Adaptive noise in an SDE sampler improved texture by only 1–2% on Flux 3 Klein

Image generators built on flow matching can run with a stochastic sampler that mixes in fresh noise at every step. The strength of that noise, the parameter eta, is usually tuned by hand once and left constant. The author of a study on Habr decided to test whether the sampler can choose eta on its own, step by step, from data it already has. The result is mostly negative: the adaptive scheme added almost nothing, and the best effect on Flux 3 Klein was about 1–2% of the texture metric value.

The idea: steer noise with free signals

At every step, a multistep solver already computes two useful numbers. The first is the trajectory's curvature, meaning how sharply it bends. The second is the discrepancy between the prediction and the correction, which serves as an estimate of local error. Neither requires extra network calls, so the scheme costs nothing computationally: no additional passes through the model and no increase in generation time.

The hypothesis was logical. If noise hurts where the field is "stiff" and helps where it is smooth, then a fixed noise budget distributed according to curvature should beat a constant. To test this, the author wrote a ComfyUI node and ran it on three open models: Anima, Flux 3 Klein and MOSS-Image. The number of steps ranged from 6 to 15.

The key methodological decision was a fair baseline. Each adaptive run was compared with a run using a constant eta and exactly the same total noise budget. Without that, as the author's previous work showed, an "adaptive" result most often turns out to mean simply "less noise," not a better distribution.

What worked and what did not

Curvature did not work. No model shows a stable gain, and the sign of the effect changes with the number of steps. In practice, this means you should not rely on curvature when choosing eta.

The solver error estimate gave a small but reproducible signal. When more noise is added where the error is small, the gain on Flux 3 Klein is about 1–2% of the texture metric value. It held up on new seeds, under criteria fixed before the run. On MOSS-Image the effect is one-sided, and on Anima the author found none. On Krea 2 the result is inconclusive because of insufficient statistical power.

The most interesting finding concerns adaptation itself. A fixed noise schedule, averaged over independent runs and applied to all images, gives the same result as step-by-step selection. Noise at the edges of the noise window helps, while the middle matters less. In other words, what matters is where the noise sits along the trajectory, not what the sampler sees at a given moment. Adaptive logic is then redundant: a precomputed schedule is enough, costs nothing at runtime and is easier to debug.

People did not notice the difference

The author ran a survey of 39 people, who had to pick, from two images, the one that looked more AI-generated. Participants noticed a large difference between weak and strong noise. They could not tell the new noise distribution from ordinary noise. Even a small measurable gain in latent space, then, does not translate into a visible improvement.

Limitations and practical takeaway

The author lists the weak points of the work. The metrics are computed in latent space, not by how the eye perceives images. By default, 16 image pairs were used for each noise budget. Anima ran on a different version of ComfyUI than the other models. The survey is informal and was conducted without participant screening. All of this means the conclusions cannot be carried over to other models and settings without re-testing.

The practical takeaway for pipeline developers is simple. If you are tuning an SDE sampler for a flow-matching model, do not complicate it with adaptive rules. First find a fixed noise schedule that accounts for the edges of the window, and check the result at an equal total budget. Otherwise you risk mistaking a simple reduction in noise for a success of the method.

The author believes such negative results are worth sharing, and has published the code, per-image numbers and survey analysis in the flow-sde-research repository on GitHub. A preprint is available on Zenodo.