Pith. sign in

REVIEW 3 major objections 4 minor 48 references

Omegance: A Single Parameter for Various Granularities in Diffusion-Based Synthesis

T0 review · 3 major / 4 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read A single scalar, omega, multiplying the predicted noise in each reverse-diffusion step controls output granularity across image and video diffusion models without retraining.

desk verdict A simple inference-time noise-scaling knob that plausibly controls granularity across many diffusion models; the empirical case is solid, the mechanism is hand-waved, but it's worth referee time. read the letter →

arxiv 2411.17769 v2 pith:KISFZCTO submitted 2024-11-26 cs.CV

Xinyu Hou , Zongsheng Yue , Xiaoming Li , Chen Change Loy This is my paper · ORCID
classification cs.CV
keywords diffusionmodelsgranularitycontrolnoisepredictionscalingtext-to-imagesynthesistext-to-videodetailenhancementomegamaskschedule
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that a single scalar $\omega$, multiplying the predicted-noise term in every reverse-diffusion step, controls the granularity of generated images and video without retraining or architectural changes. Setting $\omega<1$ removes less noise, leaving residual high-frequency content that the model responds to by generating more intricate textures and more complex scenes; setting $\omega>1$ removes more noise and yields smoother, simpler outputs. The same scalar can be applied globally, as a spatial mask for region-specific control, or as a time-varying schedule to affect layout versus fine detail. This matters because existing detail-control tools usually require retraining, guidance, or architecture-specific tricks, whereas Omegance is a single knob that works across the schedulers and models tested.

What carries the argument

The central object is the scalar $\omega$ inserted into the reverse-diffusion update $z'_{t-1} = \delta_t z_t + \zeta_t \epsilon_\theta(z_t,t)\,\omega$, with an unbounded-input rescaling used in practice. The load-bearing identity is the modified SNR expression derived in Equation (5) for the DDIM scheduler (a deterministic diffusion sampling rule): any deviation of $\omega$ from 1 moves the sampling SNR away from the forward SNR, either leaving residual high-frequency content in the latent ($\omega<1$) or stripping it out ($\omega>1$). This identity turns a one-number change into a quantitative statement about detail, and it is what lets the method operate across different schedulers without retraining.

What would settle it

Run a fixed prompt and base model with $\omega<1$ and measure high-frequency energy, artifact rate, and text-alignment across a sweep of $\omega$ values; if some model produces visible noise or content drift instead of richer detail, or if high-frequency energy does not rise as $\omega$ decreases, the claimed control fails for that model.

Watch

Extended reading notes

Core claim

The central discovery is that the denoising network's noise prediction $\epsilon_\theta(z_t,t)$ can be rescaled by $\omega$ without breaking generation, and that this rescaling shifts the effective signal-to-noise ratio of each reverse step. The paper derives a modified SNR schedule (Equation 5) for the DDIM scheduler: $\omega<1$ makes the sampling SNR lower than the forward-process SNR, so the latent $z'_{t-1}$ retains extra high-frequency information; $\omega>1$ pushes the SNR higher and suppresses high frequencies. The authors state that the model 'justifies' residual noise by generating more intricate structures, which is the empirical mechanism behind richer detail. The same operation extends to Euler and flow-matching schedulers, to spatial $\omega$ masks, and to temporal $\omega$ schedules, and experiments across image and video models show consistent granularity shifts.

Load-bearing premise

The load-bearing premise is that a pretrained diffusion network responds to a latent carrying extra residual high-frequency noise by generating more complex structures and textures, rather than by producing artifacts or shifting content; this behavior is asserted for the tested models, not derived.

Editorial extensions

If this is right

  • Users can dial global granularity continuously: $\omega<1$ gives busier scenes and richer textures, while $\omega>1$ gives cleaner, smoother images.
  • Spatial omega masks let a character be sharpened while its background is simplified, or vice versa, with untouched regions left nearly unchanged.
  • Temporal omega schedules separate layout from texture: adjusting $\omega$ in roughly the first 10 of 50 denoising steps affects scene complexity, while later adjustments affect fine detail.
  • The method transfers across DDIM, Euler, and flow-matching schedulers and across text-to-image, image-to-image, inpainting, editing, and text-to-video models without retraining and with negligible extra compute.
  • The same knob can incidentally fix artifacts on weaker models ($\omega>1$ suppresses malformed fingers) and restore realism on over-smoothed models ($\omega<1$ re-adds texture in FLUX).

Reading between the lines

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

  • Editorial inference: because the stated mechanism is residual high-frequency content, the same omega scaling should transfer to any iterative denoiser with an explicit noise or velocity prediction, such as consistency models or discrete diffusion; testing there would map the true boundary of the claim.
  • Editorial inference: Equation (5) could be inverted to design omega schedules that steer the sampling SNR along any target curve, turning the hand-picked EXP and COS schedules in the paper into a principled scheduling problem.
  • Editorial inference: the mask results suggest a composable editing tool for existing images: an omega mask derived from segmentation could re-balance detail between subject and background during real-image editing, which the paper demonstrates for generation but not for post-hoc editing.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper introduces Omegance, a single scalar parameter omega that multiplies the predicted noise term in each reverse-diffusion step. The central claim is that omega<1 retains additional high-frequency residual noise, which the pretrained model interprets as a cue to generate more intricate structures and richer textures, while omega>1 removes more noise and produces smoother, simpler outputs. The same parameter is extended to spatial masks and time-dependent schedules, and experiments are reported across SDXL, FLUX, ControlNet, SDEdit, ReNoise, inpainting, and video diffusion models. Quantitative evidence includes FID, Inception Score, CLIP score, Q-Align, PickScore, SSIM, high-frequency energy measurements, and a user study, along with code released publicly.

Significance. If the central claim holds, Omegance is a practically valuable training-free and architecture-agnostic control for diffusion-based synthesis, with negligible computational cost and simple integration into existing samplers. The paper has concrete strengths: it releases code, evaluates on 1,000 prompts from DiffusionDB, includes multiple base models and tasks, and includes an explicit limitation statement. The claimed mechanism, however, is an empirical property of pretrained networks rather than a consequence of the SNR algebra, and the quantitative tables do not report variance or multiple-seed statistics. The mask and schedule extensions are plausible but would benefit from stronger regional and temporal consistency measurements.

major comments (3)
  1. [Sec. 3.2, Eq. (5)] The load-bearing mechanism is asserted rather than tested. Equation (5) computes the one-step SNR under the assumption that the input latent is on the forward noising manifold. After a modified step with omega != 1, the noise coefficient of the latent is no longer sqrt(1-alpha_{t-1}), so subsequent latents are off-manifold and the SNR identity does not directly apply. The key statement in Sec. 3.2 that the model 'justifies' residual noise by generating more intricate structures is an empirical assumption about how a pretrained network responds to off-manifold inputs, not a derivation from the SNR formula. If a network instead removes the residual as ordinary noise, omega<1 would produce little change or smoothing rather than richer detail. Please add a quantitative test of this assumption, for example by measuring the distribution of predicted noise and residual latents before and after modified steps, by comparing against explicit injection of Gaussian noise of comparable magnitude, and by reporting at least one failure case across the tested models.
  2. [Tables 1 and 2] The quantitative comparisons lack error bars, confidence intervals, or multiple-seed statistics. Table 1 reports FID, IS, CLIP, Q-Align, and PickScore on 1,000 prompts, but several differences are small relative to likely sampling noise, and Table 2 reports SSIM and high-frequency energy for single runs. Without multiple seeds or bootstrap intervals, the claims that Omegance 'outperforms' structure-modification and scheduler-based methods, or that HFE changes align with schedule design, are not firmly supported. Please report mean and standard deviation over at least three seeds, or use paired bootstrap confidence intervals.
  3. [Sec. 3.2.1] The locality claim for omega masks is stronger than the evidence provided. The text states that the omega mask 'leverages the locality of the denoising process, ensuring that adjustments to omega in one region do not affect the SNR or visual properties of neighboring areas.' However, latent diffusion backbones commonly have global receptive fields through self-attention or large convolutions, so the assertion is not obvious. The ControlNet results in Fig. 7 are qualitative only. Please add a quantitative region-consistency evaluation, for example measuring SSIM or LPIPS in the unmodified regions with and without masking, or temper the claim to state that cross-region effects are empirically small in the tested configurations.
minor comments (4)
  1. [Table 2] The column header reads 'HDE (Changes)' while the caption and text refer to High-Frequency Energy (HFE); please align the notation.
  2. [Fig. 4] The caption introduces S1(t) and S2(t) but does not define them; please add explicit definitions or point to the schedule curves shown in Fig. 6.
  3. [Eq. (5)] Equation (5) is typeset in a way that is difficult to parse; a simplified closed form or a step-by-step derivation in the main text would help readers verify the monotonicity argument.
  4. [General] There are minor typos and undefined abbreviations, for example 'denosing' in Sec. 3.2.1 and 'SD3' used without expansion; please proofread and expand all model names.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Omegance is a user-controlled inference-time scaling parameter, and the granularity effect is validated by external metrics and user studies rather than being built into the definition.

full rationale

The paper's central derivation is a direct algebraic consequence of inserting the scalar omega into a single DDIM denoising step (Eq. 4) and recomputing the resulting SNR (Eq. 5). This is not circular: omega is a freely chosen user parameter, not a constant fitted to the target outputs, and the SNR formula does not presuppose that lower or higher omega produces more or less detail. The qualitative claim that the pretrained network interprets residual high-frequency noise as a cue for richer structure is an empirical behavioral assumption, explicitly stated in Sec. 3.2 as the model 'justifies' the residual noise; this assumption is not derived from the equations, but it is also not assumed in the equations themselves. It is instead tested against external evidence: HFE measurements in Table 2, FID/IS/CLIP/Q-Align/PickScore in Table 1, and a 101-participant user study in Table 3, all compared against SDXL and other baselines. No load-bearing self-citation appears: the authors' own prior work [23] is cited only in related work and is not used to justify the method. There is no imported uniqueness theorem, no ansatz smuggled in by citation, and no renaming of a known result presented as a derivation. The limitation stated in the conclusion, that Omegance does not inherently improve base-model quality, further confirms that the paper does not overclaim a fitted or definitionally forced result. Concerns about whether the behavioral assumption holds for arbitrary future models are legitimate correctness risks, but they are not circularity.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The method introduces one user-facing control parameter, omega, with no fitted hidden constants. The key non-standard assumption is the network's behavioral response to residual noise, which is not derived and is the main risk. No new entities, forces, or dimensions are postulated.

free parameters (1)
  • omega (ϖ) = user-specified; examples use global ϖ=6.0 (suppress) and ϖ=-6.0 (enhance), plus hand-designed schedules…
    The method's effect depends on the user-chosen value of omega; it is not a fitted hidden parameter, but the demonstration values and schedule shapes are manually selected.
assumptions (3)
  • standard math Standard diffusion forward and reverse formulations (Equ. 1 and 2) hold for the target model.
    The SNR derivation and the scheduler-specific formulas rely on the standard DDPM/DDIM formulation adopted in Sec. 3.1.
  • domain assumption Denoising dynamics: layouts and object shapes emerge in early steps, fine details in late steps.
    Invoked in Sec. 3.1 and used to justify omega schedules; this is an empirical characterization from prior diffusion work, not proven in this paper.
  • ad hoc to paper A pretrained model responds to residual high-frequency noise by generating more intricate structures rather than artifacts.
    Stated in Sec. 3.2 as the model 'justifies' residual noise with more detail; this is the load-bearing behavioral assumption and is argued from examples, not derived.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Omegance: A Single Parameter for Various Granularities in Diffusion-Based Synthesis." pith.science (2026). https://pith.science/paper/KISFZCTO

@misc{pith2026241117769,
  author       = {Pith},
  title        = {Pith review of: Omegance: A Single Parameter for Various Granularities in Diffusion-Based Synthesis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KISFZCTO}},
  note         = {Machine review of arXiv:2411.17769}
}
abstract

In this work, we show that we only need a single parameter $\omega$ to effectively control granularity in diffusion-based synthesis. This parameter is incorporated during the denoising steps of the diffusion model's reverse process. This simple approach does not require model retraining or architectural modifications and incurs negligible computational overhead, yet enables precise control over the level of details in the generated outputs. Moreover, spatial masks or denoising schedules with varying $\omega$ values can be applied to achieve region-specific or timestep-specific granularity control. External control signals or reference images can guide the creation of precise $\omega$ masks, allowing targeted granularity adjustments. Despite its simplicity, the method demonstrates impressive performance across various image and video synthesis tasks and is adaptable to advanced diffusion models. The code is available at https://github.com/itsmag11/Omegance.

Figures

Figures reproduced from arXiv: 2411.17769 by the authors.

Figure 1
Figure 1. Omegance enables flexible granularity control over generation results. The control can be implemented globally, spatially with an omega mask, or temporally with an omega schedule. (Zoom-in for best view) arXiv:2411.17769v2 [cs.CV] 21 Jul 2025 [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Effects of Omegance on the frequency spectrum of the intermediate latent [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Global effects of Omegance. The models indicated below are the base models. The middle row shows the original base model [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Illustration of ω effect during denosing process. During the early stage (t ∈ [T, τ ]), a higher ω reduces layout complex￾ity (blue region), while a lower ω enhances it (red region). In the late stage (t ∈ [τ, 0]), a higher ω suppresses fine-grained details, whereas a …
Figure 6
Figure 6. Figure 6: Temporal effects of schedule-based Omegance. Four [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Spatial effects of mask-based Omegance in ControlNet results. [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: Effects of Omegance in Text-to-Video results. The em [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 28 canonical work pages

  1. [1]

    Self-rectifying diffu- sion sampling with perturbed-attention guidance

    Donghoon Ahn, Hyoungwon Cho, Jaewon Min, Wooseok Jang, Jungwoo Kim, SeonHwa Kim, Hyun Hee Park, Ky- ong Hwan Jin, and Seungryong Kim. Self-rectifying diffu- sion sampling with perturbed-attention guidance. In ECCV,

    work page

  2. [2]

    Art and Visual Perception

    Rudolf Arnheim. Art and Visual Perception. University of California Press, Berkeley, CA, 2nd, rev. and exp. ed., reprint 2020 edition, 2020. 2

    work page 2020

  3. [3]

    Chan, Yichang Chen, Sander Dieleman, and Yuqing Du et al

    Jason Baldridge, Jakob Bauer, Mukul Bhutani, Nicole Brich- tova, Andrew Bunner, Kelvin C.K. Chan, Yichang Chen, Sander Dieleman, and Yuqing Du et al. Imagen 3. arXiv preprint arXiv:2408.07009, 2024. 2

    arXiv 2024

  4. [4]

    SEGA: Instructing text-to-image models using semantic guidance

    Manuel Brack, Felix Friedrich, Dominik Hintersdorf, Lukas Struppek, Patrick Schramowski, and Kristian Kersting. SEGA: Instructing text-to-image models using semantic guidance. In NeurIPS, 2023. 3

    work page 2023

  5. [5]

    In- structPix2Pix: Learning to follow image editing instructions

    Tim Brooks, Aleksander Holynski, and Alexei A Efros. In- structPix2Pix: Learning to follow image editing instructions. In CVPR, 2023. 2

    2023

  6. [6]

    Deep reinforcement learn- ing from human preferences

    Paul F Christiano, Jan Leike, Tom Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep reinforcement learn- ing from human preferences. In NeurIPS, 2017. 3

    work page 2017

  7. [7]

    Following the masters: Portrait viewing and appreciation is guided by selective detail

    Steve Dipaola, Caitlin Riebe, and James Enns. Following the masters: Portrait viewing and appreciation is guided by selective detail. Perception, 2013. 2

    work page 2013

  8. [8]

    Scaling rec- tified 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, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling rec- tified flow transformers for high-resolution image synthesis. In ICML, 2024. 2, 5, 6

    work page 2024

Show all 48 references
  1. [9]

    DPOK: Reinforcement learning for fine-tuning text-to-image diffu- sion models

    Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Moham- mad Ghavamzadeh, Kangwook Lee, and Kimin Lee. DPOK: Reinforcement learning for fine-tuning text-to-image diffu- sion models. In NeurIPS, 2023. 3

    2023

  2. [10]

    ReNoise: Real im- age inversion through iterative noising

    Daniel Garibi, Or Patashnik, Andrey V oynov, Hadar Averbuch-Elor, and Daniel Cohen-Or. ReNoise: Real im- age inversion through iterative noising. In ECCV, 2024. 2, 6, 8

    2024

  3. [11]

    AnimateDiff: Animate your personalized text-to- image diffusion models without specific tuning

    Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. AnimateDiff: Animate your personalized text-to- image diffusion models without specific tuning. In ICLR,

  4. [12]

    Prompt-to-prompt image editing with cross-attention control

    Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In ICLR, 2023. 2

    2023

  5. [13]

    Gans trained by a two time-scale update rule converge to a local nash equilib- rium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilib- rium. In NeurIPS, 2017. 7

    2017

  6. [14]

    Classifier-free diffusion guidance

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 3

    2021

  7. [15]

    Denoising diffu- sion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffu- sion probabilistic models. In NeurIPS, 2020. 2, 3, 4

    2020

  8. [16]

    Imagen video: High definition video generation with diffusion mod- els

    Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion mod- els. arXiv preprint arXiv:2210.02303, 2022. 2

    arXiv

  9. [17]

    Improving sample quality of diffusion models us- ing self-attention guidance

    Susung Hong, Gyuseong Lee, Wooseok Jang, and Seungry- ong Kim. Improving sample quality of diffusion models us- ing self-attention guidance. In ICCV, 2023. 3

    2023

  10. [18]

    Elucidating the design space of diffusion-based generative models

    Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In NeurIPS, 2022. 5

    2022

  11. [19]

    Elucidating the design space of diffusion-based generative models

    Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In NeurIPS, 2022. 3

    2022

  12. [20]

    Pick-a-pic: An open dataset of user preferences for text-to-image generation

    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. In NeurIPS, 2023. 7

    2023

  13. [21]

    Hunyuanvideo: A systematic framework for large video generative models

    Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603, 2024. 8

    2024 arXiv

  14. [22]

    Black Forest Labs. FLUX. https://github.com/ black-forest-labs/flux, 2023. 2, 6

    2023

  15. [23]

    When stylegan meets stable diffusion: a W+ adapter for person- alized image generation

    Xiaoming Li, Xinyu Hou, and Chen Change Loy. When stylegan meets stable diffusion: a W+ adapter for person- alized image generation. In CVPR, 2024. 2

    2024

  16. [24]

    Common diffusion noise schedules and sample steps are flawed

    Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed. In WACV, 2024. 3, 7

    2024

  17. [25]

    Latte: Latent diffusion transformer for video generation

    Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zi- wei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 2, 6, 8

    2024 arXiv

  18. [26]

    SDEdit: Guided image synthesis and editing with stochastic differential equa- tions

    Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia- jun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equa- tions. In ICLR, 2022. 2, 3, 6, 8

    2022

  19. [27]

    Midjourney

    Inc Midjourney. Midjourney. https : / / www . midjourney.com/home, 2022. 2

    2022

  20. [28]

    Improved denoising diffusion probabilistic models

    Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 3, 7

    2021

  21. [29]

    Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Car- roll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, a...

    2022

  22. [30]

    SDXL: Improving latent diffusion mod- els for high-resolution image synthesis

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas M ¨uller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion mod- els for high-resolution image synthesis. In ICLR, 2024. 2, 6, 7, 8

    2024

  23. [31]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 7

    2021

  24. [32]

    Direct preference optimization: Your language model is secretly a reward model

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christo- pher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. In NeurIPS, 2023. 3

    2023

  25. [33]

    Seyedmorteza Sadat, Manuel Kansy, Otmar Hilliges, and Romann M. Weber. No training, no problem: Rethinking classifier-free guidance for diffusion models. arXiv preprint arXiv:2407.02687, 2024. 3

    2024 arXiv

  26. [34]

    Photorealistic text-to-image diffusion models with deep language understanding

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 2

    2022

  27. [35]

    Improved techniques for training gans

    Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. In NeurIPS, 2016. 7

    2016

  28. [36]

    Proximal policy optimization algo- rithms

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Rad- ford, and Oleg Klimov. Proximal policy optimization algo- rithms. arXiv preprint arXiv:1707.06347, 2017. 3

    2017 arXiv

  29. [37]

    RealVisXL V5.0

    SG161222. RealVisXL V5.0. https://civitai.com/ models/139562/realvisxl-v50, 2024. 6

    2024

  30. [38]

    FreeU: Free lunch in diffusion U-Net

    Chenyang Si, Ziqi Huang, Yuming Jiang, and Ziwei Liu. FreeU: Free lunch in diffusion U-Net. In CVPR, 2024. 3, 6, 7

    2024

  31. [39]

    Denois- ing diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denois- ing diffusion implicit models. In ICLR, 2021. 3, 4, 5

    2021

  32. [40]

    Score-based generative modeling through stochastic differential equa- tions

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Ab- hishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equa- tions. In ICLR, 2021. 2, 4

    2021

  33. [41]

    Genmo Team. Mochi 1. https: // github. com/ genmoai/models, 2024. 8

    2024

  34. [42]

    Bovik, H.R

    Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. In IEEE TIP, 2004. 7

    2004

  35. [43]

    Wang, Evan Montoya, David Munechika, Haoyang Yang, Benjamin Hoover, and Duen Horng Chau

    Zijie J. Wang, Evan Montoya, David Munechika, Haoyang Yang, Benjamin Hoover, and Duen Horng Chau. Diffu- sionDB: A large-scale prompt gallery dataset for text-to- image generative models. arXiv preprint arXiv:2210.14896,

    arXiv

  36. [44]

    Q-align: Teaching lmms for visual scoring via discrete text-defined levels

    Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Chunyi Li, Liang Liao, Annan Wang, Erli Zhang, Wenxiu Sun, Qiong Yan, Xiongkuo Min, Guangtao Zhai, and Weisi Lin. Q-align: Teaching lmms for visual scoring via discrete text-defined levels. In ICML, 2023. 7

    2023

  37. [45]

    Uncovering the disentanglement capability in text- to-image diffusion models

    Qiucheng Wu, Yujian Liu, Handong Zhao, Ajinkya Kale, Trung Bui, Tong Yu, Zhe Lin, Yang Zhang, and Shiyu Chang. Uncovering the disentanglement capability in text- to-image diffusion models. In CVPR, 2023. 3

    2023

  38. [46]

    ImageRe- ward: Learning and evaluating human preferences for text- to-image generation

    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. In NeurIPS, 2023. 3

    2023

  39. [47]

    Using human feedback to fine-tune diffusion models without any reward model

    Kai Yang, Jian Tao, Jiafei Lyu, Chunjiang Ge, Jiaxin Chen, Qimai Li, Weihan Shen, Xiaolong Zhu, and Xiu Li. Using human feedback to fine-tune diffusion models without any reward model. In CVPR, 2023. 3

    2023

  40. [48]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2, 6, 8

    2023

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.