In generative modeling, creating a transition between two images—such as morphing a cat into a dog or a photograph into a sketch—is often framed as a simple walk along a line. If we have a starting point $A$ and an ending point $B$, we might assume that the midpoint of the transition is simply $0.5A + 0.5B$.
However, researchers have consistently found that mathematical linearity does not equal perceptual linearity. When we interpolate linearly through the latent space of a diffusion model, the resulting animation often "stalls" at the beginning and end, then "bursts" with sudden, jarring changes in the middle. To solve this, a new direction in generative research focuses on perceptually-uniform sampling: remapping the interpolation schedule so that the rate of visual change remains constant to the human eye.
The Non-Linearity of Latent Space
The core problem, as identified by Kaiwen Zhang and colleagues in DiffMorpher, is that the "functional gap" between two points in a model's parameter space is not uniform. In a 100-frame morphing sequence, frames 1 through 40 might look nearly identical to the start image, while frames 45 through 55 contain almost the entire transformation, leading to a "snapping" effect.
This phenomenon isn't limited to image morphing. Yael Vinker observed in CLIPascene that when simplifying a scene into a sketch, the loss of detail is non-linear; a small mathematical change in abstraction can lead to a sudden, abrupt loss of the scene's "essence." Similarly, Hongwei Lin noted in visual curve completion that human perception of smoothness varies dynamically based on distance, a biological reality that static mathematical models often fail to capture.
Measuring Visual Velocity
To fix these jerky transitions, we first need a way to measure "visual velocity." Instead of measuring the distance in the model's hidden layers, researchers use perceptual distance metrics (like LPIPS) that are designed to mimic human judgment.
In the IMPUS framework, Zhaoyuan Yang and Zhengyang Yu propose using these metrics to guide a search for the "true" midpoints. If the goal is to produce a sequence of $N$ frames, the system doesn't just pick $N$ equally spaced mathematical points. Instead, it looks for points where the perceptual distance between each consecutive frame is identical.
Adaptive Scheduling and Search
How do we find these perceptually-even points? Several strategies have emerged:
- Binary Search: Zhaoyuan Yang in the ICLR 2024 version of IMPUS uses a binary search algorithm. By iteratively testing different interpolation weights and measuring the resulting image's perceptual distance from the previous frame, the algorithm "hones in" on the weight that provides exactly the right amount of visual progress.
- Rescheduling Functions: In SOYO, Qifan Yu and Binghe Yu apply this to video style morphing. They pre-sample the styles to measure their visual divergence and then calculate a non-linear blending curve. This ensures that a transition from, say, Impressionism to Cubism doesn't "jump" into the new style prematurely.
- Statistical Distributions: Qiyuan He and Jinghao Wang suggest a more direct approach in AID. They use a Beta distribution to sample interpolation coefficients. By concentrating more sampling steps in the "critical transition zones" where image content shifts most rapidly, they achieve a smoother perceived result without the need for intensive iterative searching.
Why It Works: Aligning Math with Intuition
The success of these methods lies in their acknowledgment that generative models are "black boxes" whose internal geometry does not naturally align with human vision.
By decoupling the temporal midpoint (the frame index) from the mathematical midpoint (the interpolation weight), these techniques allow the model to spend more "computational time" on the difficult parts of a transformation. As Kaiwen Zhang notes in DiffMorpher (CVPR 2024), this redistribution of steps ensures that the transformation appears to progress at a constant speed to the observer, effectively "unleashing" the capability of diffusion models to create high-fidelity, natural animations that feel as smooth as a cinematic dissolve.