Paper deep dive
Guidance Matters: Rethinking the Evaluation Pitfall for Text-to-Image Generation
Dian Xie, Shitong Shao, Lichen Bai, Zikai Zhou, Bojun Cheng, Shuo Yang, Jun Wu, Zeke Xie
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 7/20/2026, 10:18:54 AM
Summary
This paper identifies a critical evaluation pitfall in text-to-image generation where human preference models (e.g., HPS v2, ImageReward) exhibit a strong bias toward large Classifier-Free Guidance (CFG) scales, leading to inflated scores for oversaturated or artifact-ridden images. The authors propose a Guidance-Aware Evaluation (GA-Eval) framework that calibrates an 'effective guidance scale' to disentangle CFG effects from other guidance methods. They demonstrate that many recent diffusion guidance methods (e.g., SAG, PAG, FreeU) largely exploit this bias, as their performance degrades significantly when compared against CFG at equivalent effective guidance scales. Additionally, they introduce a Transcendent Diffusion Guidance (TDG) method to highlight this flaw.
Entities (10)
Relation Signals (5)
Human Preference Models → exhibitsbiastowards → Large Guidance Scales
confidence 95% · common human preference models exhibit a strong bias towards large guidance scales. Simply increasing the CFG scale can easily improve quantitative evaluation scores
TDG → demonstrates → Evaluation Pitfall
confidence 94% · we design Transcendent Diffusion Guidance (TDG) method that can significantly improve human preference scores in the conventional evaluation framework but actually does not work in practice.
GA-Eval → enablesfaircomparison → CFG
confidence 92% · introduce a novel guidance-aware evaluation (GA-Eval) framework that employs effective guidance scale calibration to enable fair comparison between current guidance methods and CFG
Recent Guidance Methods → suffersfrom → Winning Rate Degradation
confidence 90% · all methods suffer severely from winning rate degradation over standard CFG
Large Guidance Scales → causes → Image Quality Damage
confidence 88% · even if image quality is severely damaged (e.g., oversaturation and artifacts)
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Classifier-free guidance (CFG) has helped diffusion models achieve great conditional generation in various fields. Recently, more diffusion guidance methods have emerged with improved generation quality and human preference. However, can these emerging diffusion guidance methods really achieve solid and significant improvements? In this paper, we rethink recent progress on diffusion guidance. Our work mainly consists of four contributions. First, we reveal a critical evaluation pitfall that common human preference models exhibit a strong bias towards large guidance scales. Simply increasing the CFG scale can easily improve quantitative evaluation scores due to strong semantic alignment, even if image quality is severely damaged (e.g., oversaturation and artifacts). Second, we introduce a novel guidance-aware evaluation (GA-Eval) framework that employs effective guidance scale calibration to enable fair comparison between current guidance methods and CFG by identifying the effects orthogonal and parallel to CFG effects. Third, motivated by the evaluation pitfall, we design Transcendent Diffusion Guidance (TDG) method that can significantly improve human preference scores in the conventional evaluation framework but actually does not work in practice. Fourth, in extensive experiments, we empirically evaluate recent eight diffusion guidance methods within the conventional evaluation framework and the proposed GA-Eval framework. Notably, simply increasing the CFG scales can compete with most studied diffusion guidance methods, while all methods suffer severely from winning rate degradation over standard CFG. Our work would strongly motivate the community to rethink the evaluation paradigm and future directions of this field.
Tags
Links
- Source: https://arxiv.org/abs/2602.22570v1
- Canonical: https://arxiv.org/abs/2602.22570v1
Trouble viewing inline? Open PDF directly →
Full Text
73,110 characters extracted from source content.
Expand or collapse full text
Guidance Matters: Rethinking the Evaluation Pitfall for Text-to-Image Generation Dian Xie1 Shitong Shao1 Lichen Bai1 Zikai Zhou1 Bojun Cheng1 Shuo Yang2 Jun Wu3 Zeke Xie1† 1The Hong Kong University of Science and Technology (Guangzhou) 2Harbin Institute of Technology (Shenzhen) 3Cogniser Information Technology dxie810, sshao213, lbai292, zzhou440@connect.hkust-gz.edu.cn zekexie@hkust-gz.edu.cn Abstract Classifier-free guidance (CFG) has helped diffusion models achieve great conditional generation in various fields. Recently, more diffusion guidance methods have emerged with improved generation quality and human preference. However, can these emerging diffusion guidance methods really achieve solid and significant improvements? In this paper, we rethink recent progress on diffusion guidance. Our work mainly consists of four contributions. First, we reveal a critical evaluation pitfall that common human preference models exhibit a strong bias towards large guidance scales. Simply increasing the CFG scale can easily improve quantitative evaluation scores due to strong semantic alignment, even if image quality is severely damaged (e.g., oversaturation and artifacts). Second, we introduce a novel guidance-aware evaluation (GA-Eval) framework that employs effective guidance scale calibration to enable fair comparison between current guidance methods and CFG by identifying the effects orthogonal and parallel to CFG effects. Third, motivated by the evaluation pitfall, we design Transcendent Diffusion Guidance (TDG) method that can significantly improve human preference scores in the conventional evaluation framework but actually does not work in practice. Fourth, in extensive experiments, we empirically evaluate recent eight diffusion guidance methods within the conventional evaluation framework and the proposed GA-Eval framework. Notably, simply increasing the CFG scales can compete with most studied diffusion guidance methods, while all methods suffer severely from winning rate degradation over standard CFG. Our work would strongly motivate the community to rethink the evaluation paradigm and future directions of this field. 1 Introduction In recent years, diffusion models (Sohl-Dickstein et al., 2015; Ho et al., 2020; Song et al., 2021a; b; Lipman et al., 2023) have become cutting-edge techniques for generating high-quality contents, including image (Nichol et al., 2022; Rombach et al., 2022; Peebles and Xie, 2023; Esser et al., 2024), video (Ho et al., 2022; Blattmann et al., 2023b; a; Shao et al., 2025b), and 3D assets (Poole et al., 2023; Yi et al., 2024). The flexibility of diffusion process enabled versatile inference and training process (Shao et al., 2024; 2025c; Zhou et al., 2025; Huang et al., 2025), and can be applied to tasks in other domains (Li et al., 2025). A keys to empirical success lies on the powerful conditional and controllable generation ability of diffusion models, empowering rich AIGC applications. Classifier-free guidance (CFG) (Ho and Salimans, 2021) is exactly a very popular and powerful guidance method that enables direct control over the denoising process with text prompts. Recent studies (Hong et al., 2023; Hong, 2024; Ahn et al., 2024; Si et al., 2024; Chung et al., 2025; Bai et al., 2025a; Shao et al., 2025a; Zhou et al., 2025) have proposed several diffusion sampling or guidance methods to improve conditional generation quality. Meanwhile, in those studies, the evaluation of text-to-image generation starts to mainly depend on recent human preference models, such as HPS v2 (Wu et al., 2023), PickScore (Kirstain et al., 2023), and ImageReward (Xu et al., 2024), which are considered to be superior to conventional evaluation metrics. This is commonly true in human preference datasets. However, researcher overlooked that human often prefer images with gorgeous colors (Lin et al., 2024), which are correlated to highly saturated images generated with a large guidance scale. This overlooked phenomenon has led to a serious evaluation pitfall that human preference models would give better ratings on images generated with a relatively large guidance scale, as illustrated in Fig 1. We present several qualitative cases in Fig 2. It is easy for CFG to achieve comparable or better performance to many methods by simply increasing the guidance scale. We raise a key question: may recent advanced diffusion guidance methods overfit the overlooked large-guidance bias in human preference models? In this work, to validate how much the recent advanced methods’ performance is biased by the evaluation pitfall, we try to disentangle their effects from large CFG effects by calculating the effective guidance scale and then comparing the winning rates between a method and its effective CFG scale. By evaluating these methods with CFG under the same effective guidance scales, we can avoid such overlooked evaluation pitfall. We surprisingly discovered that most methods mainly exploited large guidance scale to achieve high performance, as their winning rates dropped considerably compared with effective CFG. Prompt: “An astronaut riding a horse.” HPS v2: 31.4546 HPS v2: 32.2331 HPS v2: 32.5681 HPS v2: 32.5991 Figure 1: The HPS v2 scores of generated images under different CFG scales ω∈5.5,10,15,20ω∈\5.5,10,15,20\, respectively. Model: Stable Diffusion-XL. HPSv2 exhibit a strong bias to large CFG scales in a wide range, even if generation quality starts to degrade due to too strong guidance. Figure 2: Visual comparison of different methods with their corresponding results under effective guidance scale ωeω^e. The e-CFG method, namely standard CFG with effective CFG scales calibrated by GA-Eval, can easily achieve performance improvement comparable to most recent diffusion guidance or sampling methods. Our contributions can be summarized as follows: • First, we reveal a critical evaluation pitfall that common human preference models, such as HPSv2 and ImageReward, exhibit a strong bias towards large guidance scales. As Fig 1 shows, simply increasing the CFG scale can easily improve quantitative evaluation scores due to strong semantic alignment, even if image quality is severely damaged (e.g., oversaturation and artifacts). • Second, we introduce a novel guidance-aware evaluation (GA-Eval) framework that employs effective guidance scale calibration to enable fair comparison between current guidance methods and CFG by identifying the effects orthogonal and parallel to CFG effects. • Third, motivated by the evaluation pitfall, we easily design Transcendent Diffusion Guidance (TDG) method that imitate the creation of weak condition in other methods during the sampling process. This is exactly an example method that can significantly improve human preference scores in the conventional evaluation framework but cannot actually improve generation quality over CFG. • Fourth, in extensive experiments, we empirically evaluated eight diffusion guidance methods within the conventional evaluation framework and the proposed GA-Eval framework. Notably, simply increasing the CFG scales can compete with most studied diffusion guidance methods, while all methods suffer severely from winning rate degradation over CFG. 2 Related Work 2.1 Diffusion Model Diffusion model (Sohl-Dickstein et al., 2015) generate high-quality samples by gradually adding noise to the data and learning the inverse process. Denoising diffusion probabilistic models (DDPM) (Ho et al., 2020) applies this process to image generation. In the reverse process, a pure Gaussian noise was gradually transformed to original data: t−1=1αt(t−βt1−α¯tϵθ(t,t))+σt,x_t-1= 1 _t (x_t- _t 1- α_tε_θ (x_t,t ) )+ _tz, (1) where ∼(0,)z (0,I), βt=1−αt _t=1- _t, and ϵθε_θ is the noise prediction network. For simplicity, we will ignore the argument t in ϵθ(t,t)ε_θ(x_t,t) in the following paragraphs. To improve sampling efficiency, denoising diffusion implicit models (DDIM) Song et al. (2021a) use a non-Markovian sampling process that reduces the number of inference steps while maintaining generation quality. The reverse process can be written as: t−1=αt−1(t−σtϵθ(t)αt)+σt−1ϵθ(t),x_t-1= _t-1 ( x_t- _t _θ (x_t ) _t )+ _t-1 _θ (x_t ), (2) where σt=1−αt _t= 1- _t. 2.2 Diffusion Guidance and Sampling Classifier-free guidance (CFG) (Ho and Salimans, 2021) interpolates conditional output and unconditional (or negative conditional) output of diffusion model, align the generated image to the corresponding condition. Given text prompt c, CFG use a same network ϵθ(⋅) _θ(·) to predict the unconditional noise ϵtuncond=ϵθ(t,∅).ε^uncond_t= _θ(x_t, ). and the conditional noise ϵtcond=ϵθ(t,).ε^cond_t= _θ(x_t,c). The updated noise in each denoising step was the interpolation between them under the guidance scale ω: ϵ~t=ϵtuncond+ω(ϵtcond−ϵtuncond). ε_t=ε^uncond_t+ω (ε^cond_t-ε^uncond_t ). (3) Self-Attention Guidance (SAG) (Hong et al., 2023), Perturbed-Attention Guidance (PAG) (Ahn et al., 2024), and Smoothed Energy Guidance (SEG) (Hong, 2024) uses blurred attention-map or perturbed attention-map to create weak condidion term in CFG. CFG++ (Chung et al., 2025) aiming at tackling the off-manifold challenges inherent in traditional CFG. They reformulate text-guidance as an inverse problem with a text-conditioned score matching loss. Z-Sampling and W2SD (Bai et al., 2025a; b) use the guidance gap between denoising and inversion to inject semantic information during sampling process. FreeU (Si et al., 2024) amplified specific features in U-Net (Ronneberger et al., 2015) to improve performance for free. Although these methods have satisfying performance on current benchmarks, it can also achieved by enlarging the guidance scale. Kynkänniemi et al. (2024) demonstrated that using CFG only when the noise intensity is within a certain range yields better results. Karras et al. (2024) used a degraded version of diffusion model to guide a complete one. But it is difficult to find a degraded model of a pretrained model. Adaptive Project Guidance (APG) (Sadat et al., 2025) project the latent parallel to the latent updated by conditional noise, which successfully eliminated over-saturation under large guidance scale. Nevertheless, it faces discrimination from many metrics, as they prefer images of high-saturation. 2.3 Evaluation for Text-to-Image Generation Traditional metrics for evaluating image quality include Inception Score (IS) (Salimans et al., 2016) and Fréchet Inception Distance (FID) (Heusel et al., 2017), but they cannot assess text-image alignment. Meanwhile, Jayasumana et al. (2024) have pointed out that FID and IS struggles to represent the rich content in modern image-to-text generation. The simplest way to evaluate images generated from a prompt is to use CLIP (Radford et al., 2021) to calculate text-image similarity, while it cannot reflect human preference. To address this problem, reward models are introduced to serve as metrics for human preference (Liu et al., 2024). HPS v2 (Wu et al., 2023), ImageReward (Xu et al., 2024), PickScore (Kirstain et al., 2023), MPS (Zhang et al., 2024), and Social Reward (Isajanyan et al., 2024) are evaluation metrics fine-tuned on human-preferenced data with reinforcement learning. These metrics are better aligned with human preferences, and have been widely used in benchmarking T2I models. Nevertheless, people naturally prefer colorful images, evaluation metrics based on human preferences tend to give higher scores to pictures with bright colors. Therefore, these metrics are easily blinded by images generated with a large guidance scale, as they also hold the same property of high-saturation. To the best of our knowledge, such evaluation pitfall have never been discussed before. 3 Guidance Matters to Evaluation Figure 3: Different evaluation metrics under different guidance scales. Model: Stable Diffusion-XL. Dataset: Pick-a-Pic. Except AES and PickScore, other metrics would give higher ratings on images generated with a larger guidance scale within ω∈[5.5,20]ω∈[5.5,20]. In this section, we first demonstrate that the performance improvements brought by many guidance methods are largly exaggerated and similar to large CFG, and then propose the GA-Eval framework for more fair and reasonable evaluation. Bias of Large Guidance Scale. We will present prevalent existence of the evaluation bias in this subsection. As some literatures (Bai et al., 2025a; Sadat et al., 2025) mentioned, diffusion models tend to generate over-saturated images under a large guidance scale, which will seriously affects the aesthetic quality of the image. Nevertheless, a larger guidance scale would amplify the predicted noise ϵ~t ε_t to the direction of conditional noise ϵtcondε^cond_t in every timestep, thus making the generated image more aligned to the prompt. However, this phenomenon could bias many evaluation metrics based on CLIP (Radford et al., 2021), as the CLIP model assumes the image is well-aligned to the prompt in this case. Moreover, for metrics fine-tuned on human-preferenced data, the overlap between the training data and over-saturated results making these metrics easily biased by large guidance scale. We give case studies and quantitative results at Fig 1, Fig 2 and Fig 3. Most of T2I evaluation metrics would prefer images generated with a large guidance scale. Such bias may lead to results that do not reflect the real performance of other guidance methods, since we can simply increasing the guidance scale to acquire comparable or better performance on these metrics. Effective Guidance Scale. Because better performance on can be achieved by increasing the guidance scale in CFG, we need a method to verify that whether a method’s performance can be achieved by large guidance scale. Since this phenomenon was closely related to guidance scale, we can decouple the CFG component from other methods to calculate the effective guidance scale ωeω^e, thus verifying whether they exploit large guidance scale to achieve better performance. For naive CFG in Eq 3, the updated noise in every timestep ϵ~t ε_t is the linear combination of the conditional noise ϵtcondε^cond_t and the unconditional noise ϵtuncondε^uncond_t. The guidance scale ω denoted as: ω=ϵ~t−ϵtuncondϵtcond−ϵtuncond=ϵ~t−ϵtuncondΔϵ.ω= ε_t-ε^uncond_tε^cond_t-ε^uncond_t= ε_t-ε^uncond_t ε. (4) For other guidance methods, their noise updated in every timestep ϵ~t∗ ε_t^* can be decomposed to ϵtuncondε^uncond_t and two other components about Δϵ ε: ϵ~t∗=ϵtuncond+ϵt∥+ϵt⟂, ε_t^*=ε^uncond_t+ _t + _t , (5) where ϵt⟂ _t is the orthogonal component of ϵ~t∗−ϵtuncond ε_t^*-ε^uncond_t in the direction of Δϵ ε, and ϵt∥ _t is the parallel component of ϵ~t∗−ϵtuncond ε_t^*-ε^uncond_t in the direction of Δϵ ε. Since any vector can be decomposed into vectors orthogonal to each other. According to the vector projection method, the orthogonal component ϵt⟂ _t can be calculated by: ϵt⟂ _t =(ϵ~t∗−ϵtuncond)−ProjΔϵ(ϵ~t∗−ϵtuncond) = ( ε_t^*-ε^uncond_t )-Proj_ ε ( ε_t^*-ε^uncond_t ) (6) =(ϵ~t∗−ϵtuncond)−⟨ϵt∗~−ϵtuncond,Δϵ⟩‖Δϵ‖2Δϵ. = ( ε_t^*-ε^uncond_t )- _t^*-ε^uncond_t, ε \| ε\|^2 ε. Therefore, the parallel component ϵt∥ _t is given by: ϵt∥=⟨ϵt∗~−ϵtuncond,Δϵ⟩⟨Δϵ,Δϵ⟩Δϵ. _t = _t^*-ε^uncond_t, ε ε, ε ε. (7) We define the effective guidance scale at timestep t is the ratio of the amplitude of the parallel component ϵt∥ _t a to the amplitude of the guiding direction Δϵ ε. ϵ~t∗=ϵtuncond+ωteΔϵ+ϵt⟂, ε_t^*=ε^uncond_t+ω^e_t ε+ _t , (8) Thus, the effective guidance scale can be acquired by: ωte=‖ϵt∥‖Δϵ‖,ω^e_t= \| _t \|\| ε\|, (9) where ∥⋅∥\|·\| represents the ℓ2 _2-norm. However, some guidance methods (Bai et al., 2025a; Chung et al., 2025) modified the sampling process during the update of the latent tx_t, rather than explicitly modifying the predicted noise ϵt~ _t. But these two approaches are interchangeable. For instance, in the case of DDIM from Eq 2, we can simply reformulate the process as follows: ϵ~t∗=αtt−1−αt−1tαtβt−1−αt−1βt. ε_t^*= _tx_t-1- _t-1x_t _t _t-1- _t-1 _t. (10) Therefore, we can acquire ϵ~t∗ ε_t^* from the original latent tx_t and the sampled latent t−1x_t-1. Note that the specific formulation of Eq 10 depends on the scheduler used by corresponding guidance method. Additionally, since ωteω^e_t may vary across different timesteps t, we use a fixed effective guidance scale for each timestep by averaging along the sampling path: ωe=1T∑tωteω^e= 1T _tω^e_t. Guidance-aware Evaluation. In this part, we introduce the GA-Eval framework that empirically evaluate diffusion guidance methods with effective guidance scales. Although such bias could leads to an evaluation pitfall, we can use the bias itself to give a fair comparison with effective guidance scale. To determine whether a method’s performance is exaggerated, we can measure how much the winning rate degrades after applying our settings to original CFG. Given prompt set =ii=1NC=\c_i\_i=1^N, the images generated from each guidance method, CFG, and CFG with effective guidance scale denoted as (i∗,iCFG,ie-CFG)i=1N \ (X_i^*,X_i^CFG,X_i^e-CFG ) \_i=1^N , where they were generated with corresponding prompt i∈c_i . Denote binary comparison functions as follows: iCFG _i^CFG =1ℳ(i*,i)≻ℳ(iCFG,i),0otherwise, = (11) ie-CFG _i^e-CFG =1ℳ(i*,i)≻ℳ(ie-CFG,i),0otherwise, = where ℳM is an evaluation metric, ≻ indicates the rating is better on the left side. The corresponding winning rate is given by: ηCFG=1N∑i=1NiCFG,ηe-CFG=1N∑i=1Nie-CFG.η^CFG= 1N _i=1^NI_i^CFG, η^e-CFG= 1N _i=1^NI_i^e-CFG. (12) By denoting the degradation of the winning rate, Δη=ηCFG−ηe-CFG η=η^CFG-η^e-CFG, we can quantify whether a method just exploits a large guidance scale to achieve high performance. Figure 4: Comparison of classifier-free guidance Ho and Salimans (2021) and transcendent diffusion guidance (TDG). 4 Transcendent Diffusion Guidance In this section, we show that, motivated by the overlooked evaluation pitfall, we can easily design an effective TDG method that significantly increases the popular human preference metrics in the conventional evaluation framework but exposures the pitfall in the proposed GA-Eval framework. Numerous recent works have demonstrated that use weak conditions or models can improve the generation performance of diffusion models, without any additional training or input. Like under-trained model (Karras et al., 2024), perturbed attention map Ahn et al. (2024); Hong et al. (2023), skipped layers (Hyung et al., 2025; Chen et al., 2025), etc. Motivated by these recent works, we propose TDG to imitate the creation of weak condition. Specifically, TDG randomly replace tokens in text prompt c with empty token ∅ , thus creating a weakened text prompt ∗c^*. Then use the noise prediction network ϵθ(⋅) _θ(·) to acquire the weak conditional score ϵweak _weak, which combined with ϵcond _cond and ϵuncond _uncond in CFG to the diffusion sampling process. We illustrate TDG in Fig 4 and present the pseudo-code in Algorithm 1. Due to page limits, we formally introduce TDG with more details and results in Appendix C. 5 Empirical Analysis and Discussion In this section, we empirically study how recent diffusion guidance methods perform in conventional text-to-image evaluation framework and our GA-Eval framework. 5.1 Experimental Settings Models: We choose Stable Diffusion-XL (Podell et al., 2024), Stable Diffusion 2.1 (Rombach et al., 2022), Stable Diffusion 3.5 (Esser et al., 2024), and DiT-XL/2 (Peebles and Xie, 2023) for image generation. We use ω=5.5ω=5.5, T=50T=50 for SD-XL and SD-2.1, ω=4.5ω=4.5, T=32T=32 for SD-3.5, and ω=4ω=4, T=50T=50 for DiT-XL/2. Datasets: We use the first 100 prompts from Pick-a-Pic dataset (Kirstain et al., 2023), 200 prompts from DrawBench dataset (Saharia et al., 2022), Human Preferenced Dataset (HPD) (Wu et al., 2023) with 3200 prompts, and GenEval dataset (Ghosh et al., 2024) with 553 prompts. We also benchmarked some methods on COCO-30K (Lin et al., 2014). And ImageNet-50K (Deng et al., 2009) for the evaluation class-to-image generation. Evaluation Metrics: We use HPS v2 (Wu et al., 2023), Aesthetics-Predictor (AES) (Schuhmann et al., 2022), PickScore (Kirstain et al., 2023), ImageReward (Xu et al., 2024), and CLIPScore (Radford et al., 2021) and Fréchet Inception Distance (FID) (Heusel et al., 2017), Inception Score (IS) (Barratt and Sharma, 2018; Salimans et al., 2016) for COCO-30K or ImageNet-50K. Baselines: We select following guidance methods for comparison: Z-Sampling (Bai et al., 2025a), CFG++ (Chung et al., 2025), Perturbed-Attention Guidance (PAG) (Ahn et al., 2024), Self-Attention Guidance (SAG) (Hong et al., 2023), Smoothed-Energy Guidance (SEG) (Hong, 2024), FreeU (Si et al., 2024), Adaptive Project Guidance (APG) (Sadat et al., 2025), and Transcendent Diffusion Guidance (TDG) proposed in Sec 3. The detailed hyper-parameters of each method can be found at Sec B. 5.2 Main Results Figure 5: The winning rate ηe-CFGη^e-CFG compared to effective CFG and their degradation Δη η of different methods on HPD dataset. Left: ηe-CFGη^e-CFG. Right: Δη η. Among all methods, applying ωeω^e has minor influence to APG. Meanwhile, AES demonstrated negative Δη η on many methods, which means AES would give worse ratings to images generated with a large guidance scale. Table 1: The winning rates ηCFG,ηe-CFGη^CFG,η^e-CFG and its degradation Δη η of different methods on various datasets. Model: Stable Diffusion-XL (ω=5.5ω=5.5). We sort these methods based on average ηe-CFGη^e-CFG, labeling those greater than 55% in blue and the others in red. Most of methods have ωe>ω^e>ω, and suffering from severe winning rate degradation. Only recent Z-Sampling can still maintain high winning rates ηe-CFGη^e-CFG after setting CFG scale as ωeω^e, while some methods even cannot compete with effective CFG. Method HPS v2 ImageReward AES PickScore CLIPScore Average η ωeω^e Pick-a-Pic (ηCFG/ηe-CFG/Δη^CFG/η^e-CFG/ η) (ηCFGη^CFG/ηe-CFGη^e-CFG) Z-Sampling 90% / 74% / 16% 75% / 63% / 12% 65% / 76% /-11% 72% / 73% / -1% 64% / 59% / 5% 73% / 69% !20 13.51 CFG++ 82% / 68% / 14% 70% / 58% / 12% 39% / 53% /-14% 56% / 57% / -1% 60% / 54% / 6% 61% / 58% !20 8.91 SAG 75% / 59% / 16% 45% / 34% / 11% 64% / 66% / -2% 62% / 56% / 6% 53% / 44% / 9% 60% / 52% !20 8.14 TDG 62% / 48% / 14% 61% / 46% / 15% 49% / 42% / 7% 51% / 51% / 0% 63% / 53% / 10% 57% / 48% !20 8.27 SEG 60% / 56% / 4% 59% / 49% / 10% 57% / 56% / 1% 35% / 33% / 2% 47% / 42% / 5% 52% / 47% !20 6.10 PAG 61% / 48% / 13% 51% / 40% / 11% 55% / 50% / 5% 45% / 42% / 3% 49% / 44% / 5% 52% / 45% !20 5.98 FreeU 52% / 42% / 10% 56% / 44% / 12% 53% / 55% / -2% 27% / 25% / 2% 42% / 33% / 9% 46% / 40% !20 7.47 APG 20% / 20% / 0% 38% / 29% / 9% 56% / 51% / 5% 45% / 45% / 0% 47% / 46% / 1% 41% / 38% !20 15.05 DrawBench (ηCFG/ηe-CFG/Δη^CFG/η^e-CFG/ η) (ηCFGη^CFG/ηe-CFGη^e-CFG) Z-Sampling 90% / 68% / 22% 76% / 59% / 17% 69% / 64% / 4% 72% / 60% / 12% 62% / 50% / 12% 74% / 60% !20 11.94 PAG 61% / 60% / 1% 54% / 50% / 5% 60% / 64% / -4% 49% / 44% / 4% 42% / 46% / -4% 53% / 53% !20 6.04 TDG 67% / 50% / 17% 64% / 56% / 8% 48% / 50% / -2% 56% / 51% / 5% 54% / 53% / 1% 58% / 52% !20 8.24 SAG 73% / 60% / 12% 58% / 43% / 15% 64% / 66% / -2% 53% / 52% / 2% 46% / 40% / 6% 59% / 52% !20 7.42 CFG++ 77% / 60% / 18% 64% / 55% / 9% 56% / 53% / 3% 52% / 42% / 10% 55% / 46% / 9% 61% / 51% !20 8.89 SEG 64% / 60% / 4% 50% / 50% / 1% 68% / 66% / 2% 40% / 38% / 2% 38% / 40% / -2% 52% / 51% !20 6.13 FreeU 48% / 37% / 11% 54% / 46% / 8% 64% / 62% / 2% 25% / 22% / 2% 46% / 45% / 1% 47% / 43% !20 7.44 APG 22% / 21% / 2% 36% / 36% / 1% 46% / 45% / 1% 46% / 44% / 1% 48% / 48% / -1% 40% / 39% !20 13.56 HPD (ηCFG/ηe-CFG/Δη^CFG/η^e-CFG/ η) (ηCFGη^CFG/ηe-CFGη^e-CFG) Z-Sampling 91% / 72% / 20% 74% / 60% / 14% 59% / 67% / -8% 70% / 65% / 5% 55% / 45% / 10% 70% / 62% !20 12.41 CFG++ 83% / 66% / 16% 67% / 57% / 10% 44% / 48% / -4% 51% / 43% / 9% 61% / 52% / 9% 61% / 53% !20 9.08 SAG 80% / 64% / 17% 54% / 42% / 11% 61% / 64% / -3% 55% / 50% / 5% 48% / 43% / 5% 60% / 53% !20 7.36 PAG 68% / 65% / 4% 51% / 49% / 3% 57% / 58% / -1% 44% / 43% / 1% 42% / 40% / 2% 53% / 51% !20 5.96 SEG 70% / 66% / 5% 54% / 50% / 3% 57% / 58% / -1% 37% / 34% / 3% 40% / 38% / 2% 52% / 49% !20 6.07 TDG 67% / 47% / 20% 61% / 51% / 10% 42% / 47% / -5% 55% / 50% / 5% 55% / 47% / 8% 56% / 48% !20 8.31 FreeU 63% / 50% / 12% 55% / 50% / 6% 51% / 55% / -4% 26% / 21% / 5% 38% / 34% / 4% 47% / 42% !20 7.38 APG 18% / 19% / -2% 38% / 39% / -1% 51% / 51% / 1% 47% / 44% / 3% 50% / 52% / -2% 41% / 41% !20 13.29 Table 2: The quantitative results of different methods on GenEval. Model: Stable Diffusion-XL (ω=5.5ω=5.5). We labeled methods surpassed effective CFG with blue, and others with red. Increasing the CFG scale can also significantly increase the GenEval scores. Method Single object (↑ ) Two object (↑ ) Counting (↑ ) Colors (↑ ) Position (↑ ) Color attribution (↑ ) Overall (↑ ) ωeω^e CFG 98.91% 75.63% 37.66% 85.11% 10.87% 26.88% 55.84% 5.5 TDG 100.00% 76.77% 53.75% 87.23% 11.00% 21.00% 58.29% !20 8.28 e-CFG 100.00% 73.74% 47.50% 86.17% 9.00% 27.00% 57.23% Z-Sampling 100.00% 75.76% 46.25% 84.04% 17.00% 20.00% 57.18% !20 16.05 e-CFG 98.75% 77.78% 36.25% 85.11% 16.00% 25.00% 56.48% CFG++ 100.00% 81.82% 46.25% 85.11% 16.00% 20.00% 58.20% !20 8.81 e-CFG 100.00% 76.77% 48.75% 87.23% 12.00% 24.00% 58.13% FreeU 98.75% 74.75% 41.25% 76.60% 9.00% 25.00% 54.22% !20 7.62 e-CFG 100.00% 76.77% 48.75% 88.30% 12.00% 25.00% 58.47% PAG 93.75% 69.70% 37.50% 82.98% 9.00% 26.00% 53.15% !20 6.13 e-CFG 100.00% 76.77% 42.50% 85.11% 9.00% 28.00% 56.90% SAG 97.50% 75.76% 46.25% 86.17% 9.00% 25.00% 56.61% !20 7.36 e-CFG 100.00% 78.79% 43.75% 86.17% 10.00% 26.00% 57.45% SEG 97.50% 71.72% 41.25% 81.91% 9.00% 18.00% 53.23% !20 6.19 e-CFG 98.75% 77.78% 41.25% 84.04% 9.00% 26.00% 56.14% APG 98.75% 70.71% 40.00% 87.23% 12.00% 27.00% 55.95% !20 14.12 e-CFG 100.00% 78.79% 46.25% 89.36% 12.00% 31.00% 59.57% According to GA-Eval, almost all methods exaggerated their real performance which cannot be reflected in conventional evaluation paradigm. We present the results of eight diffusion guidance methods for SD-XL in Table 1. We also give an illustration of winning rate and its degradation on HPD dataset in Fig 5. Except APG that even cannot compete with standard CFG, most methods suffer from significant winning rate degradations over effective CFG in GA-Eval. The HPS v2 degradation in CFG++, SAG, TDG, and Z-Sampling are even greater than 15%15\%. Most of baselines, except Z-Sampling, have the averaged ηe-CFGη^e-CFG near or worse than 50%. In the experiment of SD-2.1, Table 4 in Sec 4, our GA-Eval also reveals the prevalently overlooked evaluation pitfall. We also use GenEval to evaluate the fine-grained property and semantic alignment. In Table 2, most methods are worse than effective CFG in terms of the average GenEval scores. Despite suffering winning rate degradations, some methods still maintained a considerable winning rate after applying ωeω^e. In Table 1, the winning rate of HPS v2 of Z-Sampling is approximately 70%, where the averaged winning rate is greater than 60% across different datasets. CFG++ also has certain but lower superiority. Meanwhile, many metrics obviously biased by large guidance scale. Given ωe>ω^e>ω, most metrics have Δη>0 η>0, which means they raised ratings to images from effective CFG with large guidance scale. Among these metrics, only AES have reversed results. GenEval also demonstrated the evaluation pitfall brought by large guidance scales. Besides, we conduct experiments on COCO-30K and ImageNet-50K with Z-Sampling and CFG++. Due to page limit, the results are presented in Table 8 and Table 9 in Sec D. After applying ωeω^e, the performance of e-CFG are getting closer to the corresponding method. However, they are still worse than the original CFG in FID and IS. 5.3 Discussion and Analysis Evaluation Metrics. The results in various metrics in Table 1 align with our observations in Fig 3. HPSv2, ImageReward, and PickScore are largely affected by large guidance scales. This might be because they were trained on human-preferenced data with high saturated images well aligned to the prompt, which are very similar to images generated with a large guidance scale. For CLIPScore, its semantic alignment may also contribute to such bias, as large guidance scales naturally improve image-prompt alignment. In GenEval, large guidance scale also contributes to the increased semantic correctness of object-level properties. Although most metrics are biased, AES provides a fair evaluation, since it only evaluates images. However, this also makes it unable to evaluate prompt-following abilities in generated images. The AIGC community eagerly needs a human preference model robust to large CFG. Guidance Methods. Since most evaluation metrics are biased, CFG with larger guidance scale of course bring severe winning rate degradations to these studied methods. In principle, many methods implicitly amplified their effective guidance scale. Z-Sampling uses a normal guidance scale in denoising, and a smaller guidance scale in inversion. The guidance gap between denoising and inversion indirectly amplified the guidance scale. For methods like PAG, SAG, SEG, and TDG, which created an additional weak condition term to the updated noise in each timestep, the difference of conditional noise and weak conditional noise is parallel to Δϵ ε with a significant portion, which result a larger ωeω^e. The formulation of CFG++ also increased the guidance scale compared to original CFG. Notably, APG holds a very low winning rate before and after apply ωeω^e, with only AES maintained normal. It is because APG alleviate over-saturation in generated images, other biased metrics would give worse evaluation ratings. So these metrics cannot reflect the real performance of APG. APG did not exploit large guidance scale, although it explicitly uses large CFG scales ω=15ω=15 for SD-XL and ω=10ω=10 for SD-21. For methods maintained effectiveness compared to effective CFG, like Z-Sampling and CFG++. This indicates that the performance improvement of these methods is not solely achieved by increasing the guidance scale. They indeed have some effective components orthogonal to CFG. Meanwhile, their FID and IS also indicate that their distribution are shifted away from ground-truth compared to CFG, since the later was trained on ground-truth images with a fixed guidance scale. Diffusion Models. Not limited to results of SD-XL in Table 1, we have conducted experiments on other models as well. The results of SD-2.1 and SD-3.5 are presented in Table 4, and Table 3 in Sec D. In SD-2.1, for CFG++ and Z-Sampling, their winning rate degradations are even larger than SD-XL (above 20%). For simple models like SD-2.1, it is quite easier to rectify the sampling trajectory to improve the generation quality, although they often collapse to a larger guidance scale. For SD-3.5, its generation capability is already very powerful, which limits the improvements due to guidance methods. 6 Conclusion In this work, we report a critical but overlooked evaluation pitfall hidden in recent advances in diffusion sampling and its evaluation for text-to-image generation. These pitfalls severely damage development of this line of research. As human preference models exhibit strong bias to large guidance scales, we are the first to reveal that most recent diffusion guidance methods even cannot outperform standard CFG with large guidance scales. To more properly evaluate diffusion guidance methods, we propose a novel GA-Eval framework that disentangles the effects orthogonal and parallel to the CFG effect. Motivated by the findings, we also propose TDG to have improved evaluation scores in the conventional evaluation framework. Through comprehensive experiments in the GA-Eval framework, we expose that most claimed improvements, including TDG, essentially mirror the effects of simply increasing the CFG scale, with all methods showing inflated performance metrics under proper evaluation. We believe that this work serves as a crucial wake-up call for the community to rethink the evaluation paradigm for AIGC tasks and more properly recognize true innovation in diffusion models. Acknowledgments This work was supported by the Science and Technology Bureau of Nansha District Under Key Field Science and Technology Plan Program No. 2024ZD002 and Dream Set Off - Kunpeng and Ascend Seed Program. References D. Ahn, H. Cho, J. Min, W. Jang, J. Kim, S. Kim, H. H. Park, K. H. Jin, and S. Kim (2024) Self-rectifying diffusion sampling with perturbed-attention guidance. In European Conference on Computer Vision, p. 1–17. Cited by: §1, §2.2, §4, §5.1. L. Bai, S. Shao, Z. Zhou, Z. Qi, Z. Xu, H. Xiong, and Z. Xie (2025a) Zigzag diffusion sampling: the path to success is zigzag. The Thirteenth International Conference on Learning Representations. Cited by: §1, §2.2, §3, §3, §5.1. L. Bai, M. Sugiyama, and Z. Xie (2025b) Weak-to-strong diffusion with reflection. In The Fourteenth International Conference on Learning Representations, Cited by: §2.2. S. Barratt and R. Sharma (2018) A note on the inception score. arXiv preprint arXiv:1801.01973. Cited by: §5.1. A. Blattmann, T. Dockhorn, S. Kulal, D. Mendelevitch, M. Kilian, D. Lorenz, Y. Levi, Z. English, V. Voleti, A. Letts, et al. (2023a) Stable video diffusion: scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127. Cited by: §1. A. Blattmann, R. Rombach, H. Ling, T. Dockhorn, S. W. Kim, S. Fidler, and K. Kreis (2023b) Align your latents: high-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 22563–22575. Cited by: §1. J. Chang, Y. Fang, P. Xing, S. Wu, W. Cheng, R. Wang, X. Zeng, G. Yu, and H. Chen (2025) OneIG-bench: omni-dimensional nuanced evaluation for image generation. arXiv preprint arXiv:2506.07977. Cited by: Appendix D. C. Chen, J. Zhu, X. Feng, N. Huang, M. Wu, F. Mao, J. Wu, X. Chu, and X. Li (2025) Sˆ 2-guidance: stochastic self guidance for training-free enhancement of diffusion models. arXiv preprint arXiv:2508.12880. Cited by: §4. H. Chung, J. Kim, G. Y. Park, H. Nam, and J. C. Ye (2025) Cfg++: manifold-constrained classifier free guidance for diffusion models. The Thirteenth International Conference on Learning Representations. Cited by: §1, §2.2, §3, §5.1. J. Deng, W. Dong, R. Socher, L. Li, K. Li, and L. Fei-Fei (2009) Imagenet: a large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, p. 248–255. Cited by: §5.1. P. Esser, S. Kulal, A. Blattmann, R. Entezari, J. Müller, H. Saini, Y. Levi, D. Lorenz, A. Sauer, F. Boesel, et al. (2024) Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first International Conference on Machine Learning, Cited by: §1, §5.1. D. Ghosh, H. Hajishirzi, and L. Schmidt (2024) Geneval: an object-focused framework for evaluating text-to-image alignment. Advances in Neural Information Processing Systems 36. Cited by: §5.1. M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter (2017) Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30. Cited by: §2.3, §5.1. J. Ho, W. Chan, C. Saharia, J. Whang, R. Gao, A. Gritsenko, D. P. Kingma, B. Poole, M. Norouzi, D. J. Fleet, et al. (2022) Imagen video: high definition video generation with diffusion models. arXiv preprint arXiv:2210.02303. Cited by: §1. J. Ho, A. Jain, and P. Abbeel (2020) Denoising diffusion probabilistic models. Advances in neural information processing systems 33, p. 6840–6851. Cited by: §1, §2.1. J. Ho and T. Salimans (2021) Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, Cited by: §1, §2.2, Figure 4. S. Hong, G. Lee, W. Jang, and S. Kim (2023) Improving sample quality of diffusion models using self-attention guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 7462–7471. Cited by: §1, §2.2, §4, §5.1. S. Hong (2024) Smoothed energy guidance: guiding diffusion models with reduced energy curvature of attention. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, Cited by: §1, §2.2, §5.1. R. Huang, S. Shao, Z. Zhou, P. Zhao, H. Guo, T. Ye, L. Bai, S. Yang, and Z. Xie (2025) Diffusion dataset condensation: training your diffusion model faster with less data. arXiv preprint arXiv:2507.05914. Cited by: §1. J. Hyung, K. Kim, S. Hong, M. Kim, and J. Choo (2025) Spatiotemporal skip guidance for enhanced video diffusion sampling. In Proceedings of the Computer Vision and Pattern Recognition Conference, p. 11006–11015. Cited by: §4. A. Isajanyan, A. Shatveryan, D. Kocharian, Z. Wang, and H. Shi (2024) Social reward: evaluating and enhancing generative ai through million-user feedback from an online creative community. In The Twelfth International Conference on Learning Representations, Cited by: §2.3. S. Jayasumana, S. Ramalingam, A. Veit, D. Glasner, A. Chakrabarti, and S. Kumar (2024) Rethinking fid: towards a better evaluation metric for image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 9307–9315. Cited by: §2.3. T. Karras, M. Aittala, T. Kynkänniemi, J. Lehtinen, T. Aila, and S. Laine (2024) Guiding a diffusion model with a bad version of itself. Advances in Neural Information Processing Systems. Cited by: Appendix D, §2.2, §4. Y. Kirstain, A. Polyak, U. Singer, S. Matiana, J. Penna, and O. Levy (2023) Pick-a-pic: an open dataset of user preferences for text-to-image generation. Advances in Neural Information Processing Systems 36, p. 36652–36663. Cited by: §1, §2.3, §5.1, §5.1. T. Kynkänniemi, M. Aittala, T. Karras, S. Laine, T. Aila, and J. Lehtinen (2024) Applying guidance in a limited interval improves sample and distribution quality in diffusion models. Advances in Neural Information Processing Systems. Cited by: §2.2. B. Li, Z. Xie, X. Zhang, X. Zhu, and Z. Lei (2025) Diffusion models are zero-shot generative text-vision retrievers. In ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), p. 1–5. Cited by: §1. C. Lin, S. Mottaghi, and L. Shams (2024) The effects of color and saturation on the enjoyment of real-life images. Psychonomic bulletin & review 31 (1), p. 361–372. Cited by: §1. T. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollár, and C. L. Zitnick (2014) Microsoft coco: common objects in context. In European conference on computer vision, p. 740–755. Cited by: §5.1. Y. Lipman, R. T. Chen, H. Ben-Hamu, M. Nickel, and M. Le (2023) Flow matching for generative modeling. In International Conference on Learning Representations, Cited by: §1. B. Liu, S. Shao, B. Li, L. Bai, Z. Xu, H. Xiong, J. Kwok, S. Helal, and Z. Xie (2024) Alignment of diffusion models: fundamentals, challenges, and future. arXiv preprint arXiv:2409.07253. Cited by: §2.3. Y. Ma, X. Wu, K. Sun, and H. Li (2025) Hpsv3: towards wide-spectrum human preference score. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 15086–15095. Cited by: Appendix D. A. Q. Nichol, P. Dhariwal, A. Ramesh, P. Shyam, P. Mishkin, B. Mcgrew, I. Sutskever, and M. Chen (2022) GLIDE: towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, p. 16784–16804. Cited by: §1. W. Peebles and S. Xie (2023) Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 4195–4205. Cited by: §1, §5.1. D. Podell, Z. English, K. Lacey, A. Blattmann, T. Dockhorn, J. Müller, J. Penna, and R. Rombach (2024) SDXL: improving latent diffusion models for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations, Cited by: §5.1. B. Poole, A. Jain, J. T. Barron, and B. Mildenhall (2023) DreamFusion: text-to-3d using 2d diffusion. In The Eleventh International Conference on Learning Representations, Cited by: §1. A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al. (2021) Learning transferable visual models from natural language supervision. In International conference on machine learning, p. 8748–8763. Cited by: §2.3, §3, §5.1. R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer (2022) High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 10684–10695. Cited by: §1, §5.1. O. Ronneberger, P. Fischer, and T. Brox (2015) U-net: convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part I 18, p. 234–241. Cited by: §2.2. S. Sadat, O. Hilliges, and R. M. Weber (2025) Eliminating oversaturation and artifacts of high guidance scales in diffusion models. The Thirteenth International Conference on Learning Representations. Cited by: §2.2, §3, §5.1. C. Saharia, W. Chan, S. Saxena, L. Li, J. Whang, E. L. Denton, K. Ghasemipour, R. Gontijo Lopes, B. Karagol Ayan, T. Salimans, et al. (2022) Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems 35, p. 36479–36494. Cited by: §5.1. T. Salimans, I. Goodfellow, W. Zaremba, V. Cheung, A. Radford, and X. Chen (2016) Improved techniques for training gans. Advances in neural information processing systems 29. Cited by: §2.3, §5.1. C. Schuhmann, R. Beaumont, R. Vencu, C. Gordon, R. Wightman, M. Cherti, T. Coombes, A. Katta, C. Mullis, M. Wortsman, et al. (2022) Laion-5b: an open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, p. 25278–25294. Cited by: §5.1. S. Shao, L. Bai, H. Xiong, Z. Xie, et al. (2025a) IV-mixed sampler: leveraging image diffusion models for enhanced video synthesis. In The Thirteenth International Conference on Learning Representations, Cited by: §1. S. Shao, H. Yi, H. Guo, T. Ye, D. Zhou, M. Lingelbach, Z. Xu, and Z. Xie (2025b) MagicDistillation: weak-to-strong video distillation for large-scale few-step synthesis. arXiv preprint arXiv:2503.13319. Cited by: §1. S. Shao, Z. Zhou, D. Xie, Y. Fang, T. Ye, L. Bai, and Z. Xie (2025c) CoReˆ 2: collect, reflect and refine to generate better and faster. arXiv preprint arXiv:2503.09662. Cited by: §1. S. Shao, Z. Zhou, T. Ye, L. Bai, Z. Xu, and Z. Xie (2024) Bag of design choices for inference of high-resolution masked generative transformer. arXiv preprint arXiv:2411.10781. Cited by: §1. C. Si, Z. Huang, Y. Jiang, and Z. Liu (2024) Freeu: free lunch in diffusion u-net. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 4733–4743. Cited by: §1, §2.2, §5.1. J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan, and S. Ganguli (2015) Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, p. 2256–2265. Cited by: §1, §2.1. J. Song, C. Meng, and S. Ermon (2021a) Denoising diffusion implicit models. In International Conference on Learning Representations, Cited by: §1, §2.1. Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole (2021b) Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, Cited by: §1. Y. Wang, Y. Zang, H. Li, C. Jin, and J. Wang (2025) Unified reward model for multimodal understanding and generation. arXiv preprint arXiv:2503.05236. Cited by: Appendix D. X. Wu, Y. Hao, K. Sun, Y. Chen, F. Zhu, R. Zhao, and H. Li (2023) Human preference score v2: a solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341. Cited by: §1, §2.3, §5.1, §5.1. J. Xu, X. Liu, Y. Wu, Y. Tong, Q. Li, M. Ding, J. Tang, and Y. Dong (2024) Imagereward: learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems 36. Cited by: §1, §2.3, §5.1. T. Yi, J. Fang, J. Wang, G. Wu, L. Xie, X. Zhang, W. Liu, Q. Tian, and X. Wang (2024) Gaussiandreamer: fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 6796–6807. Cited by: §1. S. Zhang, B. Wang, J. Wu, Y. Li, T. Gao, D. Zhang, and Z. Wang (2024) Learning multi-dimensional human preference for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 8018–8027. Cited by: §2.3. Z. Zhou, S. Shao, L. Bai, S. Zhang, Z. Xu, B. Han, and Z. Xie (2025) Golden noise for diffusion models: a learning framework. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 17688–17697. Cited by: §1, §1. Appendix A LLM Usage In this paper, no Large Language Model (LLM) was employed as a research or writing tool. All research conception, data analysis, writing, and editing were independently accomplished by the author. Appendix B Experimental Settings All our experiments were conducted on NVIDIA GeForce RTX 4090 GPUs with 24GB VRAM. In SD-XL, we use guidance scale ω=5.5ω=5.5 and inference step T=50T=50. For Z-Sampling, we use inversion guidance scale ωinv=0 _inv=0, optimization steps λ=Tλ=T. For PAG, we use PAG scale s=3s=3, and apply PAG on layer “mid” of the backbone of SD-XL. For CFG++, we use guidance scale ω=0.4ω=0.4. For TDG, we use g=1.8g=1.8 and β=2.6β=2.6. For APG, we use guidance scale ω=15ω=15. For FreeU, we use b1=1.3b_1=1.3, b2=1.4b_2=1.4, s1=0.9s_1=0.9, and s2=0.2s_2=0.2. For SAG, we use SAG scale s=0.75s=0.75. For SEG, we use SEG scale γseg=3 _seg=3, standard deviation of Gauusian blur σ=100σ=100, and apply on layer “mid” of the backbone of SD-XL. In SD-2.1, we use guidance scale ω=5.5ω=5.5 and inference step T=50T=50. For Z-Sampling, we use inversion guidance scale ωinv=0 _inv=0, optimization steps λ=Tλ=T. For PAG, we use PAG scale s=3s=3, and apply PAG on layer “mid” of the backbone of SD-XL. For CFG++, we use guidance scale ω=0.4ω=0.4. For FreeU, we use b1=1.4b_1=1.4, b2=1.6b_2=1.6, s1=0.9s_1=0.9, and s2=0.2s_2=0.2. For SAG, we use SAG scale s=0.75s=0.75. In SD-3.5, we use 4-bit NormalFloat quantization to run on RTX-4090, with guidance scale ω=4.5ω=4.5 and inference step T=32T=32. For Z-Sampling, we use inversion guidance scale ωinv=0 _inv=0, optimization steps λ=Tλ=T. For PAG, we use PAG scale s=3s=3, and apply PAG on layer “13” of the backbone of SD-3.5. For TDG, we use g=1.6g=1.6 and β=2.4β=2.4. In DiT-XL/2, we use guidance scale ω=4ω=4 and timestep T=50T=50. For Z-Sampling, we use inversion guidance scale ωinv=0 _inv=0, optimization steps λ=Tλ=T. For CFG++, we use guidance scale ω=0.5ω=0.5. Appendix C Transcedent Diffusion Guidance In this section, we will present limitations of previous guidance techniques. After that, we will introduce the proposed method transcendent diffusion guidance (TDG), and illustrate how can we improve the sampling process of diffusion models by enlarging search space. C.1 Motivation Previous improvements to CFG have been limited to the guidance scale ω or use self-defined functions to substitute ϵuncond _uncond or ϵcond _cond. However, these methods are limited to the operation in Eq 3: interpolation between the two outputs of the diffusion model. To solve this problem, we need to expand the search space of the sampling process. Obviously, due to the search space of CFG restricted on a line, we only need to have a point outside of this line to obtain the whole hyperplane as the search space. Besides, the point should be acquired through the text prompt c without introducing any additional inputs. Driven by the analysis above, we propose transcendent diffusion guidance to improve the sampling process of diffusion models, which illustrated in Fig. 4. Specifically, TDG corrupts the text prompt c to create a weakened text prompt ∗c^*. Then use the noise prediction network ϵθ(⋅) _θ(·) to acquire the weak conditional score ϵweak _weak. Combining with ϵcond _cond and ϵuncond _uncond in CFG, we can expand the search space to the entire hyperplane. During the sampling process, larger search space will accelerate the convergence of latent towards the optimal solution, thereby improving the generation quality of the model. C.2 Weak Condition In this subsection, we will provide a detailed explanation on how to obtain a weak condition. As mentioned in the main text, many works demonstrated that introduce an additional weak model or weak condition could improve the generation performance. Meanwhile, they should avoid introducing any additional input and be computationally efficient. Therefore, we obtain the weak condition outside of CFG by modifying the text prompt. Specifically, consider a text prompt =[c1,c2,…,cn]c=[c_1,c_2,…,c_n], where cic_i is the i-th token of c. We randomly choose some tokens with index i∈ℐi , and replace them with empty token ∅ : ci∗=∅,i∈ℐci,i∉ℐ.c^*_i= cases ,&i \\ c_i,&i cases. (13) By combining these new tokens together, we obtained the weakened prompt ∗=[c1∗,…,cn∗]c^*=[c_1^*,…,c_n^*]. In practice, |ℐ|=n2|I|= n2, which means we randomly replace half tokens in c with ∅ . On the one hand, the modified prompt only requires the original prompt to be obtained without introducing additional input. On the other hand, to obtain ∗c^*, it only needs to be calculated at the beginning, which is computationally efficient compared to those methods need to manipulating attention maps at every denoising step. Therefore, the weak conditional score is given by: ϵtweak=ϵθ(t,∗).ε^weak_t= _θ(x_t,c^*). (14) C.3 Algorithm Algorithm 1 The Sampling of Transcendent Diffusion Guidance Input: text condition c, guidance scale factor g, balance scale factor β, random replace ratio λ ∗=random_replace(,λ)c^*=random\_replace(c,λ) T∼(0,)x_T (0,I) for t in T,T−1,…,1T,T-1,…,1 do ϵuncondt=ϵθ(t,∅) _uncond^t= _θ(x_t, ) ϵcondt=ϵθ(t,) _cond^t= _θ(x_t,c) ϵweakt=ϵθ(t,∗) _weak^t= _θ(x_t,c^*) ϵ~t=12(ϵtuncond+ϵtweak)+ω⋅g⋅β+1(ϵtcond−ϵtuncond)+ω⋅gβ+1(ϵtcond−ϵtweak)⋅‖ϵtcond−ϵtuncond‖ϵtcond−ϵtweak‖ ε_t= 12(ε^uncond_t+ε^weak_t)+ ω· g·β+1(ε^cond_t-ε^uncond_t)+ ω· gβ+1(ε^cond_t-ε^weak_t)· \|ε^cond_t-ε^uncond_t \| \|ε^cond_t-ε^weak_t \| t−1∼(1α¯t(xt−βt1−α¯tϵ~t),Σt)x_t-1 ( 1 α_t(x_t- _t 1- α_t ε_t), _t) end for return 0x_0 After obtaining the weak condition, we can construct the hyperplane through an extra guidance term. Compared to CFG in Eq 3, TDG reformulated the diffusion score in each denoising step with following add-on: ϵ~t= ε_t= 12(ϵtuncond+ϵtweak)+ω⋅g⋅β+1(ϵtcond−ϵtuncond)+ 12(ε^uncond_t+ε^weak_t)+ ω· g·β+1(ε^cond_t-ε^uncond_t)+ (15) ω⋅gβ+1(ϵtcond−ϵtweak)⋅‖ϵtcond−ϵtuncond‖ϵtcond−ϵtweak‖, ω· gβ+1(ε^cond_t-ε^weak_t)· \|ε^cond_t-ε^uncond_t \| \|ε^cond_t-ε^weak_t \|, where g and β are guidance scale factor and balance scale factor, respectively. For noise prediction network ϵθ(⋅) _θ(·), the linear combination of conditional score ϵtcondε^cond_t, the weak conditional score ϵtweakε^weak_t, and the unconditional score ϵtuncondε^uncond_t will create three non-collinear points, thus creating a hyperplane. While the original CFG only limits its search space the combination of ϵtcondε^cond_t and ϵtuncondε^uncond_t, which is a straight line. Therefore, the model can have a hyperplane as the search space, which transcend the restrictions of the original search space of CFG. We give a brief illustration in Fig 8. And the hyperparameter grid search result was illustrated in Fig 9. The pseudo-code is provided in Alg 1. Fig 4 also briefly illustrates the differences between TDG and CFG. Qualitative comparison results are presented in Fig 6. The hyperparameters of TDG are determined through grid search, and the criteria is HPSv2 winning rate compared to original CFG. The results can be found in Fig 9. Meanwhile, we give an analysis on the number of inference step T to demonstrate the superiority of TDG. In Fig 7, TDG with T=16T\!=\!16 achieved comparable performance on CFG with T=20T\!=\!20, and TDG always maintain better performance than CFG when T≥16T\!≥\!16. We also present ablation results of TDG in Fig 10. There is no significant performance degradation when mask ratio smaller than 0.6. Figure 6: The qualitative results of TDG. Figure 7: Comparison of HPS v2 Score of TDG and CFG at different timesteps. Model: SD-3.5. Figure 8: TDG vs. CFG in manifold space. Compared to CFG, TDG enlarges the search space to a hyperplane. Figure 9: TDG hyperparameter grid search heatmap, with HPSv2 winning rate compared to CFG as the criteria. Model: SD-XL. Figure 10: Different mask ratio of prompt tokens in TDG. Model: SD-XL. Appendix D Supplementary Empirical Results We give more cases of HPS v2 scores increasing with ω in Fig 11, the phenomenon can be verified across different prompts, and has further verified in Fig 3. For simplicity, we give the quantitative performance of SD-XL, SD-2.1, and SD-3.5 are given in Table 5, Table 6, and Table 7, respectively. Specifically, the winning rates and the degradation of different methods on SD-3.5 is given in Table 3. The results are consistent with our findings in the maintext. More cases of visual comparison between other methods and their corresponding effective CFG are presented below. In addition, we test recent reward models and benchmark, like HPSv3 (Ma et al., 2025), UnifiedReward (Wang et al., 2025), and OneIG-Bench (Chang et al., 2025), to valid their robustness on the evaluation pitfall. And the results are presented in Table 10, 11, and 12. HPSv3 still hold this evaluation pitfall, with most methods suffer winning rate degradation. Meanwhile, due to OneIG-Bench require to use theirs prompts to evaluate, we only conduct experiments on SD-XL with different guidance scales. The Alignment score, Reasoning score, and Text score are increasing with guidance scale. But Diversity score decreases on the contrary, this also meet the discovery in (Karras et al., 2024). To further demonstrate that human-preferenced metrics favor saturated images, we give 2 toy experiments on images with different saturation: 1) Real images with different saturation; 2) Generated images with different guidance scale; We also use Spearman’ test to valid the correlations between the image saturation and metric score, the alpha value was set to 0.05. • 1) For real images, we randomly selected 20 Ground-Truth images from MS-COCO dataset, and changed their saturation with different levels. The scatter map are presented in Fig 12. Most prompts have metric score positively correlated with image saturation. In HPSv2, Significant positive correlations: 14, Positive correlation not significant: 1, No significant positive correlation: 5. In ImageReward, Significant positive correlations: 14, Positive correlation not significant: 2, No significant positive correlation: 4. • 2) For generated images, we randomly select 20 prompts from Pick-a-Pic, DrawBench, and HPD. Then generated then with different guidance scales. The scatter map are presented in Fig 13. And the problem still remains: In HPSv2, Significant positive correlations: 15, Positive correlation not significant: 2, No significant positive correlation: 3. In ImageReward, Significant positive correlations: 12, Positive correlation not significant: 1, No significant positive correlation: 7. We analyzed the ωteω^e_t distribution over timestep t. In Fig 14 (a), every method have their own ωteω^e_t distribution on t, but the ωteω^e_t averaged on t is still greater than the baseline ω=5.5ω=5.5. In Fig 14 (b), the different scheduler have minor effect on the ωteω^e_t. Prompt: “A beautiful blonde woman with freckles eating desserts with a dog at a dining table. The background of Eiffel Tower with neon fireworks lighting up the night sky.” HPS v2: 29.7943 HPS v2: 30.5297 HPS v2: 30.3956 HPS v2: 30.7722 Prompt: “A cat, chubby, very fine wispy and extremely long swirly wavy fur, under water, Kuniyoshi Utagawa, Hishida Shuns.” HPS v2: 28.6557 HPS v2: 29.2244 HPS v2: 29.6573 HPS v2: 29.7427 Prompt: “A brown colored giraffe.” HPS v2: 29.6685 HPS v2: 29.9727 HPS v2: 30.2884 HPS v2: 30.3286 Figure 11: The HPS v2 scores of generated images under different CFG scales ω∈5.5,10,15,20ω∈\5.5,10,15,20\, respectively. Model: Stable Diffusion-XL. HPSv2 exhibit a strong bias to large CFG scales in a wide range, even if generation quality starts to degrade due to too strong guidance. Table 3: The winning rate and its degradation of different methods on various datasets. Model: Stable Diffusion-3.5 (ω=4.5ω=4.5). Method HPS v2 ImageReward AES PickScore CLIPScore Average η ωeω^e Pick-a-Pic (ηCFG/ηe-CFG/Δη^CFG/η^e-CFG/ η) (ηCFGη^CFG/ηe-CFGη^e-CFG) PAG 53% / 53% / 0% 44% / 43% / 1% 53% / 57% / -4% 36% / 43% / -7% 44% / 41% / 3% 46% / 47% !20 5.33 TDG 61% / 48% / 13% 50% / 51% / -1% 45% / 51% / -6% 38% / 47% / -9% 53% / 47% / 6% 49% / 49% !20 6.62 Z-Sampling 40% / 36% / 4% 41% / 54% /-13% 46% / 49% / -3% 31% / 37% / -6% 53% / 58% / -5% 42% / 47% !20 9.57 DrawBench (ηCFG/ηe-CFG/Δη^CFG/η^e-CFG/ η) (ηCFGη^CFG/ηe-CFGη^e-CFG) PAG 51% / 50% / 2% 44% / 40% / 3% 61% / 64% / -3% 42% / 40% / 2% 39% / 42% / -3% 48% / 47% !20 5.41 TDG 57% / 46% / 12% 54% / 45% / 9% 47% / 48% / -1% 41% / 44% / -4% 48% / 49% / -1% 50% / 46% !20 6.59 Z-Sampling 50% / 41% / 9% 50% / 45% / 4% 51% / 47% / 4% 34% / 44% / -9% 46% / 48% / -1% 46% / 45% !20 9.34 Table 4: The winning rates and the degradation of different methods on various datasets. Model: Stable Diffusion-2.1 (ω=5.5ω=5.5). Method HPS v2 ImageReward AES PickScore CLIPScore Average η ωeω^e Pick-a-Pic (ηCFG/ηe-CFG/Δη^CFG/η^e-CFG/ η) (ηCFGη^CFG/ηe-CFGη^e-CFG) Z-Sampling 90% / 60% / 30% 80% / 62% / 18% 70% / 63% / 7% 84% / 73% / 11% 71% / 46% / 25% 79% / 61% !20 11.80 CFG++ 82% / 62% / 20% 73% / 58% / 15% 58% / 55% / 3% 65% / 49% / 16% 57% / 55% / 2% 67% / 56% !20 9.09 FreeU 64% / 61% / 3% 73% / 67% / 6% 63% / 57% / 6% 52% / 45% / 7% 50% / 47% / 3% 60% / 55% !20 6.57 PAG 48% / 45% / 3% 49% / 41% / 8% 54% / 55% / -1% 49% / 45% / 4% 46% / 44% / 2% 49% / 46% !20 5.62 APG 11% / 31% /-20% 21% / 43% /-22% 49% / 53% / -4% 26% / 34% / -8% 45% / 58% /-13% 30% / 44% !20 11.19 SAG 42% / 28% / 14% 52% / 37% / 15% 47% / 46% / 1% 30% / 23% / 7% 56% / 39% / 17% 45% / 35% !20 6.53 DrawBench (ηCFG/ηe-CFG/Δη^CFG/η^e-CFG/ η) (ηCFGη^CFG/ηe-CFGη^e-CFG) FreeU 76% / 68% / 8% 68% / 64% / 5% 67% / 65% / 2% 57% / 54% / 3% 50% / 48% / 2% 64% / 60% !20 6.65 Z-Sampling 88% / 62% / 25% 73% / 65% / 8% 66% / 52% / 14% 74% / 66% / 8% 65% / 52% / 13% 73% / 59% !20 11.74 CFG++ 80% / 62% / 17% 70% / 57% / 12% 67% / 66% / 2% 58% / 48% / 11% 60% / 56% / 4% 67% / 58% !20 9.08 SAG 42% / 30% / 11% 54% / 42% / 11% 62% / 66% / -4% 38% / 38% / 1% 55% / 48% / 7% 50% / 45% !20 6.45 PAG 46% / 42% / 4% 42% / 42% / 1% 50% / 45% / 4% 46% / 48% / -1% 42% / 43% / -2% 45% / 44% !20 5.63 APG 16% / 29% /-13% 30% / 42% /-12% 38% / 48% /-10% 28% / 40% /-12% 38% / 44% / -6% 30% / 41% !20 10.48 Table 5: Quantitative results of different methods. Model: Stable Diffusion-XL (ω=5.5ω=5.5). Dataset Method HPS v2 (↑ ) ImageReward (↑ ) AES (↑ ) PickScore (↑ ) CLIPScore (↑ ) Pick-a-Pic CFG 28.51 0.845 6.112 22.50 31.01 APG 28.70 0.937 5.995 22.53 32.28 e-CFG 29.30 1.032 5.992 22.61 32.23 CFG++ 29.21 1.025 6.091 22.60 31.46 e-CFG 29.05 0.990 6.103 22.64 31.39 FreeU 28.88 0.929 6.156 22.11 30.01 e-CFG 28.90 0.955 6.115 22.63 31.23 PAG 28.92 0.863 6.174 22.47 30.54 e-CFG 28.68 0.884 6.126 22.56 31.09 SAG 28.87 0.884 6.087 22.55 30.97 e-CFG 28.70 0.931 6.022 22.55 31.26 SEG 28.96 0.875 6.178 22.35 30.47 e-CFG 28.70 0.893 6.123 22.56 31.12 TDG 28.91 0.972 6.092 22.61 31.19 e-CFG 28.96 0.955 6.104 22.61 31.33 Z-Sampling 29.31 1.037 6.118 22.67 31.30 e-CFG 28.80 0.909 6.003 22.40 31.56 DrawBench CFG 28.44 0.514 5.569 22.22 29.26 APG 28.67 0.672 5.576 22.31 33.27 e-CFG 29.23 0.770 5.609 22.36 33.37 CFG++ 29.14 0.742 5.622 22.34 29.71 e-CFG 29.05 0.718 5.600 22.40 29.73 FreeU 28.59 0.567 5.648 21.86 28.92 e-CFG 28.84 0.657 5.603 22.34 29.45 PAG 28.75 0.530 5.682 22.21 29.06 e-CFG 28.66 0.584 5.581 22.30 29.35 SAG 28.75 0.599 5.586 22.30 32.54 e-CFG 28.53 0.629 5.521 22.28 32.90 SEG 28.86 0.533 5.700 22.14 28.63 e-CFG 28.66 0.592 5.588 22.30 29.31 TDG 28.95 0.723 5.566 22.39 29.69 e-CFG 28.95 0.643 5.584 22.36 29.49 Z-Sampling 29.21 0.724 5.656 22.37 29.70 e-CFG 28.64 0.556 5.538 22.13 29.55 HPD CFG 28.51 0.845 6.112 22.50 31.01 APG 28.70 0.937 5.995 22.53 32.28 e-CFG 29.30 1.032 5.992 22.61 32.23 CFG++ 29.21 1.025 6.091 22.60 31.46 e-CFG 29.05 0.990 6.103 22.64 31.39 FreeU 28.88 0.929 6.156 22.11 30.01 e-CFG 28.90 0.955 6.115 22.63 31.23 PAG 28.92 0.863 6.174 22.47 30.54 e-CFG 28.68 0.884 6.126 22.56 31.09 SAG 28.87 0.884 6.087 22.55 30.97 e-CFG 28.70 0.931 6.022 22.55 31.26 SEG 28.96 0.875 6.178 22.35 30.47 e-CFG 28.70 0.893 6.123 22.56 31.12 TDG 28.91 0.972 6.092 22.61 31.19 e-CFG 28.96 0.955 6.104 22.61 31.33 Z-Sampling 29.31 1.037 6.118 22.67 31.30 e-CFG 28.80 0.909 6.003 22.40 31.56 Table 6: Quantitative results of different methods. Model: Stable Diffusion-2.1 (ω=5.5ω=5.5). Dataset Method HPS v2 (↑ ) ImageReward (↑ ) AES (↑ ) PickScore (↑ ) CLIPScore (↑ ) Pick-a-Pic CFG 26.25 -0.353 5.658 20.14 27.08 APG 26.03 -0.441 5.675 20.05 31.97 e-CFG 26.00 -0.420 5.537 20.17 30.29 CFG++ 27.13 0.045 5.751 20.43 27.72 e-CFG 26.99 -0.030 5.731 20.45 27.49 FreeU 26.93 -0.004 5.759 20.34 27.08 e-CFG 26.58 -0.251 5.706 20.30 27.35 PAG 26.46 -0.274 5.679 20.23 27.24 e-CFG 26.51 -0.258 5.672 20.24 27.46 SAG 26.18 -0.330 5.653 19.96 31.66 e-CFG 26.53 -0.240 5.704 20.25 32.29 Z-Sampling 27.09 0.009 5.762 20.53 27.88 e-CFG 26.66 -0.124 5.684 20.20 28.01 DrawBench CFG 27.09 -0.218 5.373 21.13 27.83 APG 26.78 -0.304 5.366 21.02 31.08 e-CFG 27.06 -0.181 5.300 21.11 30.64 CFG++ 27.96 0.116 5.474 21.35 28.68 e-CFG 27.80 0.065 5.430 21.33 28.47 FreeU 27.88 0.155 5.512 21.36 28.23 e-CFG 27.55 -0.079 5.405 21.29 28.04 PAG 27.28 -0.150 5.378 21.18 27.85 e-CFG 27.26 -0.155 5.387 21.19 27.99 SAG 26.92 -0.213 5.446 21.00 31.27 e-CFG 27.17 -0.188 5.361 21.15 31.30 Z-Sampling 27.83 0.082 5.467 21.42 28.62 e-CFG 26.96 -0.288 5.377 21.04 27.98 Table 7: Quantitative results of different methods. Model: Stable Diffusion-3.5 (ω=4.5ω=4.5). Dataset Method HPS v2 (↑ ) ImageReward (↑ ) AES (↑ ) PickScore (↑ ) CLIPScore (↑ ) Pick-a-Pic CFG 28.43 1.004 5.921 21.94 28.72 PAG 28.49 0.918 5.994 21.77 28.19 e-CFG 28.54 1.053 5.925 21.96 28.92 TDG 28.56 0.951 5.890 21.82 28.57 e-CFG 28.56 1.016 5.921 21.87 28.83 Z-Sampling 28.30 0.980 5.853 21.63 28.74 e-CFG 28.53 0.990 5.881 21.71 28.38 DrawBench CFG 29.23 0.962 5.429 22.72 30.39 PAG 29.25 0.894 5.531 22.55 29.70 e-CFG 29.42 1.031 5.429 22.77 30.62 TDG 29.42 0.996 5.435 22.66 30.50 e-CFG 29.47 1.005 5.446 22.71 30.52 Z-Sampling 29.32 0.974 5.434 22.61 30.17 e-CFG 29.50 1.040 5.420 22.68 30.27 Table 8: Quantitative results of different methods on COCO-30K. Model: Stable Diffusion-XL (ω=5.5ω=5.5). Method FID (↓ ) HPS v2 (↑ ) AES (↑ ) ωeω^e CFG 13.56 28.73 5.558 5.5 Z-Sampling 17.32 29.56 5.666 10.81 e-CFG 14.21 29.22 5.550 CFG++ 17.27 29.46 5.643 8.99 e-CFG 16.07 29.28 5.609 Table 9: Quantitative results of different methods on ImageNet-50K. Model: DiT-XL/2 (ω=4ω=4). Method FID (↓ ) IS (↑ ) AES (↑ ) ωeω^e CFG 21.22 474.9 4.911 4 Z-Sampling 25.43 471.7 4.853 9.17 e-CFG 24.04 465.1 4.834 CFG++ 24.33 480.1 4.950 8.01 e-CFG 25.55 480.0 4.791 Table 10: The winning rate and its degradation of different methods on Pick-a-Pic datasets. Model: Stable Diffusion-XL (ω=5.5ω=5.5). Method HPS v3 OneIG Pick-a-Pic (ηCFG/ηe-CFG/Δη^CFG/η^e-CFG/ η) PAG 71% / 68% / 3% 44% / 44% / 0% CFG++ 55% / 54% / 1% 36% / 26% / 10% Z-Sampling 79% / 67% / 12% 54% / 55% / -1% APG 44% / 39% / 5% 24% / 26% / -2% FreeU 39% / 35% / 4% 26% / 21% / 5% TDG 59% / 49% / 10% 41% / 32% / 9% SAG 66% / 56% / 10% 45% / 30% / 15% SEG 62% / 56% / 6% 23% / 25% / -2% Table 11: Quantitative results of different methods on Pick-a-Pic dataset. Model: Stable Diffusion-XL (ω=5.5ω=5.5). Dataset Model HPS v3 (↑ ) UnifiedReward (to CFG) UnifiedReward (to e-CFG) Pick-a-Pic CFG 6.837 / / PAG 7.781 0.16 0.16 e-CFG 7.341 / / CFG++ 7.648 0.1 -0.02 e-CFG 7.612 / / Z-Sampling 7.776 0.38 0.4 e-CFG 7.203 / / APG 7.263 -0.24 -0.19 e-CFG 7.615 / / FreeU 7.038 -0.25 -0.32 e-CFG 7.660 / / TDG 7.493 0.14 -0.03 e-CFG 7.461 / / SAG 7.432 0.25 -0.03 e-CFG 7.231 / / SEG 7.534 -0.18 -0.14 e-CFG 7.359 / / Table 12: Quantitative results of SD-XL on OneIG-Bench with different guidance scales. Guidance Scale Alignment (↑ ) Diversity (↑ ) Reasoning (↑ ) Style (↑ ) Text (↑ ) 5.5 .5638 .4399 .1932 .3325 .0128 6.0 .5738 .4357 .1910 .3328 .0178 6.5 .5771 .4294 .1970 .3374 .0170 7.0 .5820 .4242 .2000 .3420 .0164 7.5 .5837 .4225 .2076 .3363 .0181 8.0 .5936 .4169 .2048 .3366 .0166 8.5 .5917 .4145 .2086 .3358 .0182 9.0 .5944 .4109 .2133 .3411 .0267 9.5 .5934 .4085 .2154 .3387 .0194 10.0 .5936 .4043 .2226 .3377 .0259 (a) Scatter map of image saturation with HPSv2 (b) Scatter map of image saturation with ImageReward Figure 12: Reward evaluation on images with different saturation, each color represents images with different saturation. 20 images were randomly picked in MS-COCO dataset with changed saturation. Model: SD-XL. (a) Scatter map of image saturation with HPSv2 (b) Scatter map of image saturation with ImageReward Figure 13: Reward evaluation on images with different saturation, each color represents images generated by a prompt. 20 prompts were randomly selected in Pick-a-Pic, DrawBench, and HPD, and corresponding images were generated with different guidance scale. Model: SD-XL. (a) ωteω^e_t over timestep t. (b) ωteω^e_t over timestep t with different scheduler. Figure 14: Effective guidance scale value over timestep. Model: SD-XL. Dataset: Pick-a-Pic. Note that except Z-Sampling, CFG++, and APG, which modified the noise scheduling process, other methods use EulerDiscreteScheduler by default. Figure 15: Qualitative Results of Z-Sampling. Figure 16: Qualitative Results of CFG++. Figure 17: Qualitative Results of SAG. Figure 18: Qualitative Results of SEG. Figure 19: Qualitative Results of PAG. Figure 20: Qualitative Results of FreeU. Figure 21: Qualitative Results of TDG. Figure 22: Qualitative Results of APG.