CSFlow: Aligning Flow Matching with Human Contrast Sensitivity
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: We're looking at the paper: "CSFlow: Aligning Flow Matching with Human Contrast Sensitivity".
Jane: Here's the gist of the paper. CSFlow introduces Contrast Sensitive Flow (CSFlow), a weighting scheme that connects the human eye’s Contrast Sensitivity Function (CSF) to the iterative denoising steps of flow matching, providing a weighting scheme grounded in human vision.
Tom: This method improves generative performance by lowering FID by 4.7%, increasing Inception Score by 2.2% and improving GenEval scores by 2.5% using inference-only timestep modification or short fine-tuning.
Jane: Core Concept of CSFlow The authors propose a weighting scheme that connects the human eye’s Contrast Sensitivity Function (CSF) to the iterative denoising steps of flow matching.
Tom: This weighting scheme is grounded in human vision.
Paper summary: Jane: So, to summarize this CSFlow paper, it introduces a weighting scheme that connects the human eye’s contrast sensitivity function to the iterative denoising steps of flow matching <ref:2606.08833#pg1>.
Tom: The thesis is that because real-world images concentrate signal at low spatial frequencies, this method allocates more capacity to frequencies that the human eye is more sensitive <ref:2606.08833#pg1>.
Lu: They define a metric that estimates which frequencies are generated at each reverse flow interval and then weight those obtained by aligning the frequencies generated at each noise level with human contrast sensitivity <ref:2606.08833#pg1>.
Meng: So they’re essentially telling the model to pay more attention to the things our eyes notice, which is a nice concept, but how does that translate into actual computational savings or stability?
Lalam: The paper analyzes how recoverable frequency content evolves over time during the denoising process by defining a fraction of expected spectral power in xt that comes from the image component <ref:2606.08833#pg1>.
Tom: They define that fraction as rsignal of f at time t, which is the expected power after cleaning divided by the noise level <ref:2606.08833#pg1>.
Jane: And then they calculate a stepwise increase in that signal, delta rsignal of f at time t plus delta t using that initial value <ref:2606.08833#pg1>.
Lu: Perceptual importance is quantified using the contrast sensitivity function CSF, defining the raw perceptual weight at time t as weCSFlow of t which is the average CSF value over frequencies that become newly recoverable around time t <ref:2606.08833#pg1>.
Meng: That sounds complex to implement on a production system. What's the final output they propose?
Lalam: They normalize those interval weights to have a mean of one over time, resulting in wCSFlow of t which is the raw perceptual weight divided by an integral over time <ref:2606.08833#pg1>.
Tom: That final weighting scheme, wCSFlow of t, can then be used in two ways: modifying the step size during inference to increase approximation accuracy for important regions, and weighting the importance of certain times during training <ref:2606.08833#pg1>.
Jane: So we get a way to guide the generation process by using human vision as our guide, which is a really interesting direction for generative modeling.
Lu: And when they applied it, they showed that for PixelGen and JiT models, these CSFlow modifications consistently improved visual quality by four point seven percent in FID, two point two percent in Inception Score and two point five percent in GenEval scores <ref:2606.08833#pg1>.
Conclusion: Tom: So looking at CSFlow by Galinska, Pogodzinski, and Lenssen, it’s essentially taking something deeply rooted in human perception—the CSF—and using it to guide how a flow matching model refines an image step by step <ref:2606.08833#pg1>.
Jane: The implication is that we can move beyond just optimizing for generic metrics and start optimizing for what actually makes an image look good to a human observer, which is pretty fundamental.
Lu: It means the AI isn't just learning noise reduction; it’s learning to prioritize generating frequencies that correspond to what our eyes are most sensitive to, which is different from just minimizing geometric errors <ref:2606.08833#pg1>.
Meng: I still wonder about the practical deployment side. The authors noted something important—that CSFlow does not impact the number of large geometric errors, suggesting that those specific types of errors aren't fixed by this weighting scheme <ref:2606.08833#pg1>.
Lalam: And they also point out a conceptual limitation: you need an assumption about a typical screen and screen distance, so if you want fully optimal weights, you might actually need to calibrate the system <ref:2606.08833#pg1>.
Tom: That calibration point is something we have to think about when we try to scale this up—if the weights depend on where you are viewing things, it adds a layer of complexity for deployment <ref:2606.08833#pg1>.
Jane: It shows that while the performance numbers are solid, the next step for this research is figuring out how to transfer these perceptual weighting schemes over to latent image generation spaces instead of just pixel space <ref:2606.08833#pg1>.
Malgorzata Galinska, Bart Pogodzinski, Jan Eric Lenssen
Max Planck Institute for Informatics, Saarland Informatics Campus
cs.CV
Submitted: 2026-06-07
Updated: 2026-11-09
Comments: Accepted at the 40th Annual Conference on Neural Information Processing Systems (NeurIPS 2026). Project page: https://gosiagal.github.io/CSFlow/ Code: https://github.com/gosiagal/CSFlow
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 90/100
The gist: CSFlow introduces Contrast Sensitive Flow (CSFlow), a weighting scheme that connects the human eye’s Contrast Sensitivity Function (CSF) to the iterative denoising steps of flow matching…
Key concepts
- Contrast Sensitivity Function (CSF)
- CSF describes how sensitive the human eye is to changes in contrast across different spatial frequencies. It quantifies which visual details are perceptible at various scales, allowing the model to prioritize generating features that match human vision.
- Reverse Flow Interval Analysis
- The authors analyze how recoverable frequency content evolves over time during denoising. They measure the fraction of expected spectral power in the noisy image that corresponds to the original clean image components at different time intervals.
- Perceptual Weight Construction
- This involves calculating a raw perceptual weight by multiplying the change in signal power by the CSF. This weight is then normalized to create interval weights, which determine how important each denoising step is based on its relevance to human perception.
Terminology
Summary
CSFlow introduces Contrast Sensitive Flow (CSFlow), a weighting scheme that connects the human eye’s Contrast Sensitivity Function (CSF) to the iterative denoising steps of flow matching, providing a weighting scheme grounded in human vision. p. 0 4
Invisible
Visible
Perception Weight
CSFlow: Perceptual Importance of Information Gain CSFlow Perceptual Importance Information Gain
t=0. 0Information Gain
Less Noise
Denoised Image
Figure 1: Contrast Sensitive Flow (CSFlow) connects the human eye’s Contrast Sensitivity Function
(CSF) [Barten, 2003] (left) to the iterative denoising steps of flow matching (right), providing a
weighting scheme grounded in human vision. We
for the first time merge these observations through two contributions: (1) a metric
that estimates which frequencies are generated at each reverse flow interval and
(2) timestep weights obtained by aligning the frequencies generated at each noise
level with human contrast sensitivity. We validate our contributions experimentally
showing that these weights can improve generative performance by lowering FID
by 4. We validate our contributions experimentally
showing that these weights can improve generative performance by lowering FID
by 4. 7%, increasing Inception Score by 2. 7%, increasing Inception Score by 2. 2% and improving GenEval scores by
2. 2% and improving GenEval scores by
2. 5% using inference-only timestep modification or short fine-tuning.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Core Concept of CSFlow
The authors propose a weighting scheme that connects the human eye’s Contrast Sensitivity Function (CSF) to the iterative denoising steps of flow matching. This weighting scheme is grounded in human vision. p. 0 4
Invisible
Visible
Perception Weight
CSFlow: Perceptual Importance of Information Gain CSFlow Perceptual Importance Information Gain
t=0. 0Information Gain
Less Noise
Denoised Image
Figure 1: Contrast Sensitive Flow (CSFlow) connects the human eye’s Contrast Sensitivity Function
(CSF) [Barten, 2003] (left) to the iterative denoising steps of flow matching (right), providing a
weighting scheme grounded in human vision. Meanwhile,
the human visual system is unequally sensitive to spatial frequencies: very low
and very high frequencies require significantly higher contrast to be perceived. We
for the first time merge these observations through two contributions: (1) a metric
that estimates which frequencies are generated at each reverse flow interval and
(2) timestep weights obtained by aligning the frequencies generated at each noise
level with human contrast sensitivity.Where the paper says it
Where the paper says it
Frequency Content Analysis
The paper analyzes how recoverable frequency content evolves over time during the denoising process in Sec. 4.1. They define the fraction of expected spectral power in xt that comes from the image component, denoted as rsignal(f, t). p. 3 For a frequency f, we denote the corresponding 2D DFT coefficient as Ff (·) , power as Pf (x) =
|Ff (x)|
2
, and define the expected clean image power Sf = E[Pf (xclean)] . The quantity we would like to measure is the fraction of expected spectral power in xt that comes
from the image component, i. The quantity we would like to measure is the fraction of expected spectral power in xt that comes
from the image component, i. , the fraction of retained image signal at denoising state t:
rsignal(f, t) = E [Pf (atxclean)]
E [Pf (xt)] . 2 Denoising step information gain from rsignal
While rsignal(f, t) describes how recoverable a frequency f is at a single denoising state, we are
interested in which frequencies become newly recoverable during each denoising interval. We
therefore define the stepwise increase
∆rsignal(f, t, ∆t) = rsignal(f, t + ∆t) − rsignal(f, t).Where the paper says it
Where the paper says it
Where the paper says it
Perceptual Weight Construction
The authors quantify perceptual importance using the contrast sensitivity function CSF. They define the raw perceptual weight at time t as weCSFlow(t) = R omega ∂trsignal(f, t) · CSF(f) df. p. 4 In continuous notation, this corresponds to the
instantaneous increase ∂trsignal(f, t). We define the raw perceptual weight at time t as the CSFweighted average of the newly recovered frequency content:
weCSFlow(t) =
R
Ω
∂trsignal(f, t) · CSF(f) df
R
Ω
∂trsignal(f, t) df . The denominator makes the score relative: intervals are
weighted according to which frequencies are recovered, rather than simply how much total spectral
content is recovered. Equivalently, weCSFlow(t) is the average CSF value over the frequencies that
become recoverable around time t. If CSF(f) ≡ 1, all times receive the same raw score. We normalize these scores over the denoising trajectory:
wCSFlow(t) = weCSFlow(t)
R 1
0 weCSFlow(τ ) dτ
.Where the paper says it
Where the paper says it
Where the paper says it
Application in Training and Inference
The derived weights can be applied in two ways: (1) to modify the step size of inference steps, increasing the approximation accuracy of important regions of the generation process; and (2) to weight the importance of certain t during model training. For training, timesteps are sampled from the CSF-weighted time distribution using inverse transform sampling. p. 2 , 2024] applies a related resolution-dependent timestep shift in latent-space rectified
flows, selecting the shift through inference-time sampling experiments and then using it during both
high-resolution training and inference. It also studies non-uniform training-time timestep sampling,
including logit-normal sampling, which emphasizes intermediate timesteps and vanishes near the
endpoints. , 2024] applies a related resolution-dependent timestep shift in latent-space rectified
flows, selecting the shift through inference-time sampling experiments and then using it during both
high-resolution training and inference. It also studies non-uniform training-time timestep sampling,
including logit-normal sampling, which emphasizes intermediate timesteps and vanishes near the
endpoints.Where the paper says it
Where the paper says it
Experimental Validation
Experiments on PixelGen and JiT models show that our CSFlow-modifications consistently improve visual quality by 4.7% in FID, 2. p. 0 We
for the first time merge these observations through two contributions: (1) a metric
that estimates which frequencies are generated at each reverse flow interval and
(2) timestep weights obtained by aligning the frequencies generated at each noise
level with human contrast sensitivity. We validate our contributions experimentally
showing that these weights can improve generative performance by lowering FID
by 4. We validate our contributions experimentally
showing that these weights can improve generative performance by lowering FID
by 4. 7%, increasing Inception Score by 2. 7%, increasing Inception Score by 2. 2% and improving GenEval scores by
2. Figure 5 shows chosen image pairs that compare the generations of the
baseline model and our best configuration model. We observe that our model generates more realistic images in terms of their textures: while the
baseline model suffers from oversmoothing objects and surfaces, we find that our weighting in
finetuning and inference fixes that issue, making
the images look less artificial and cartoonish.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Limitations
The authors find that CSFlow does not impact the number of large geometric errors, suggesting that the reasons for these types of errors are different from suboptimal weighting of timesteps. A conceptual limitation is the required assumption of a typical screen and screen distance, leading to the necessity of calibration if fully optimal weights are desired. p. 8 This suggests that the reasons for these types of errors
are different from suboptimal weighting of timesteps. A conceptual limitation of our weighting
derivation is the required assumption of a typical screen and screen distance, leading to the necessity
of calibration if fully optimal weights are desired. A conceptual limitation of our weighting
derivation is the required assumption of a typical screen and screen distance, leading to the necessity
of calibration if fully optimal weights are desired. A promising venture for the future would be to
analyze how image frequencies translate into common VAE latent spaces to transfer appropriate
weighting schemes to latent image generation.Where the paper says it
Where the paper says it
CSF Definition Details
The CSF is defined for spatial frequencies measured in cycles per degree of visual angle, using the Barten CSF formula. The parameters used are specified as Mopt(f) = exp−2π/2σ2f2 (10) and various constants like σ = 0.5/60, k = 3.0, T = 0.1. The conversion from cycles per pixel to cycles per degree is defined by fcpd = f / θpix where θpix is the visual angle subtended by one pixel. p. 4 We quantify perceptual importance
using the contrast sensitivity function CSF. Since CSF is defined over frequencies in cycles per
degree of visual angle, while our image frequencies are measured in cycles per pixel, we first convert
image frequency bins to cycles per degree before evaluating CSF.Where the paper says it
Proof of Theorem 1
The proof for Theorem 1 relies on the linearity of the DFT, showing that E [Pf (xt)] = a2/tSf + b2/t. This leads to E [Pf (atxclean)] / E [Pf (xt)] = a2/tSf. p. 3 Assume that ϵ ∼ N (0, 1) is independent of xclean and
Sf , at, and bt are defined as above. Then,
rsignal(f, t) = E [Pf (atxclean)]
E [Pf (xt)] =
a
2
tSf
a
2
tSf + b
2
t
. Assume that ϵ ∼ N (0, 1) is independent of xclean and
Sf , at, and bt are defined as above. Then,
rsignal(f, t) = E [Pf (atxclean)]
E [Pf (xt)] =
a
2
tSf
a
2
tSf + b
2
t
.Where the paper says it
Where the paper says it
Conversion from Cycles Per Pixel to Cycles Per Degree
The conversion of units is described in Appendix D. The visual angle subtended by one pixel is θpix = 2 arctan p / (2d) · 180/π degrees/pixel. p. 4 We quantify perceptual importance
using the contrast sensitivity function CSF. Since CSF is defined over frequencies in cycles per
degree of visual angle, while our image frequencies are measured in cycles per pixel, we first convert
image frequency bins to cycles per degree before evaluating CSF. [2026] and JiT-H Li
and He [2025]. In all our experiments, we use the original inference setups (e. We assume a pixel size of 0. 0114
cm and a viewing distance of 50 cm.Where the paper says it
Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Biased training time sampling can be implemented by using inverse transform sampling based on a smooth inverse CDF derived from CSF-based weights to bias training toward denoising times with larger wCSFlow values. This biases training toward
intervals where the newly recovered frequencies are more perceptually important.
-
Non-uniform inference step sizes can be constructed using the smooth inverse CDF to create a denser sampling grid in perceptually important denoising intervals and smaller step sizes where wCSFlow is large. This results in
a denser inference grid in perceptually important denoising intervals.
-
Training modifications can favor specific generation intervals by weighting the importance of certain t during model training, which is achieved by aligning frequencies generated at each noise level with human contrast sensitivity. This leads to
favoring the allocation of more model capacity to specific generation intervals.
Abstract
We introduce Contrast Sensitive Flow (CSFlow), a weighting scheme that connects the human eye's Contrast Sensitivity Function (CSF) to the iterative denoising steps of flow matching. Because real-world images concentrate signal at low spatial frequencies, these components reach high signal-to-noise ratio earlier during continuous diffusion than high-frequency components. When generating images with diffusion or flow matching models, this induces a soft autoregressive structure in Fourier space, where coarse image content stabilizes before fine detail. Meanwhile, the human visual system is unequally sensitive to spatial frequencies: very low and very high frequencies require significantly higher contrast to be perceived. We for the first time merge these observations through two contributions: (1) a metric that estimates which frequencies are generated at each reverse flow interval and (2) timestep weights obtained by aligning the frequencies generated at each noise level with human contrast sensitivity. We validate our contributions experimentally showing that these weights can improve generative performance by lowering FID by up to 12.15%, increasing Inception Score by up to 3.2% and improving GenEval scores by up to 3.0% using inference-only timestep modification or short fine-tuning. Qualitatively, we find that our CSFlow weights lead to better visual realism and less cartoonish appearance of generated images. In a two-alternative forced-choice study, participants preferred CSFlow generations over the baseline in 65% of comparisons. Additional visualisations, code, models and the example CSFlow weights are available on our project website.
Sources
- eDiff-I: Text-to-Image Diffusion Models with an Ensemble of Expert Denoisers
- BLIP3-o: A Family of Fully Open Unified Multimodal Models-Architecture, Training and Dataset
- A Fourier Space Perspective on Diffusion Models
- Back to Basics: Let Denoising Generative Models Denoise
- PixelGen: Improving Pixel Diffusion with Perceptual Supervision
Related papers
- TEA: Text Encoder Alignment for Robust Concept Erasure in Text-to-Image Models
- Do Egocentric Video-Language Models Capture Both Hand- and Object-Centric Cues?
- Aura: Consistent Multi-Subject Video Generation via VLM-Grounded Semantic Alignment
- PVRF: All-in-one Adverse Weather Removal via Prior-modulated and Velocity-constrained Rectified Flow
- DLWM: Diverse Latent World Models for Efficient Multimodal Reasoning
- Out of Sight, Still in Mind: Token Compression for Omni-LLMs