Pith. sign in

REVIEW 4 major objections 4 minor 44 references

A single learned scaling coefficient per block, found by black-box reward optimization, improves DiT image generation quality and cuts the sampling steps needed.

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

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

T0 review · deepseek-v4-flash

2026-08-02 17:25 UTC pith:YTC7H27B

load-bearing objection Clever lightweight calibration idea, but the headline gains are confounded with reduced inference steps and the artifacts are missing — needs controlled comparisons before it convinces. the 4 major comments →

arxiv 2603.24800 v2 pith:YTC7H27B submitted 2026-03-25 cs.CV

Calibri: Enhancing Diffusion Transformers via Parameter-Efficient Calibration

Danil Tokhchukov , Aysel Mirzoeva , Andrey Kuznetsov , Konstantin Sobolev This is my paper
classification cs.CV
keywords diffusion transformersparameter-efficient calibrationreward optimizationCMA-EStext-to-image generationblock scalinginference steps reductionmodel ensembling
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper claims that the standard Diffusion Transformer (DiT) architecture weights its own blocks sub-optimally, and that re-weighting each block's output with one learned scalar markedly improves generation quality. The authors show that ablating or scaling individual blocks (multiplying their residual output by a constant) can raise reward-model scores above the original model, and that optimizing roughly 100 such coefficients with an evolutionary strategy consistently improves three open text-to-image models. The method, Calibri, requires no fine-tuning of the base model and also reduces the number of inference steps needed—from 30–100 down to 10–30 in their tests—so calibrated models are both better-scoring and faster. A human study on two models reports decisively higher preference for calibrated outputs, which the authors offer as evidence that the gains reflect genuine perceptual quality rather than reward-model blind spots.

Core claim

On the paper's own terms, the central discovery is that DiT blocks are not equally useful as originally weighted: disabling certain blocks can improve reward-model scores, and for every block there exists an output scaling factor that improves the model over its baseline, as shown by scaling experiments on FLUX. This motivates treating the block/layer/gate outputs as learnable constants and optimizing them. Concretely, Calibri parameterizes the model's forward pass with a vector of scalars c = ω ∪ {s_i}, where ω are output-level weights and s_i are internal block/layer/gate coefficients, and solves c* = argmax_c R(c) with CMA-ES, where R is a text-to-image reward model such as HPSv3. Optimiz

What carries the argument

The load-bearing object is the scalar 'gate' multiplier γ applied to each sub-layer's residual output in the DiT/MM-DiT forward pass (Eq. 1–2). By freezing the base weights and treating these scalars (plus an output-level weight ω) as free parameters, Calibri converts the whole model into a low-dimensional, black-box function R(c) that can be maximized by CMA-ES without backpropagating through the diffusion process. The same scalar machinery is reused in Calibri Ensemble, where multiple sets of coefficients are optimized together and combined in a weighted sum, giving an ensemble interpretation that also covers classifier-free guidance as a two-model special case.

Load-bearing premise

The load-bearing premise is that reward-model scores are a trustworthy proxy for true generation quality; if the fitted scalars exploit reward-model blind spots instead of fixing real flaws, the central 'quality improvement' claim overstates what has been demonstrated.

What would settle it

A concrete test would be to run Calibri on a reward model that is explicitly designed to penalize anatomical inconsistencies (or to use artifact-specific detectors on the outputs), then measure human preference on prompts rich in such failures. If human preference does not follow the reward improvement — or if the calibrated model produces more extra limbs or distorted hands while scoring higher on HPSv3 — the claim that calibration improves genuine quality would be falsified.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If Calibri is right, any DiT-based text-to-image model can be improved post-hoc without retraining, by fitting about 100 scalars to a reward model.
  • The inference-step reduction means the same quality can be reached with roughly half to a third of the sampling steps, so calibrated models are cheaper to deploy.
  • Because calibration acts at the block output level, it composes with existing alignment methods (e.g., Flow-GRPO): applying Calibri to an already-aligned checkpoint further raises the target reward (and several other rewards).
  • The method is effectively training-free in the sense that no gradients flow through the base model; the one-time cost is offline CMA-ES search (tens to hundreds of GPU-hours), after which the scalars are baked into inference.
  • Calibri Ensemble suggests that combining differently calibrated versions of the same model behaves like a lightweight ensembling strategy, and for N=2 with block scaling it generalizes skip-layer guidance.

Where Pith is reading between the lines

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

  • A likely risk is reward hacking: because the coefficients are selected by a reward model that the authors themselves note is insensitive to artifacts like extra limbs, the improved scores may partly reflect blind spots. The reported human study on FLUX and Qwen-Image suggests the effect is not purely an artifact, but the mechanism limits how far one can trust metric-only comparisons on prompts whe
  • The discovery that some blocks can be disabled with net benefit hints at deeper architectural redundancy in DiTs; one testable extension is pruning or re-weighting blocks at training time, rather than post-hoc, to obtain the gains for free.
  • The 10–15 step sweet spot for calibrated ensembles suggests that block scaling can act as a type of guidance term that stabilizes sampling with few steps; combining this with step-distillation or consistency models may yield even larger speedups.
  • Because only the scalar multipliers are changed, the fixed points or Jacobian of the diffusion ODE are altered in a structured way; one could analyse how the optimal scalars depend on prompt distribution, potentially leading to prompt-conditional or adaptive calibration.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes Calibri, a post-hoc calibration method that reweights the outputs of DiT blocks, layers, or gates via ~100 scalar coefficients optimized with CMA-ES against a reward model (mainly HPSv3). The authors present motivation experiments showing that block ablation or scaling can improve ImageReward on FLUX, then report quality improvements across FLUX, SD-3.5M, and Qwen-Image, a 200-user human study, combination with Flow-GRPO alignment, calibration cost estimates, and supplementary diversity and reward-choice analyses. The central claim is that standard DiT architectures are sub-optimally weighted and that simple scalar calibration can improve generation quality and reduce inference steps.

Significance. If the central claim holds, the result is practically significant: a training-free (post-hoc), parameter-efficient way to improve several state-of-the-art text-to-image models, with an interesting architectural insight about uneven block contributions. The paper's strengths include breadth (three models, multiple reward metrics), the human evaluation, the explicit calibration-cost accounting, and the supplementary diversity analysis. However, the headline comparisons are confounded by simultaneous changes in inference steps, and the primary reported metric is also the optimization objective, so the current evidence does not cleanly separate the calibration effect from step-count reduction or reward overfitting.

major comments (4)
  1. [§4.2, Table 2, Table 3] The main cross-model comparison is confounded by different inference step counts. Calibri is evaluated at 15 NFE for FLUX and 30 NFE for SD-3.5M and Qwen-Image, while baselines are evaluated at 30, 80, and 100 NFE respectively. Since the calibration coefficients are optimized at 15 steps (Section 4, Implementation Details), the reported HPSv3/IR/Q-Align gains are a joint effect of calibration and step-count reduction. A same-step baseline is shown only for FLUX in Figure 6; no such control is provided for SD-3.5M or Qwen-Image. The human study in Table 3 also appears to compare faster Calibri models against baselines at their default step counts (the text notes Calibri is 2–3.3× faster), so it does not isolate the calibration effect. Please provide same-NFE baselines for all models, or otherwise disentangle the two factors.
  2. [§4 Implementation Details, Table 1, Supp. A] The paper optimizes exactly HPSv3 during calibration and then reports HPSv3 as the headline metric in Tables 1, 2, and 4. The HPSv3 gains (e.g., 11.41→13.48 in Table 2) are therefore partly fitted values, not independent predictions. The alternative-reward evidence is much weaker and inconsistent: in Table 1, gate scaling achieves the highest HPSv3 (13.48) but lower ImageReward (1.18) and Q-Align (4.88) than layer scaling (1.24 and 4.90). The statement in Supp. C that Calibri is 'not designed as a reward hacking method' is not fully supported by these numbers. The limitation in Supp. A (reward models are insensitive to artifacts such as extra limbs) further underscores the risk. Please add same-step evaluation on held-out reward metrics that were not used as objectives, and ideally report per-prompt significance for the human study.
  3. [§4.2, Table 2, Table 5] The granularity of calibration used for the main results in Table 2 is not stated. Table 5 lists different scaling types per model (Block/Layer/Gate/Gate/Gate) and different parameter counts, but the main text does not specify which scaling type produced the HPSv3/IR/Q-Align numbers in Table 2 or the human study in Table 3. This makes it difficult to interpret the comparisons and to reproduce the results. Please state explicitly which calibration scheme is used for each model and whether the same scheme was used across all reported metrics.
  4. [§4.2, Table 2] There is an internal discrepancy in the reported baseline step counts. The text says Calibri uses '15 steps compared to 30 for Flux, 40 for SD-3.5M, and 50 for Qwen-Image,' but Table 2 lists baseline NFE as 30, 80, and 100 respectively. If the text is correct, then the baseline step counts are lower than what is tabulated; if the table is correct, the text is wrong. This discrepancy directly affects the size of the claimed efficiency gain and must be corrected.
minor comments (4)
  1. [§3.3, Eq. (3)] The notation is inconsistent: Eq. (3) says c ∈ R^(L+1), but in the next paragraph c is defined as ω ∪ {s_i}_{i=1}^{L_s}. Please unify the notation for the number of calibration parameters and define L_s explicitly. Also, in the ensemble formula in §3.4, the sum over i is missing a summation symbol in the typeset equation.
  2. [§3.2 / Figure 3] The motivation experiment reports ImageReward scores but does not provide error bars or statistical significance despite using 5 seeds. This would strengthen the claim that scaling improves over the baseline for each block.
  3. [Table 6 (Supp.)] The supplementary diversity table is typeset in a way that is hard to parse; the rows for different Flow-GRPO/Calibri configurations are not visually separated, and 'NFE' values appear inconsistently. Please reformat.
  4. [Throughout] There are several typos and formatting artifacts: 'Stantdard' in §3.1, 'The The Ministry' in the Acknowledgments, the affiliation placeholder 'AXXX', and the reference numbering in the references list (e.g., [30] appears after [17] in the text but is numbered out of order).

Circularity Check

1 steps flagged

Partial circularity: the headline HPSv3 gains are the optimized objective; independent rewards and human evaluation keep the central claim from being circular by construction.

specific steps
  1. fitted input called prediction [Section 4 (Implementation Details; Evaluation and Metrics), Table 2]
    "We used HPSv3 [25] to track image preference and Q-Align [38] to track image quality during training. ... To measure the final metrics, we used HPSv3 [25], Q-Align [38] and ImageReward [39]."

    The calibration coefficients are selected by CMA-ES to maximize HPSv3 on candidate buckets; the same HPSv3 score is then reported as the primary evidence of improvement (Table 2, e.g., FLUX 11.41→13.48). Thus the HPSv3 gain is partly a fitted-value artifact, not an independent prediction. Circularity is partial: HPSv3 is evaluated on held-out HPDv3 prompts, non-target rewards (IR, Q-Align, PickScore) and a human study also improve.

full rationale

The paper's central claim (standard DiT is sub-optimally weighted; scalar calibration improves quality) does not reduce to a single fitted parameter because it is supported by multiple reward models and by a human study (Table 3). However, the most prominent quantitative evidence in Tables 2 and 4 is the HPSv3 score, which is also the optimization objective; that portion of the evidence is circular in the 'fitted input called prediction' sense. The supplementary cross-objective comparison (Table 7) mitigates reward-hacking concerns, although Section A concedes reward models are insensitive to artifacts such as extra limbs, so the reward-as-truth assumption remains load-bearing for backbones not covered by the human study. The NFE comparison (Calibri at 15/30/30 steps vs baselines at 30/80/100) is a confound for attribution, but it is not a formal circularity by construction. The only self-citation ([30] ImageRefl, sharing author Andrey Kuznetsov) is used for a diversity caveat and is not load-bearing. Overall, partial circularity on the headline metric, with independent content elsewhere: score 4.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

All free parameters are calibration coefficients and CMA-ES hyperparameters that are fitted or chosen rather than derived. The method rests on domain assumptions about reward-model validity and about scalar reweighting sufficiency. No new physical or conceptual entities are introduced.

free parameters (4)
  • Calibration coefficients c = ω ∪ {s_i} (block/layer/gate scalars) = not reported in paper
    The core of Calibri: 57-482 scalar multipliers fitted by CMA-ES to maximize reward; final coefficient values are not listed, so exact replication is impossible.
  • CMA-ES initial sigma = 0.25
    Hand-set exploration step size (Section 4, Implementation Details); affects which coefficients are found.
  • CMA-ES candidate count = 4+⌊3 ln d⌋ ≈ 20±3
    Population size chosen by standard convention; influences convergence and final objective value.
  • Bucket size and training inference steps = 16 images; 15 steps
    Evaluation batch size and generation-step budget during calibration; authors say 15 steps was found to be the lowest acceptable, making this a hand-chosen operating point.
axioms (4)
  • domain assumption DiT forward pass is exactly the residual form in Eq. (1)/(2), so multiplying block outputs by scalars preserves semantics.
    Calibration coefficients are inserted as multipliers on the gamma outputs in Eq. (1)-(2); if real implementations have extra nonlinear paths or skip connections not modeled, the optimized scalars may not correspond to actual components.
  • domain assumption Reward models HPSv3, Q-Align, ImageReward, PickScore measure human-preferred image quality sufficiently for optimization and evaluation.
    Calibri's objective R(c) is one of these reward models (Sections 3.3 and 4); Section A admits reward models miss anatomical artifacts, so optimization can exploit blind spots.
  • ad hoc to paper The suboptimality of a DiT can be captured by independent scalar multipliers on block/layer/gate outputs.
    The motivation (Fig. 3) shows scalar reweighting helps, but gives no reason that diagonal scalar reweighting is sufficient for optimal calibration rather than some other low-rank intervention.
  • domain assumption The prompt sets used for optimization (T2I-Compbench++ train/test) and evaluation (HPDv3) are representative of general text-to-image use.
    Coefficients are searched and selected on T2I-Compbench++ prompts and evaluated on HPDv3 prompts; if these sets are narrow, reward-model gains may not transfer to other prompts.

pith-pipeline@v1.3.0-alltime-deepseek · 12101 in / 14476 out tokens · 145518 ms · 2026-08-02T17:25:59.831462+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Calibri: Enhancing Diffusion Transformers via Parameter-Efficient Calibration." pith.science (2026). https://pith.science/paper/YTC7H27B

@misc{pith2026260324800,
  author       = {Pith},
  title        = {Pith review of: Calibri: Enhancing Diffusion Transformers via Parameter-Efficient Calibration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YTC7H27B}},
  note         = {Machine review of arXiv:2603.24800}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

In this paper, we uncover the hidden potential of Diffusion Transformers (DiTs) to significantly enhance generative tasks. Through an in-depth analysis of the denoising process, we demonstrate that introducing a single learned scaling parameter can significantly improve the performance of DiT blocks. Building on this insight, we propose Calibri, a parameter-efficient approach that optimally calibrates DiT components to elevate generative quality. Calibri frames DiT calibration as a black-box reward optimization problem, which is efficiently solved using an evolutionary algorithm and modifies just ~100 parameters. Experimental results reveal that despite its lightweight design, Calibri consistently improves performance across various text-to-image models. Notably, Calibri also reduces the inference steps required for image generation, all while maintaining high-quality outputs.

Figures

Figures reproduced from arXiv: 2603.24800 by Andrey Kuznetsov, Aysel Mirzoeva, Danil Tokhchukov, Konstantin Sobolev.

Figure 1
Figure 1. Figure 1: Introducing Calibri – a parameter-efficient method for diffusion transformer alignment. By optimizing only ∼ 102 parameters, Calibri significantly enhances the model’s generation quality. Abstract In this paper, we uncover the hidden potential of Dif￾fusion Transformers (DiTs) to significantly enhance gen￾erative tasks. Through an in-depth analysis of the de￾noising process, we demonstrate that introducing… view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of DiT architectural components. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Motivational Experiment: Our findings reveal that the contribution of DiT blocks is not fully optimized. We demonstrate that their performance can be enhanced through a straightforward output scaling using a scalar multiplier. prompts. To ensure statistical validity, we repeated each ex￾perimental configuration across 5 different random seeds. To assess the impact of the layer, we compute Image Re￾ward [39… view at source ↗
Figure 4
Figure 4. Figure 4: Illustration of calibration parameter search procedure. [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Quantitative comparison of various granularity levels for [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison between Calibri Ensemble and original model across several inference steps. the appropriate scaling strategy. N models. The Calibri Ensemble method (Section 3.4) allows us to aggregate multiple differently calibrated mod￾els into a single sampler. To validate this approach, we eval￾uate Calibri Ensemble on FLUX guided by HPSv3 reward with N \in \{1, 2\} models using HPDv3 prompts. In the ex￾peri… view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative evaluation of generation quality improvements across various baseline models. Models have same NFE as in Table [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative comparison of Calibri and Flow-GRPO [24] on SD-3.5M. Calibri achieves comparable performance with 105 fewer parameters and can be combined with alignment methods to boost either the same or different target metrics [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Limitations of modern reward models. The results demonstrate that the generation diversity of SD-3.5M with 40 inference timesteps remains compa￾rable to that of the model optimized by Calibri, which achieves comparable diversity while requiring only 15 in￾ference timesteps. Importantly, despite its reduced infer￾ence time, the model optimized by Calibri exhibits signif￾icantly higher generation quality com… view at source ↗
Figure 11
Figure 11. Figure 11: Comparison of CMA-ES and Flow-GRPO performance [PITH_FULL_IMAGE:figures/full_fig_p012_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: CMA-ES algorithm optimizes Calibri coefficients for layer scale FLUX [PITH_FULL_IMAGE:figures/full_fig_p013_12.png] 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

44 extracted references · 10 linked inside Pith

  1. [1]

    Stable flow: Vital layers for training-free image editing

    Omri Avrahami, Or Patashnik, Ohad Fried, Egor Nemchi- nov, Kfir Aberman, Dani Lischinski, and Daniel Cohen-Or. Stable flow: Vital layers for training-free image editing. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 7877–7888, 2025. 1, 2, 3

    2025

  2. [2]

    Training diffusion models with reinforce- ment learning

    Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforce- ment learning. InThe Twelfth International Conference on Learning Representations, 2024. 2

    2024

  3. [3]

    Align your latents: High-resolution video synthesis with la- tent diffusion models

    Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dock- horn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with la- tent diffusion models. InProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, pages 22563–22575, 2023. 2

    2023

  4. [4]

    Pixart-alpha : Fast training of diffusion transformer for photorealistic text-to-image syn- thesis

    Junsong Chen, YU Jincheng, GE Chongjian, Lewei Yao, Enze Xie, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-alpha : Fast training of diffusion transformer for photorealistic text-to-image syn- thesis. InThe Twelfth International Conference on Learning Representations. 2

  5. [5]

    Training-free layout control with cross-attention guidance

    Minghao Chen, Iro Laina, and Andrea Vedaldi. Training-free layout control with cross-attention guidance. InProceedings of the IEEE/CVF winter conference on applications of com- puter vision, pages 5343–5353, 2024. 2

    2024

  6. [6]

    Directly fine-tuning diffusion models on differentiable re- wards

    Kevin Clark, Paul Vicol, Kevin Swersky, and David J Fleet. Directly fine-tuning diffusion models on differentiable re- wards. InThe Twelfth International Conference on Learning Representations, 2024. 2

    2024

  7. [7]

    Diffusion self-guidance for control- lable image generation.Advances in Neural Information Processing Systems, 36:16222–16239, 2023

    Dave Epstein, Allan Jabri, Ben Poole, Alexei Efros, and Aleksander Holynski. Diffusion self-guidance for control- lable image generation.Advances in Neural Information Processing Systems, 36:16222–16239, 2023. 2

    2023

  8. [8]

    Scaling recti- fied flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas M ¨uller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling recti- fied flow transformers for high-resolution image synthesis. InForty-first international conference on machine learning,

  9. [9]

    Geneval: An object-focused framework for evaluating text- to-image alignment.Advances in Neural Information Pro- cessing Systems, 36:52132–52152, 2023

    Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text- to-image alignment.Advances in Neural Information Pro- cessing Systems, 36:52132–52152, 2023. 6

    2023

  10. [10]

    Reducing the time complexity of the derandomized evolution strategy with covariance matrix adaptation (cma- es).Evolutionary computation, 11(1):1–18, 2003

    Nikolaus Hansen, Sibylle D M ¨uller, and Petros Koumout- sakos. Reducing the time complexity of the derandomized evolution strategy with covariance matrix adaptation (cma- es).Evolutionary computation, 11(1):1–18, 2003. 2, 4

    2003

  11. [11]

    CMA-ES/pycma on Github

    Nikolaus Hansen, Youhei Akimoto, and Petr Baudis. CMA-ES/pycma on Github. Zenodo, DOI:10.5281/zenodo.2559634, 2019. 2, 4

    doi:10.5281/zenodo.2559634 2019

  12. [12]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceed- ings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. 2

    2016

  13. [13]

    Prompt-to-prompt im- age editing with cross attention control.arXiv preprint arXiv:2208.01626, 2022

    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

    Pith/arXiv arXiv 2022

  14. [14]

    Denoising dif- fusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020. 1

    2020

  15. [15]

    Denoising dif- fusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

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

    2020

  16. [16]

    Video dif- fusion models.Advances in neural information processing systems, 35:8633–8646, 2022

    Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video dif- fusion models.Advances in neural information processing systems, 35:8633–8646, 2022. 2

    2022

  17. [17]

    Kaiyi Huang, Chengqi Duan, Kaiyue Sun, Enze Xie, Zhen- guo Li, and Xihui Liu. T2i-compbench++: An enhanced and comprehensive benchmark for compositional text-to-image generation.IEEE Transactions on Pattern Analysis and Ma- chine Intelligence, 47:3563–3579, 2024. 5, 2

    2024

  18. [18]

    Spatiotemporal skip guidance for enhanced video diffusion sampling.IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), 2025

    Junha Hyung, Kinam Kim, Susung Hong, Min-Jung Kim, and Jaegul Choo. Spatiotemporal skip guidance for enhanced video diffusion sampling.IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), 2025. 6

    2025

  19. [19]

    Guiding a diffu- sion model with a bad version of itself.Advances in Neural Information Processing Systems, 2024

    Tero Karras, Miika Aittala, Tuomas Kynk ¨a¨anniemi, Jaakko Lehtinen, Timo Aila, and Samuli Laine. Guiding a diffu- sion model with a bad version of itself.Advances in Neural Information Processing Systems, 2024. 6

    2024

  20. [20]

    Pick-a-pic: An open dataset of user preferences for text-to-image generation.Ad- vances in neural information processing systems, 36:36652– 36663, 2023

    Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Ma- tiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation.Ad- vances in neural information processing systems, 36:36652– 36663, 2023. 2, 6

    2023

  21. [21]

    Flux.https://github.com/ black-forest-labs/flux, 2024

    Black Forest Labs. Flux.https://github.com/ black-forest-labs/flux, 2024. 1, 3, 5, 2

    2024

  22. [22]

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

    Pith/arXiv arXiv

  23. [23]

    Flow matching for genera- tive modeling

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maxim- ilian Nickel, and Matthew Le. Flow matching for genera- tive modeling. InThe Eleventh International Conference on Learning Representations. 1

  24. [24]

    Flow-grpo: Training flow matching models via on- line rl.arXiv preprint arXiv:2505.05470, 2025

    Jie Liu, Gongye Liu, Jiajun Liang, Yangguang Li, Jiaheng Liu, Xintao Wang, Pengfei Wan, Di Zhang, and Wanli Ouyang. Flow-grpo: Training flow matching models via on- line rl.arXiv preprint arXiv:2505.05470, 2025. 2, 5, 6, 8, 1

    Pith/arXiv arXiv 2025

  25. [25]

    Hpsv3: Towards wide-spectrum human preference score

    Yuhang Ma, Xiaoshi Wu, Keqiang Sun, and Hongsheng Li. Hpsv3: Towards wide-spectrum human preference score. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15086–15095, 2025. 2, 5, 1

    2025

  26. [26]

    Scalable diffusion models with transformers

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

  27. [27]

    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. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 1, 2

    2022

  28. [28]

    U- net: Convolutional networks for biomedical image segmen- tation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U- net: Convolutional networks for biomedical image segmen- tation. InInternational Conference on Medical image com- puting and computer-assisted intervention, pages 234–241. Springer, 2015. 1, 2

    2015

  29. [29]

    Freeu: Free lunch in diffusion u-net

    Chenyang Si, Ziqi Huang, Yuming Jiang, and Ziwei Liu. Freeu: Free lunch in diffusion u-net. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4733–4743, 2024. 2

    2024

  30. [30]

    Imagerefl: Balancing quality and di- versity in human-aligned diffusion models.arXiv preprint arXiv:2505.22569, 2025

    Dmitrii Sorokin, Maksim Nakhodnov, Andrey Kuznetsov, and Aibek Alanov. Imagerefl: Balancing quality and di- versity in human-aligned diffusion models.arXiv preprint arXiv:2505.22569, 2025. 1

    Pith/arXiv arXiv 2025

  31. [31]

    What the daam: Interpreting stable dif- fusion using cross attention

    Raphael Tang, Linqing Liu, Akshat Pandey, Zhiying Jiang, Gefei Yang, Karun Kumar, Pontus Stenetorp, Jimmy Lin, and Ferhan T ¨ure. What the daam: Interpreting stable dif- fusion using cross attention. InProceedings of the 61st An- nual Meeting of the Association for Computational Linguis- tics (Volume 1: Long Papers), pages 5644–5659, 2023. 2

    2023

  32. [32]

    Attention is all you need.Advances in neural information processing systems, 30, 2017

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko- reit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in neural information processing systems, 30, 2017. 2

    2017

  33. [33]

    Diffusion model align- ment using direct preference optimization

    Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model align- ment using direct preference optimization. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8228–8238, 2024. 2

    2024

  34. [34]

    Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314, 2025

    Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianx- iao Yang, et al. Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314, 2025. 1

    Pith/arXiv arXiv 2025

  35. [35]

    Unified reward model for multimodal understanding and generation.arXiv preprint arXiv:2503.05236, 2025

    Yibin Wang, Yuhang Zang, Hao Li, Cheng Jin, and Jiaqi Wang. Unified reward model for multimodal understanding and generation.arXiv preprint arXiv:2503.05236, 2025. 2

    Pith/arXiv arXiv 2025

  36. [36]

    Freeflux: Understanding and exploiting layer-specific roles in rope-based mmdit for versatile image editing.arXiv preprint arXiv:2503.16153, 2025

    Tianyi Wei, Yifan Zhou, Dongdong Chen, and Xingang Pan. Freeflux: Understanding and exploiting layer-specific roles in rope-based mmdit for versatile image editing.arXiv preprint arXiv:2503.16153, 2025. 2, 3

    Pith/arXiv arXiv 2025

  37. [37]

    Qwen-image technical report.arXiv preprint arXiv:2508.02324, 2025

    Chenfei Wu, Jiahao Li, Jingren Zhou, Junyang Lin, Kaiyuan Gao, Kun Yan, Sheng-ming Yin, Shuai Bai, Xiao Xu, Yilei Chen, et al. Qwen-image technical report.arXiv preprint arXiv:2508.02324, 2025. 1, 5

    Pith/arXiv arXiv 2025

  38. [38]

    Q-align: Teaching lmms for vi- sual scoring via discrete text-defined levels.arXiv preprint arXiv:2312.17090, 2023

    Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Yixuan Gao, Annan Wang, Erli Zhang, Wenxiu Sun, Qiong Yan, Xiongkuo Min, Guang- tao Zhai, and Weisi Lin. Q-align: Teaching lmms for vi- sual scoring via discrete text-defined levels.arXiv preprint arXiv:2312.17090, 2023. 5

    Pith/arXiv arXiv 2023

  39. [39]

    Imagere- ward: Learning and evaluating human preferences for text- to-image generation.Advances in Neural Information Pro- cessing Systems, 36:15903–15935, 2023

    Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagere- ward: Learning and evaluating human preferences for text- to-image generation.Advances in Neural Information Pro- cessing Systems, 36:15903–15935, 2023. 2, 3, 5

    2023

  40. [40]

    Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025. 3 Calibri: Enhancing Diffusion Transformers via Parameter-Efficient Calibration Supplementary Material Supplementary Material Structure This supplementary document is or...

    Pith/arXiv arXiv 2025

  41. [41]

    Section A elaborates on the limitations of the proposed methodology, providing a detailed analysis

  42. [42]

    Section B analyzes the diversity of images generated by the method, both before and after incorporating the Cal- ibri technique

  43. [43]

    Section C explains the rationale behind the chosen re- ward model, highlighting its impact on the system’s per- formance

  44. [44]

    Section D discusses the motivation for using the CMA- ES approach as the parameter search method, justifying its effectiveness. A. Limitations Our calibration coefficients selection method,Calibri, leverages a reward model [25] as its objective function. Re- ward models are trained to approximate user preferences for generated images, which enablesCalibri...