1. Bibliographic Information

• Title: Pixel-Perfect Depth with Semantics-Prompted Diffusion Transformers
• Authors: Gangwei Xu, Haotong Lin, Hongcheng Luo, Xianqi Wang, Jingfeng Yao, Lianghui Zhu, Yuechuan Pu, Cheng Chi, Haiyang Sun, Bing Wang, Guang Chen, Hangjun Ye, Sida Peng, Xin Yang.
• Affiliations: The authors are from Huazhong University of Science and Technology, Xiaomi EV, and Zhejiang University.
• Journal/Conference: The paper is available on arXiv, an open-access repository for preprints.
• Publication Year: The arXiv identifier 2510.07316v1 suggests a submission date in October 2025. Given the current date is October 11, 2025, this is a very recent paper.
• Abstract: The paper introduces "Pixel-Perfect Depth," a monocular depth estimation model designed to eliminate "flying pixels"—artifacts commonly found at object edges in point clouds generated from depth maps. Unlike existing generative models that fine-tune Stable Diffusion and rely on a Variational Autoencoder (VAE) which causes these artifacts, this model operates directly in the pixel space. To manage the high complexity of pixel-space diffusion, the authors propose two innovations: 1) Semantics-Prompted Diffusion Transformers (SP-DiT), which use semantic features from vision foundation models to guide the diffusion process, ensuring global consistency and fine detail. 2) A Cascade DiT Design that progressively refines the depth map from coarse to fine for better efficiency and accuracy. The model achieves state-of-the-art performance among generative models on five benchmarks and excels in a new edge-aware point cloud evaluation.
• Original Source Link: The paper is available as a preprint at https://arxiv.org/pdf/2510.07316v1.

2. Executive Summary

• Background & Motivation (Why):

  • Core Problem: State-of-the-art monocular depth estimation models produce depth maps that, when converted to 3D point clouds, suffer from flying pixels. These are stray points, typically found around object boundaries, that float between the foreground and background, corrupting the 3D geometry.
  • Importance & Gaps: Flying pixels are a major obstacle for practical applications like 3D reconstruction, robotics, and augmented reality, which require clean and accurate geometry. The paper identifies two root causes:
    1. Discriminative Models: These models, trained with regression losses, tend to predict an "average" depth at sharp edges, smoothing over discontinuities and creating flying pixels.
    2. Generative Models: While better at preserving sharp edges in theory, current models are based on latent diffusion (e.g., Stable Diffusion). They use a VAE to compress the depth map into a low-dimensional latent space. This lossy compression inevitably introduces artifacts and blurs details, leading to flying pixels upon reconstruction.
  • Innovation: The paper proposes a radical departure from the latent diffusion paradigm. By performing the diffusion process directly in the pixel space, it completely bypasses the VAE, thereby eliminating its associated artifacts as the source of flying pixels. The core challenge then becomes making this high-dimensional generation process computationally tractable and accurate.

• Main Contributions / Findings (What):

  1. Pixel-Perfect Depth Framework: A novel monocular depth estimation model based on pixel-space diffusion that produces high-quality, flying-pixel-free point clouds.
  2. Semantics-Prompted Diffusion Transformer (SP-DiT): A new technique to stabilize and guide high-resolution pixel-space diffusion. It injects high-level semantic features from pre-trained Vision Foundation Models (VFMs) into the diffusion transformer, helping it maintain global structural consistency while generating fine-grained details.
  3. Cascade DiT Design (Cas-DiT): An efficient, coarse-to-fine architecture. It uses fewer tokens (larger patches) in early stages to model global layout and more tokens (smaller patches) in later stages to refine details, improving both speed and accuracy.
  4. Edge-Aware Point Cloud Evaluation: A new evaluation protocol to specifically measure the quality of point clouds at object edges, providing a quantitative way to assess the flying pixel problem. The proposed model significantly outperforms all competitors on this metric.

3. Prerequisite Knowledge & Related Work

• Foundational Concepts:

  • Monocular Depth Estimation (MDE): The task of predicting the depth (distance from the camera) for every pixel in a single 2D image. It's an ill-posed problem, as a single 2D image can correspond to infinitely many 3D scenes.
  • Flying Pixels: Erroneous 3D points that appear detached from surfaces, typically at the boundary between a foreground object and the background. They arise from inaccurate depth values at these discontinuities.
  • Discriminative vs. Generative Models:
    • Discriminative Models: Learn a direct mapping from an input (image) to an output (depth map). They are often trained to minimize a regression loss like Mean Squared Error, which can lead to blurry or averaged predictions at sharp edges. Examples: Depth Anything v2, DPT.
    • Generative Models: Learn the underlying probability distribution of the data (depth maps). They can generate new samples from this distribution. Diffusion models are a prominent type. They are theoretically better at capturing complex distributions, like the bimodal depth values at an object edge.
  • Diffusion Models: A class of generative models that work in two steps. First, a "forward process" gradually adds Gaussian noise to a clean data sample until it becomes pure noise. Second, a "reverse process" trains a neural network (often a U-Net or, in this case, a Transformer) to denoise the sample step-by-step, starting from pure noise and ending with a clean sample.
  • Variational Autoencoder (VAE): A neural network architecture with an encoder that compresses input data into a compact latent space and a decoder that reconstructs the data from this latent representation. This compression is lossy, meaning some information is lost. Stable Diffusion and its derivatives use a VAE to make diffusion computationally cheaper by working in the smaller latent space.
  • Diffusion Transformer (DiT): A specific type of diffusion model where the denoising network is a Transformer, which has shown excellent scalability and performance in generative modeling.
  • Vision Foundation Models (VFMs): Large-scale models (e.g., DINOv2, MAE) pre-trained on enormous image datasets. Their encoders are powerful feature extractors that capture rich semantic and structural information from images.

• Previous Works & Differentiation:

  • Prior discriminative models like Depth Anything v2 [76] and Depth Pro [4] achieve strong performance but are prone to smoothing edges due to their regression-based nature.
  • Prior generative models like Marigold [31] and GeoWizard [15] leverage the power of pre-trained Stable Diffusion. However, their reliance on a VAE to work in latent space is their Achilles' heel, as VAE reconstruction is a primary source of flying pixels. Figure 2 in the paper compellingly shows that even a ground-truth depth map (GT(VAE)) becomes corrupted with flying pixels after a round trip through the VAE.
  • This paper's key differentiation is its move to pixel-space diffusion. It avoids the VAE entirely. To overcome the known challenges of pixel-space generation (high computational cost and difficulty modeling global structures), it introduces SP-DiT and Cas-DiT. Instead of relying on the generative priors of Stable Diffusion, it uses semantic priors from VFMs to guide the process.

4. Methodology (Core Technology & Implementation)

The proposed model, Pixel-Perfect Depth, is a diffusion model that generates a depth map directly in pixel space, conditioned on an input RGB image.

该图像为模型结构示意图，展示了论文中Pixel-Perfect Depth的工作流程。输入图片经过视觉基础模型提取语义特征和噪声叠加后的深度图拼接输入，进入级联DiT模块中，先通过标准DiT块处理，再结合语义提示的DiT块处理，最后输出深度预测图。图中还标注了真实深度图仅用于训练阶段。

Figure 1: This diagram illustrates the architecture of Pixel-Perfect Depth. An input image is processed in two ways: (1) It's fed into a Vision Foundation Model (e.g., DINOv2) to extract high-level semantic features, which are then normalized. (2) It's concatenated with a noisy version of the target depth map (during training) or pure noise (during inference). This combined input enters the Cascade DiT Design. The initial DiT Blocks process a coarse representation, which is then refined by the Semantics-Prompted DiT Blocks that are guided by the normalized semantic features. The final output is the predicted depth map. The Ground Truth (GT) depth is only used for training supervision.

• Principles: The core idea is that modeling depth distributions directly at the pixel level can capture sharp discontinuities without the artifacts introduced by VAEs. The main challenge is guiding this high-dimensional generative process to be both globally coherent (understanding the overall scene layout) and locally precise (generating fine details). The solution is to use explicit semantic guidance from powerful VFMs.

• Steps & Procedures:

  1. Generative Formulation (Flow Matching) The model uses Flow Matching, an efficient alternative to traditional diffusion models. It learns a velocity field that transforms a simple noise distribution into the target data distribution.

  • An interpolated sample $\mathbf{x}_t$ is defined for any time $t \in [0, 1]$ between a clean depth sample $\mathbf{x}_0$ and Gaussian noise $\mathbf{x}_1$: $$\mathbf { x } _ { t } = t \cdot \mathbf { x } _ { 1 } + ( 1 - t ) \cdot \mathbf { x } _ { 0 } .$$ This creates a straight path from the data to the noise.
  • The velocity field $\mathbf{v}_t$ along this path is constant: $${ \bf v } _ { t } = \frac { d { \bf x } _ { t } } { d t } = { \bf x } _ { 1 } - { \bf x } _ { 0 } ,$$ This vector points from the clean depth map to the noise map.
  • The model, a network $\mathbf{v}_{\theta}$, is trained to predict this velocity field given the noisy sample $\mathbf{x}_t$, the time step $t$, and the conditioning image $\mathbf{c}$. The training loss is the Mean Squared Error (MSE): $$\mathcal { L } _ { \mathrm { v e l o c i t y } ( \theta ) } = \mathbb { E } _ { \mathbf { x } _ { 0 } , \mathbf { x } _ { 1 } , t } \left[ \left\| \mathbf { v } _ { \theta } ( \mathbf { x } _ { t } , t , \mathbf { c } ) - \mathbf { v } _ { t } \right\| ^ { 2 } \right] .$$
  • At inference, the model starts with pure noise $\mathbf{x}_1$ and iteratively updates it by following the predicted velocity field backward in time from $t=1$ to $t=0$, using an ODE solver: $$\mathbf { x } _ { t _ { i - 1 } } = \mathbf { x } _ { t _ { i } } + \mathbf { v } _ { \theta } ( \mathbf { x } _ { t _ { i } } , t _ { i } , \mathbf { c } ) ( t _ { i - 1 } - t _ { i } ) ,$$ This process gradually transforms the noise into a coherent depth map.

  2. Semantics-Prompted Diffusion Transformers (SP-DiT) This is the core innovation to make pixel-space diffusion work. A standard DiT struggles to model both global structure and fine details simultaneously. SP-DiT enhances the DiT with semantic guidance.

  • Problem: Without guidance, the model produces poor results (see Figure 6), failing to capture the scene's overall structure.
  • Solution: Use a pre-trained VFM encoder $f$ (e.g., the ViT from DINOv2) to extract a sequence of semantic tokens $\mathbf{e}$ from the input image $\mathbf{c}$: $$\mathbf { e } = f ( \mathbf { c } ) \in \mathbb { R } ^ { T ^ { \prime } \times D ^ { \prime } } ,$$
  • Normalization: The authors found that the magnitude of these semantic features $\mathbf{e}$ was mismatched with the internal representations of their DiT, causing training instability. They solve this with a simple but effective L2 normalization along the feature dimension: $${ \hat { \mathbf { e } } } = { \frac { \mathbf { e } } { \| \mathbf { e } \| _ { 2 } } } .$$
  • Integration: The normalized semantic features $\hat{\mathbf{e}}$ are bilinearly upsampled to match the spatial resolution of the DiT's internal tokens $\mathbf{z}$. They are then concatenated with $\mathbf{z}$ and passed through a multilayer perceptron (MLP) $h_{\phi}$ to produce the "prompted" tokens $\mathbf{z}'$: $$\mathbf { z } ^ { \prime } = h _ { \phi } ( \mathbf { z } \oplus B ( \hat { \mathbf { e } } ) ) ,$$ These prompted tokens are then processed by the subsequent DiT blocks, which are now conditioned on high-level semantic information, enabling them to generate globally consistent and detailed outputs.

  3. Cascade DiT Design (Cas-DiT) This is an architectural optimization for efficiency and accuracy, based on the observation that different layers of a vision transformer learn features at different scales.

  • Coarse Stage: The first 12 DiT blocks operate on a smaller number of tokens, created using a large patch size (e.g., 16x16). This encourages the model to focus on global layout and low-frequency information, reducing computational cost.
  • Fine Stage: After the 12th block, an MLP upsamples the feature representation, increasing the number of tokens (equivalent to using a smaller 8x8 patch size). These tokens are fed into the remaining 12 SP-DiT blocks. This fine stage, now prompted by semantics, focuses on refining high-frequency details. This coarse-to-fine progression improves both speed and performance.

5. Experimental Setup

• Datasets:

  • Training: The models were trained exclusively on synthetic datasets to leverage their high-quality, artifact-free ground-truth geometry.
    • 512x512 model: Trained on Hypersim (~54K images), a photorealistic indoor scene dataset.
    • 1024x768 model: Trained on a larger mix including Hypersim, UrbanSyn, UnrealStere04K, VKITTI, and TartanAir.
  • Evaluation: The model's zero-shot generalization capability was tested on five diverse real-world benchmarks:
    • NYUv2 (indoor scenes)
    • KITTI (outdoor driving scenes)
    • ETH3D (high-resolution indoor/outdoor)
    • ScanNet (indoor scenes)
    • DIODE (diverse indoor and outdoor)

• Evaluation Metrics:

  1. Absolute Relative Error (AbsRel): Measures the average relative error between the predicted depth $d_y$ and ground truth depth $d_y^*$. Lower is better.

     • Conceptual Definition: It indicates, on average, what percentage of the true depth the prediction is off by. It is scale-invariant after alignment.
     • Formula: $\text{AbsRel} = \frac{1}{|P|} \sum_{y \in P} \frac{|d_y - d_y^*|}{d_y^*}$
     • Symbol Explanation: $P$ is the set of all valid pixels, $d_y$ is the predicted depth at pixel $y$, and $d_y^*$ is the ground truth depth at pixel $y$.

  2. $\delta_1$ Accuracy: Measures the percentage of pixels where the ratio between the predicted and ground truth depth is within a certain threshold (1.25). Higher is better.

     • Conceptual Definition: It represents the fraction of pixels that are "accurately" predicted.
     • Formula: $$\delta_1 = \% \text{ of pixels } y \text{ such that } \max\left(\frac{d_y}{d_y^*}, \frac{d_y^*}{d_y}\right) < 1.25$$
     • Symbol Explanation: Same as above.

  3. Edge-Aware Chamfer Distance: This is a novel metric proposed by the authors to specifically evaluate flying pixels.

     • Conceptual Definition: It quantifies the 3D point cloud error specifically at object boundaries. First, edges are detected in the ground truth depth map using a Canny edge detector. Then, the Chamfer Distance is computed between the predicted and ground truth point clouds only within a narrow band around these edges. A lower value means fewer flying pixels and more accurate edges.
     • Formula (Chamfer Distance): $$CD(S_1, S_2) = \sum_{x \in S_1} \min_{y \in S_2} \|x-y\|_2^2 + \sum_{y \in S_2} \min_{x \in S_1} \|x-y\|_2^2$$
     • Symbol Explanation: $S_1$ and $S_2$ are the two point clouds being compared (in this case, the subsets of points near edges). The metric finds the average squared distance from each point in one set to its nearest neighbor in the other set.

• Baselines: The paper compares against a wide range of state-of-the-art models, including:

  • Discriminative Models: MiDaS, DPT, Depth Anything v2, Depth Pro.
  • Generative (Latent Diffusion) Models: Marigold, GeoWizard, DepthFM, Lotus.

6. Results & Analysis

• Core Results (Zero-Shot Relative Depth):

  Note: This table is a transcription of the original data from Table 1, not the original image.

  Type	Method	Training Data	NYUv2		KITTI		ETH3D		ScanNet		DIODE
  			AbsRel↓	δ1↑	AbsRel↓	δ1↑	AbsRel↓	δ1↑	AbsRel↓	δ1↑	AbsRel↓	δ1↑
  Discriminative	DiverseDepth[81]	320K	11.7	87.5	19.0	70.4	22.8	69.4	10.9	88.2	-	-
  	MiDaS[46]	2M	11.1	88.5	23.6	63.0	18.4	75.2	12.1	84.6	-	-
  	LeReS[83]	354K	9.0	91.6	14.9	78.4	17.1	77.7	9.1	91.7	-	-
  	Omnidata[11]	12M	7.4	94.5	14.9	83.5	16.6	77.8	7.5	93.6	-	-
  	HDN[85]	300K	6.9	94.8	11.5	86.7	12.1	83.3	8.0	93.9	-	-
  	DPT[45]	1.2M	9.8	90.3	10.0	90.1	7.8	94.6	8.2	93.4	-	-
  	DepthAny. v2[76]	54K	5.4	97.2	8.6	92.8	12.3	88.4	-	-	8.8	93.7
  	DepthAny. v2[76]	62M	4.5	97.9	7.4	94.6	13.1	86.5	6.5	97.2	6.6	95.2
  	Marigold[31]	74K	5.5	96.4	9.9	91.6	6.5	96.0	6.4	95.1	10.0	90.7
  Generative	GeoWizard[15]	280K	5.2	96.6	9.7	92.1	6.4	96.1	6.1	95.3	12.0	89.8
  	DepthFM[18]	74K	5.5	96.3	8.9	91.3	5.8	96.2	6.3	95.4	-	-
  	GenPercept[72]	90K	5.2	96.6	9.4	92.3	6.6	95.7	5.6	96.5	-	-
  	Lotus[20]	54K	5.4	96.8	8.5	92.2	5.9	97.0	5.9	95.7	9.8	92.4
  	Ours (512)	54K	4.3	97.4	8.0	93.1	4.5	97.7	4.5	97.3	7.0	95.5
  	Ours (1024)	125K	4.1	97.7	7.0	95.5	4.3	98.0	4.6	97.2	6.8	95.9

  • Analysis: The proposed model, Ours (1024), achieves the best performance among all published generative models across all five benchmarks. It is also highly competitive with the best discriminative model (Depth Anything v2 trained on 62M images), even outperforming it on KITTI and ETH3D despite being trained on far less data (125K). This demonstrates excellent zero-shot generalization from synthetic to real-world data.

• Qualitative Results:

  该图像为插图，展示了不同方法在单目深度估计及点云重建上的对比。每组第一列为输入图像，后续列展示各方法生成的深度图（下方小图）和对应的点云渲染。可以看出“ Ours”方法生成的点云更加完整，边缘细节更清晰，且飞行像素明显更少，体现了本文提出的像素级扩散生成优势。

  Figure 2: This figure provides a direct visual comparison of point clouds. "Ours" produces a clean, complete point cloud, faithfully representing the scene's geometry. In contrast, both the discriminative (Depth Anything v2) and generative (Marigold) baselines show significant flying pixels and noisy artifacts around object boundaries.

  该图像为对比图，共五行五列。每行第一列为输入彩色图像，后续四列展示五种单目深度估计模型生成的深度图，分别是Depth Anything v2、Depth Pro、MoGe 2和本文提出的方法（Ours）。从视觉上看，本文方法的深度图在边缘和细节处更为清晰准确，无飞行像素，深度变化平滑且语义一致性更好。

  Figure 4: In these open-world scenes, the proposed model ("Ours") generates depth maps with remarkably sharp edges and fine details (e.g., the bridge railings, tree branches, building structures). It appears more robust than Depth Pro, which struggles in some areas, and preserves more detail than Depth Anything v2 and MoGe 2.

• Ablations and Analysis:

  Note: This table is a transcription of the original data from Table 2, not the original image.

  Method	NYUv2		KITTI		ETH3D		ScanNet		DIODE		Time(s)
  	AbsRel↓	δ1↑	AbsRel↓	δ1↑	AbsRel↓	δ1↑	AbsRel↓	δ1↑	AbsRel↓	δ1↑
  DiT (baseline)	22.5	72.8	27.3	63.9	12.1	87.4	25.7	65.1	23.9	76.5	0.19
  SP-DiT	4.8	96.7	8.6	92.2	4.6	97.5	6.2	94.8	8.2	94.1	0.20
  SP-DiT+Cas-DiT	4.3	97.4	8.0	93.1	4.5	97.7	4.5	97.3	7.0	95.5	0.14

  • Analysis: This ablation study is crucial. The DiT (baseline) without any proposed modules performs extremely poorly, confirming that naive pixel-space diffusion is ineffective. Adding SP-DiT leads to a massive performance improvement (e.g., AbsRel on NYUv2 drops from 22.5 to 4.8, a 78% improvement), proving that semantic prompting is the key enabler. Further adding Cas-DiT provides an additional accuracy boost while also reducing inference time by ~30% (from 0.20s to 0.14s), demonstrating its effectiveness for both accuracy and efficiency.

    该图像为示意图，展示了不同数据集（NYUv2、KITTI、ETH3D、ScanNet、DIODE）下的单目深度估计结果对比。第一行为原始图像，第二行为未使用语义提示扩散变换器（w/o SP-DiT）的深度估计结果，第三行为使用语义提示扩散变换器（w/ SP-DiT）后的深度估计，后者在边缘细节和深度连续性上表现更佳，更准确地反映了场景深度变化。

  Figure 6: This figure visually demonstrates the impact of SP-DiT. The top row (w/o SP-DiT) shows blurry, incoherent depth maps that fail to capture basic scene geometry. The bottom row (w/ SP-DiT) shows sharp, globally consistent depth maps, aligning perfectly with the quantitative results in Table 2.

• Edge-Aware Point Cloud Evaluation:

  Note: This table is a transcription of the original data from Table 4, not the original image.

  	Marigold[31]	GeoWizard[15]	DepthAny. v2[76]	DepthPro[4]	GT(VAE)	Ours
  Chamfer Dist.↓	0.17	0.16	0.18	0.14	0.12	0.08

  • Analysis: This result is arguably the paper's most compelling evidence. "Ours" achieves a Chamfer Distance of 0.08, significantly outperforming all other methods. Critically, the GT(VAE) experiment, where ground truth depth is passed through a VAE, results in a Chamfer Distance of 0.12. This shows that the VAE compression itself introduces more edge artifacts than the entire proposed pipeline, decisively proving the paper's central hypothesis that avoiding VAEs is key to eliminating flying pixels.

    该图像为多幅深度估计结果的插图，展示了从左到右依次为输入原图、Marigold方法预测的深度点云及其色彩编码、使用VAE的真实深度图（GT VAE）、本文提出方法（Ours）生成的深度图及其色彩编码，以及真实深度图（GT）。可见，本文方法较Marigold在细节和边缘处表现更为清晰和准确，且接近真实深度图。

  Figure 3: This figure visually supports Table 4. The point cloud from GT(VAE) clearly shows flying pixels and blurred edges, similar to Marigold. In contrast, "Ours" is much cleaner and closer to the ground truth ("GT"), demonstrating the benefit of the pixel-space approach.

7. Conclusion & Reflections

• Conclusion Summary: The paper successfully presents "Pixel-Perfect Depth," a novel model for monocular depth estimation that generates high-quality point clouds free from flying pixels. By moving diffusion from latent space to pixel space, it avoids the artifacts caused by VAE compression. The key technical contributions, Semantics-Prompted DiT (SP-DiT) and Cascade DiT Design (Cas-DiT), make this challenging approach feasible, enabling the model to achieve state-of-the-art results among generative models and to significantly outperform all competitors on a new, more revealing edge-aware evaluation metric.

• Limitations & Future Work: The authors acknowledge two main limitations:

  1. Lack of Temporal Consistency: When applied to video, the model processes each frame independently, which can lead to flickering in the resulting depth sequence.
  2. Inference Speed: The iterative nature of diffusion makes it slower than single-pass discriminative models like Depth Anything v2. Future work could explore video-based diffusion techniques to enforce temporal consistency and adopt DiT acceleration methods to improve inference speed.

• Personal Insights & Critique:

  • Novelty & Impact: The paper's core idea is both simple and powerful. It correctly identifies a fundamental flaw (VAE compression) in the dominant paradigm for generative depth estimation and proposes a direct solution. This work could inspire a new wave of research into pixel-space generative models for dense prediction tasks, shifting focus away from simply fine-tuning existing latent diffusion models. The SP-DiT concept of using VFMs for explicit semantic guidance is a very strong contribution that could be applied to other generative tasks.
  • Strengths: The experimental validation is exceptionally strong. The GT(VAE) comparison is a "smoking gun" experiment that makes their central claim undeniable. The introduction of the edge-aware metric is also a valuable contribution to the field, as it provides a better tool for evaluating a problem that many researchers have observed qualitatively but struggled to measure.
  • Potential Questions:
    • The model's performance is tied to the quality of the VFM used for semantic prompting. How much does the choice of VFM matter, and would a weaker VFM significantly degrade performance?
    • The model is trained entirely on synthetic data. While it shows impressive zero-shot results, there might be subtle real-world phenomena (e.g., specific sensor noise, atmospheric effects) that it doesn't capture.
    • While faster than a baseline DiT, the inference time of 140ms for the main model (Table 6) is still a barrier for real-time applications like robotics or autonomous driving, where discriminative models running at <20ms have a clear advantage. The PPD-Small variant at 40ms is a good step towards addressing this.