Abstract
Diffusion models typically generate images by following a path from random noise to a clean data state. However, these paths often intersect during training—a phenomenon termed "flow crossing" or XFLOW. Such intersections lead to semantic ambiguity, where the model predicts an average of multiple potential outcomes, resulting in inconsistent generations and artifacts during fast sampling. This paper proposes Non-Cross Diffusion, a method that untangles these generative trajectories by augmenting the input space with predicted noise. By "ascending the dimension," the model distinguishes between overlapping paths, significantly improving semantic consistency and image quality across varying sampling steps.
The Problem: Flow Crossing (XFLOW)
In standard diffusion training, different data points (e.g., a "cat" and a "bird") may pass through the same noisy state $x_t$ at the same timestep $t$. Because the training objective minimizes mean squared error, the model learns to predict the average of these disparate targets. This geometric limitation, known as XFLOW, causes several issues:
- Semantic Inconsistency: Changing the number of sampling steps can lead to entirely different subjects being generated from the same initial noise.
- Out-Of-Distribution (OOD) Samples: Fast, deterministic sampling (like DDIM) often encounters these "average" states, leading to blurry or nonsensical artifacts.
- Distillation Challenges: The inherent ambiguity makes it difficult to compress models into efficient, few-step versions.
Methodology: Untangling the Flow
The core idea of Non-Cross Diffusion is to "ascend the dimension" of the input. Just as two lines crossing in 2D can be separated in 3D, adding an extra condition to the diffusion model allows it to differentiate between intersecting paths.
Dimensionality Augmentation via Predicted Noise
The authors introduce the model's own previous estimate of noise, $\hat{\epsilon}$, as a "directional marker." Instead of predicting noise based solely on $(x_t, t)$, the model now considers $(x_t, t, \hat{\epsilon})$. This additional information helps the model identify which specific generative path it is currently traversing.
Training and Architecture
To implement this, the researchers utilize a ControlNet-like architecture. An auxiliary U-Net branch processes the noise condition, which is then fused with the main backbone. This allows the model to learn the untangling logic while preserving its original generative capacity.
A Bootstrap Training Strategy is employed to ensure robustness. During training, the model randomly alternates between a "zero" condition and its own predicted noise. This prevents the model from relying on ground-truth noise that is unavailable during inference and helps it handle early-stage estimation errors.
Key Results and Metrics
The authors introduced a new metric, Inference Flow Consistency (IFC), which measures the semantic shift of an image as it progresses through different sampling timesteps using PSNR.
- Superior Consistency: On CIFAR-10, Non-Cross Diffusion significantly outperformed standard iDDPM in the IFC metric, particularly at low step counts (5, 10, and 20 steps).
- Improved Image Quality: The method achieved lower Fréchet Inception Distance (FID) scores across nearly all step counts. For instance, at 10 steps on CIFAR-10, the proposed method reached an FID of 12.85, compared to 15.91 for the baseline.
- Artifact Suppression: Qualitative analysis demonstrates that while baseline models often produce "blobs" or semantically inconsistent images when steps are reduced, Non-Cross Diffusion maintains the same subject regardless of sampling speed.
Conclusion and Future Work
Non-Cross Diffusion addresses a fundamental geometric bottleneck in diffusion models. By straightening generative trajectories, it enables more reliable and semantically stable image synthesis. While the current implementation requires training from scratch or significant fine-tuning, future work aims to integrate these "untangling" principles into pre-trained text-to-image models more efficiently.
Summary updated