds proportional, integral and derivative residuals: uk+1 = u* + Kp ek + Ki sum en + Kd (ek - e k-1). This residual compensator supplies the continuous bias a memoryless generative plant needs while still correcting observed drift.

Load-bearing premise

The pre-trained sensor must produce an error vector that, when residual-added to the control input, the frozen generative model can actually follow toward the target; if the sensor is noisy or the plant response is highly non-linear the loop can stall or amplify artifacts.

What would settle it

On a held-out set of identity, pose and depth references, run the same number of full diffusion passes for both the closed-loop PID wrapper and a best-of-N open-loop baseline; if facial similarity, MPJPE and depth MAE of the closed-loop outputs do not exceed those of the best open-loop samples, the central claim fails.

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

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes a training-free, model-agnostic test-time framework that reformulates reference-consistent controllable generation as a closed-loop tracking problem. A pre-trained encoder/sensor measures the discrepancy between a reference target state and the current generated image; a modified PID controller then updates the latent control input as a residual around the target state (Eq. 3), treating the frozen diffusion pipeline as a black-box plant. The method is instantiated for ID-preserving, pose-controlled, and depth-controlled generation and is evaluated against original and best-of-N open-loop baselines on multiple backbones. Reported relative gains reach up to 25.36% facial similarity and up to 27.71%/28.50% reductions in pose MPJPE and depth MAE under iteration-matched budgets, with ablations on PID terms, prompts, seeds, and coefficient sensitivity.

Significance. If the empirical gains hold under the stated matching protocol, the work is a useful systems contribution: a lightweight plug-and-play wrapper that improves reference fidelity without retraining and that generalizes across identity and spatial control adapters. Strengths include computation-matched best-of-N baselines, deliberate decoupling of feedback sensors from evaluation models, multi-backbone evaluation, PID-component and coefficient-sensitivity analyses, multi-seed stability statistics, and public code. The control-theoretic framing is primarily conceptual rather than a deep stability analysis, but it productively organizes an external generation-and-check loop that existing open-loop adapters lack. The practical impact is clearest for applications that prioritize strict reference consistency and can tolerate multi-pass latency.

major comments (3)
  1. [§5.3, Tables 2–3, §5.6] §5.3 and Tables 2–3 frame best-of-20 / best-of-15 open-loop sampling as computation-matched to closed-loop iteration. These protocols match the number of full denoising runs but not wall-clock cost under realistic deployment: independent samples can be batched/parallelized, whereas the closed loop is strictly sequential and also incurs sensor cost each step (§5.6). The paper should either report wall-clock and FLOPs under a fixed latency budget, or qualify the claim as “iteration-matched” rather than fully computation-matched, and discuss how gains change when best-of-N is allowed the same wall-clock time.
  2. [§4.2 Eq. (3), Table 4, §2] The central algorithmic claim is that a modified PID residual controller is an effective external feedback mechanism (Eq. 3, §4.2). Table 4 shows that P alone already recovers much of the gain and that I/D help, but the paper does not compare against other test-time uses of the same sensor error—e.g., pure residual iteration without integral/derivative memory, simple fixed-step residual ascent, or energy-guided methods such as FreeDoM/UGD/DAS that could optimize the same tracking objective. Without at least one such control, it remains unclear how much of the reported improvement is due specifically to the PID structure versus any iterative residual correction driven by F.
  3. [§4.3, Limitations, Tables 2–3] Limitations correctly flag dependence on sensor accuracy, but the main results do not quantify failure modes when the sensor is biased or inconsistent with the plant’s controllable degrees of freedom. A short stress test—e.g., injecting controlled noise into F, swapping a weaker feedback encoder, or reporting cases where MPJPE/MAE plateaus or worsens—would make the weakest assumption (that residual updates in the sensor space are plant-followable) falsifiable rather than only acknowledged. This is load-bearing for the universality claim across ID/pose/depth.
minor comments (6)
  1. [Fig. 1, Eq. (3)] Figure 1 and Eq. (3) use u_k for both the latent control input and, in the caption, “current state”; keep a single consistent terminology (control input vs. measured state F(I_k)).
  2. [Table 2, §5.1 Metrics] In Table 2, Structure Diversity sometimes drops slightly under +CL. (e.g., PuLID on CelebA300). A one-sentence discussion of whether closed-loop optimization trades pose/structure diversity for identity would help readers interpret the metric.
  3. [§2] Related work cites ControlNet-XS as feedback-control interpretation of adapters; a clearer sentence distinguishing external task-level PID loops from internal adapter dynamics would sharpen novelty.
  4. [§4.2, Algorithm S1] Supplementary Algorithm S1 is helpful; consider moving a condensed version into the main paper near §4.2 for reproducibility without the supplement.
  5. [§4.1, Table 2] Minor polish: “AsshowninEquation(2)” and similar spacing/typos appear in §4.1; also normalize “PhotoMaker v2” vs “PMv2” on first use in tables.
  6. [Fig. 6, Fig. S1] Fig. 6 is referenced as optimization dynamics in the main text while the supplement also uses Fig. S1 for trajectories; ensure numbering and cross-references are consistent in the camera-ready version.

Circularity Check

0 steps flagged

No significant circularity: empirical wrapper whose gains are measured by held-out external sensors, not forced by construction or self-citation.

full rationale

The paper's load-bearing claims are empirical performance gains (Tables 2–3) of a training-free PID residual wrapper around frozen diffusion plants. The control law (Eq. 3) simply adds a hand-tuned residual to the reference embedding u*; it does not algebraically equal the evaluation metrics (cosine similarity of a different face encoder, MPJPE of DWPose, MAE of ZoeDepth). Feedback and evaluation deliberately use distinct models to avoid sensor-evaluator coupling. PID gains are manually chosen per task (Table 1) and shown robust over grids (Fig. 7), not fitted to the reported numbers. No uniqueness theorem, self-citation chain, or ansatz is invoked to force the result; classical PID is adapted for memoryless plants and then measured against computation-matched best-of-N open-loop baselines. The derivation is therefore self-contained against external benchmarks and exhibits none of the six circularity patterns.

Axiom & Free-Parameter Ledger

2 free parameters · 3 axioms · 1 invented entities

The central empirical claim rests on a small set of hand-chosen controller gains, the modeling assumption that the generative plant is memoryless, and the availability of accurate external sensors. No new physical entities are postulated; the ‘modified PID’ is an algorithmic adaptation of a classical controller.

free parameters (2)
  • Kp, Ki, Kd (per-task PID gains) = task-specific triples listed in Table 1
    Manually selected for each task (Table 1: ID 0.30/0.05/0.01, Pose 0.05/0.02/0.01, Depth 0.01/0.005/0.005). Performance surfaces in Fig. 7 show a usable range but the reported numbers are still free parameters chosen by the authors.
  • iteration budget N = 20 / 15
    Fixed at 20 for ID and 15 for pose/depth; directly multiplies compute and is chosen by hand to reach steady state.
axioms (3)
  • domain assumption Image generators behave as static (memoryless) plants: output depends only on the current control input uk and prompt c.
    Stated in §4.2 to justify replacing classical PID with a residual form that always adds the target state u*.
  • domain assumption A pre-trained encoder F can serve simultaneously as target extractor and as online sensor whose residual is a useful control error.
    Core of the architecture (Eq. 2–3); validity depends on sensor accuracy, noted as a limitation.
  • ad hoc to paper Standard discrete-time PID update (with unit sampling period) remains stable when applied to high-dimensional latent residuals of diffusion models.
    No Lyapunov or frequency-domain stability proof is given; stability is observed empirically within the chosen gain ranges.
invented entities (1)
  • Modified PID residual controller for generative plants no independent evidence
    purpose: Keeps a persistent base control signal u* so that the subject does not disappear when tracking error approaches zero.
    Classical positional PID would drive the control input to zero; the authors add u* as a bias term (Eq. 3). This is an algorithmic invention local to the paper.

pith-pipeline@v1.1.0-grok45 · 21185 in / 2785 out tokens · 26196 ms · 2026-07-11T15:04:16.070161+00:00 · methodology

0 comments
read the original abstract

While controllable image generation has made significant strides by incorporating visual reference conditions, existing methods predominantly operate as open-loop systems. They inject control signals in a strictly feed-forward manner, failing to guarantee strict fidelity to the reference due to the absence of active feedback and error correction mechanisms. To address this fundamental limitation, we propose a novel test-time iterative optimization framework that reformulates reference-consistent generation as a closed-loop dynamic tracking problem. By treating the pre-trained generative model as a control plant, our framework employs a sensor-controller architecture driven by a modified Proportional-Integral-Derivative (PID) algorithm. This mechanism iteratively optimizes the latent control signals at test time based on the sensed discrepancy between the generated output and the reference target. Notably, this approach is entirely training-free, model-agnostic, and integrates seamlessly around existing diffusion pipelines. Extensive evaluations across ID-preserving, pose-controlled, and depth-controlled generation tasks validate the universality of our method. Empirical results demonstrate improvements over computation-matched open-loop baselines, achieving relative performance gains of up to 25.36\% for facial similarity, alongside spatial error reductions of up to 27.71\% for pose alignment and 28.50\% for depth consistency. More broadly, this work offers a new conceptual perspective: it demonstrates that controllable generation can be effectively managed as a dynamic feedback system, bringing the rigorous principles of classical control theory into the optimization of generative models. Code is available at https://github.com/zzdrill/From-Open-Loop-to-Closed-Loop.

Figures

Figures reproduced from arXiv: 2607.04691 by Baixuan Zhao, Guangtao Zhai, Huayu Zheng, Shuaicheng Liu, Xiaohong Liu, Xinyu Zhang, Xiongkuo Min.

Figure 1
Figure 1. Figure 1: Overview of the proposed closed-loop optimization framework. We reformu￾late reference-consistent generation as feedback control over a generative model. The encoder and sensor are the same pre-trained module F, reused in different roles for target extraction and feedback measurement. A modified PID controller iteratively updates the latent control input uk to reduce the generation error. This closed-loop … view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative comparisons of ID-preserving portrait generation. We integrate our closed-loop optimization (+CL.) into four representative baselines. The results demonstrate noticeable improvements in both identity fidelity and fine-grained detail preservation compared to the standard open-loop generations. occasionally smoothed out by standard open-loop methods. Overall, by estab￾lishing an effective closed-… view at source ↗
Figure 3
Figure 3. Figure 3: Recurrent generation experiment. The output of the previous round serves as the reference for the next. Our closed-loop optimization effectively mitigates identity degradation across iterations. Facial similarity (↑) are provided below each image. 5.3 Quantitative Evaluation Optimization Dynamics. Figure S1 illustrates the convergence behavior of the proposed closed-loop system. Across all three tasks (ide… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparisons for pose-controlled generation. Compared to the stan￾dard open-loop baseline (w/o), our closed-loop optimization (with) drastically rectifies structural misalignments and complex articulations (highlighted in boxes). Extracted skeletal maps and the corresponding Mean Per Joint Position Error (MPJPE, ↓) are provided below each result to quantify the geometric improvements. by 28.50%.… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparisons for depth-controlled generation. The open-loop base￾line (w/o) frequently struggles with ambiguous geometric boundaries. In contrast, our closed-loop system (with) ensures precise depth alignment and sharp foreground￾background separation (highlighted by arrows/boxes). The Mean Absolute Error (MAE, ↓) evaluated on the extracted depth maps is reported below. identity and depth. These… view at source ↗
Figure 6
Figure 6. Figure 6: Optimization dynamics across three generation tasks: (a) ID-preserving, (b) pose-controlled, and (c) depth-controlled generation. The plots illustrate the progres￾sion of Facial Similarity (↑), MPJPE (↓), and MAE (↓) over iterations, demonstrating stable and monotonic convergence across diverse baselines [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Coefficient sensitivity analysis. With Kd fixed, we perform a 2D grid search over Kp and Ki. The closed-loop controller consistently improves ID similarity and reduces pose/depth errors over a broad coefficient range. 6 Conclusion In this work, we propose a test-time iterative optimization framework that brings automatic control principles into controllable image generation. By introducing a closed-loop fe… 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

47 extracted references · 15 linked inside Pith

  1. [1]

    Princeton university press (2010)

    Åström, K.J., Murray, R.M.: Feedback systems: An introduction for scientists and engineers. Princeton university press (2010)

    2010

  2. [2]

    In: International Conference on Learning Representations

    Bansal,A.,Chu,H.M.,Schwarzschild,A.,Sengupta,R.,Goldblum,M.,Geiping,J., Goldstein, T.: Universal guidance for diffusion models. In: International Conference on Learning Representations. vol. 2024, pp. 51304–51323 (2024)

    2024

  3. [3]

    arXiv preprint arXiv:2302.12288 (2023)

    Bhat, S.F., Birkl, R., Wofk, D., Wonka, P., Müller, M.: ZoeDepth: Zero-shot trans- fer by combining relative and metric depth. arXiv preprint arXiv:2302.12288 (2023)

    Pith/arXiv arXiv 2023

  4. [4]

    com / black - forest - labs / flux (2024), accessed on March 5, 2026

    Black Forest Labs: FLUX.https : / / github . com / black - forest - labs / flux (2024), accessed on March 5, 2026

    2024

  5. [5]

    International Journal of Dynamics and Control9(2), 818–827 (2021)

    Borase, R.P., Maghade, D., Sondkar, S., Pawar, S.: A review of PID control, tuning methods and applications. International Journal of Dynamics and Control9(2), 818–827 (2021)

    2021

  6. [6]

    Springer science & business media (2013)

    Camacho, E.F., Alba, C.B.: Model predictive control. Springer science & business media (2013)

    2013

  7. [7]

    IEEE Transactions on Pattern Analysis and Machine Intelligence43(1), 172–186 (2019)

    Cao, Z., Hidalgo, G., Simon, T., Wei, S.E., Sheikh, Y.: OpenPose: Realtime multi- person 2D pose estimation using part affinity fields. IEEE Transactions on Pattern Analysis and Machine Intelligence43(1), 172–186 (2019)

    2019

  8. [8]

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

    Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9650–9660 (2021)

    2021

  9. [9]

    arXiv e-prints pp

    Choi, Y., Kwak, S., Lee, K., Choi, H., Shin, J.: Improving diffusion models for virtual try-on. arXiv e-prints pp. arXiv–2403 (2024)

    2024

  10. [10]

    arXiv preprint arXiv:2411.19246 (2024)

    Cui, X., Wu, G., Gan, Z., Zhai, G., Liu, X.: Face2QR: A unified framework for aesthetic, face-preserving, and scannable QR code generation. arXiv preprint arXiv:2411.19246 (2024)

    Pith/arXiv arXiv 2024

  11. [11]

    In: European Conference on Computer Vision

    Dam, P., Jeong, J., Tran, A., Kim, D.: Time-efficient and identity-consistent virtual try-on using a variant of altered diffusion models. In: European Conference on Computer Vision. pp. 35–51. Springer (2024)

    2024

  12. [12]

    In: Advances in Neural Information Processing Systems (2024)

    Guo, Z., Wu, Y., Chen, Z., Chen, L., Zhang, P., He, Q.: PuLID: Pure and lightning ID customization via contrastive alignment. In: Advances in Neural Information Processing Systems (2024)

    2024

  13. [13]

    ACM Transactions on Graphics (TOG)42(6), 1– 14 (2023)

    Jiang, H., Luo, A., Fan, H., Han, S., Liu, S.: Low-light image enhancement with wavelet-based diffusion models. ACM Transactions on Graphics (TOG)42(6), 1– 14 (2023)

    2023

  14. [14]

    arXiv preprint arXiv:1710.10196 (2017)

    Karras, T., Aila, T., Laine, S., Lehtinen, J.: Progressive growing of GANs for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196 (2017)

    Pith/arXiv arXiv 2017

  15. [15]

    arXiv preprint arXiv:2501.05803 (2025)

    Kim,S.,Kim,M.,Park,D.:Test-timealignmentofdiffusionmodelswithoutreward over-optimization. arXiv preprint arXiv:2501.05803 (2025)

    Pith/arXiv arXiv 2025

  16. [16]

    In: Proceedings of the 32nd ACM International Conference on Multimedia

    Li, C., Wu, H., Hao, H., Zhang, Z., Kou, T., Chen, C., Bai, L., Liu, X., Lin, W., Zhai, G.: G-Refine: A general quality refiner for text-to-image generation. In: Proceedings of the 32nd ACM International Conference on Multimedia. pp. 7375– 7384 (2024)

    2024

  17. [17]

    arXiv preprint arXiv:2508.04228 (2026)

    Li, G., Cen, K., Zhao, B., Xin, Y., Luo, S., Zhai, G., Zhang, L., Liu, X.: LayerT2V: A unified multi-layer video generation framework. arXiv preprint arXiv:2508.04228 (2026)

    arXiv 2026

  18. [18]

    In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition

    Li, G., Yang, Y., Song, C., Liu, X., Zhang, C.: FlowDirector: Training-free flow steering for precise text-to-video editing. In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition. pp. 7805–7815 (2026) From Open Loop to Closed Loop 17

    2026

  19. [19]

    In: European Conference on Computer Vision

    Li, M., Yang, T., Kuang, H., Wu, J., Wang, Z., Xiao, X., Chen, C.: ControlNet++: Improving conditional controls with efficient consistency feedback. In: European Conference on Computer Vision. pp. 129–147. Springer (2024)

    2024

  20. [20]

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

    Li, Z., Cao, M., Wang, X., Qi, Z., Cheng, M.M., Shan, Y.: PhotoMaker: Cus- tomizing realistic human photos via stacked ID embedding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8640– 8650 (2024)

    2024

  21. [21]

    arXiv preprint arXiv:2408.02657 (2024)

    Liu, D., Zhao, S., Zhuo, L., Lin, W., Xin, Y., Li, X., Qin, Q., Qiao, Y., Li, H., Gao, P.: Lumina-mGPT: Illuminate flexible photorealistic text-to-image generation with multimodal generative pretraining. arXiv preprint arXiv:2408.02657 (2024)

    Pith/arXiv arXiv 2024

  22. [22]

    arXiv preprint arXiv:2408.06070 (2024)

    Peng, B., Wang, J., Zhang, Y., Li, W., Yang, M.C., Jia, J.: ControlNext: Powerful and efficient control for image and video generation. arXiv preprint arXiv:2408.06070 (2024)

    Pith/arXiv arXiv 2024

  23. [23]

    arXiv preprint arXiv:2307.01952 (2023)

    Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: SDXL: Improving latent diffusion models for high-resolution im- age synthesis. arXiv preprint arXiv:2307.01952 (2023)

    Pith/arXiv arXiv 2023

  24. [24]

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

    Qin, Q., Zhuo, L., Xin, Y., Du, R., Li, Z., Fu, B., Lu, Y., Li, X., Liu, D., Zhu, X., et al.: Lumina-Image 2.0: A unified and efficient image generative framework. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 20031–20042 (2025)

    2025

  25. [25]

    In: International Conference on Machine Learning

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748–8763. PMLR (2021)

    2021

  26. [26]

    arXiv preprint arXiv:2204.061251(2), 3 (2022)

    Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: DALL-E 2: Hier- archical text-conditional image generation with CLIP latents. arXiv preprint arXiv:2204.061251(2), 3 (2022)

    Pith/arXiv arXiv 2022

  27. [27]

    IEEE Transactions on Pattern Analysis and Machine Intelligence44(3), 1623–1637 (2020)

    Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence44(3), 1623–1637 (2020)

    2020

  28. [28]

    arXiv preprint arXiv:2211.01364 (2023)

    Richter, L., Berner, J., Bötte, L.: An optimal control perspective on diffusion-based generative modeling. arXiv preprint arXiv:2211.01364 (2023)

    Pith/arXiv arXiv 2023

  29. [29]

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

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10684–10695 (2022)

    2022

  30. [30]

    In: Interna- tional Conference on Learning Representations (ICLR) (2021)

    Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score- based generative modeling through stochastic differential equations. In: Interna- tional Conference on Learning Representations (ICLR) (2021)

    2021

  31. [31]

    In: Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision

    Tan, Z., Liu, S., Yang, X., Xue, Q., Wang, X.: OMNI-control: Minimal and uni- versal control for diffusion transformer. In: Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision. pp. 14940–14950 (2025)

    2025

  32. [32]

    arXiv preprint arXiv:2401.07519 (2024)

    Wang, Q., Bai, X., Wang, H., Qin, Z., Chen, A., Li, H., Tang, X., Hu, Y.: InstantID: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519 (2024)

    Pith/arXiv arXiv 2024

  33. [33]

    arXiv preprint arXiv:2503.19262 (2025)

    Wang, R., Zheng, Y., Zhang, Z., Li, C., Liu, S., Zhai, G., Liu, X.: Learning hazing to dehazing: Towards realistic haze generation for real-world image dehazing. arXiv preprint arXiv:2503.19262 (2025)

    Pith/arXiv arXiv 2025

  34. [34]

    In: Pro- 18 B

    Wu, G., Liu, X., Jia, J., Cui, X., Zhai, G.: Text2QR: Harmonizing aesthetic cus- tomization and scanning robustness for text-guided QR code generation. In: Pro- 18 B. Zhao et al. ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion. pp. 8456–8465 (2024)

    2024

  35. [35]

    Advances in Neural Information Processing Systems38, 3472–3490 (2026)

    Wu, G., Zheng, H., Luo, S., Zhai, G., Liu, X.: AnimateQR: Bridging aesthetics and functionality in dynamic QR code generation. Advances in Neural Information Processing Systems38, 3472–3490 (2026)

    2026

  36. [36]

    arXiv preprint arXiv:2312.17090 (2023), equal Contribution by Wu, Haoning and Zhang, Zicheng

    Wu, H., Zhang, Z., Zhang, W., Chen, C., Li, C., Liao, L., Wang, A., Zhang, E., Sun, W., Yan, Q., Min, X., Zhai, G., Lin, W.: Q-Align: Teaching LMMs for visual scoring via discrete text-defined levels. arXiv preprint arXiv:2312.17090 (2023), equal Contribution by Wu, Haoning and Zhang, Zicheng. Project Lead by Wu, Haoning. Corresponding Authors: Zhai, Guan...

    Pith/arXiv arXiv 2023

  37. [37]

    arXiv preprint arXiv:2510.06308 (2025)

    Xin, Y., Qin, Q., Luo, S., Zhu, K., Yan, J., Tai, Y., Lei, J., Cao, Y., Wang, K., Wang, Y., et al.: Lumina-DiMOO: An omni diffusion large language model for multi-modal generation and understanding. arXiv preprint arXiv:2510.06308 (2025)

    arXiv 2025

  38. [38]

    arXiv preprint arXiv:2406.08713 (2024)

    Yang, X., Wang, Z., Hu, A.: Batch-instructed gradient for prompt evolution: Sys- tematic prompt optimization for enhanced text-to-image synthesis. arXiv preprint arXiv:2406.08713 (2024)

    Pith/arXiv arXiv 2024

  39. [39]

    arXiv preprint arXiv:2407.10459 (2024)

    Yang, Y., Liu, Z., Jia, J., Gao, Z., Li, Y., Sun, W., Liu, X., Zhai, G.: DiffStega: To- wards universal training-free coverless image steganography with diffusion models. arXiv preprint arXiv:2407.10459 (2024)

    Pith/arXiv arXiv 2024

  40. [40]

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

    Yang, Z., Zeng, A., Yuan, C., Li, Y.: DWPose: Effective whole-body pose estima- tion with two-stage distillation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4210–4220 (2023)

    2023

  41. [41]

    arXiv preprint arXiv:2308.06721 (2023)

    Ye, H., Zhang, J., Liu, S., Han, X., Yang, W.: IP-adapter: Text-compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721 (2023)

    Pith/arXiv arXiv 2023

  42. [42]

    Yu, J., Wang, Y., Zhao, C., Ghanem, B., Zhang, J.: FreeDoM: Training-free energy- guidedconditionaldiffusionmodel.In:ProceedingsoftheIEEE/CVFInternational Conference on Computer Vision. pp. 23174–23184 (2023)

    2023

  43. [43]

    In: European Confer- ence on Computer Vision

    Zavadski, D., Feiden, J.F., Rother, C.: ControlNet-XS: Rethinking the control of text-to-image diffusion models as feedback-control systems. In: European Confer- ence on Computer Vision. pp. 343–362. Springer (2024)

    2024

  44. [44]

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

    Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3836–3847 (2023)

    2023

  45. [45]

    Machine Intelligence Research22(5), 817–848 (2025)

    Zhang, X., Wei, X., Hu, W., Wu, J., Wu, J., Zhang, W., Zhang, Z., Lei, Z., Li, Q.: A survey on personalized content synthesis with diffusion models. Machine Intelligence Research22(5), 817–848 (2025)

    2025

  46. [46]

    Advances in Neu- ral Information Processing Systems36, 11127–11150 (2023)

    Zhao, S., Chen, D., Chen, Y.C., Bao, J., Hao, S., Yuan, L., Wong, K.Y.K.: Uni- ControlNet: All-in-one control to text-to-image diffusion models. Advances in Neu- ral Information Processing Systems36, 11127–11150 (2023)

    2023

  47. [47]

    Zheng, H., Li, G., Zhao, B., Luo, S., Jiang, H., Zhai, G., Liu, X.:A2-Edit: Precise reference-guided image editing of arbitrary objects and ambiguous masks. arXiv preprint arXiv:2603.10685 (2026) Supplementary Material From Open Loop to Closed Loop: A Test-Time Iterative Optimization Framework for Reference-Consistent Image Generation Baixuan Zhao1 , Xiny...

    arXiv 2026