Original title: SPARK: Input-Conditioned Sparse Activation Modulation for Frozen DiT-based Super-Resolution
Federico Putamorsi, Leonardo Zini, Marcella Cornia, Lorenzo Baraldi · September 2026
Making blurry photos sharp using AI often relies on powerful models called Diffusion Transformers. Usually, improving these models means retraining them or adding heavy extra parts, which takes a lot of work. The researchers noticed that inside these models, a tiny fraction of massive signal pathways, called channels, control most of the activity. They found that tweaking these specific channels heavily changes image quality. So, they built SPARK. SPARK is a lightweight controller that leaves the main model frozen. It identifies the most active channels as the model runs. Then, it uses the low-resolution input to predict small adjustments just for those key channels. This simple tweak reliably boosted image sharpness and detail across tests, without needing to change the rest of the network.
Imagine you're baking a cake with dozens of ingredients, but the taste depends mostly on sugar and vanilla. If the cake isn't sweet enough, you don't remake the entire batter—you just adjust those two key ingredients. Here, the super-resolution model has many internal components, but a few dominant channels control the output quality. SPARK tweaks only these channels based on the input, leaving the rest untouched.
This helps developers and users of super-resolution tools get clearer images without expensive model updates. By targeting only a handful of channels, it makes improving quality more efficient and accessible for real-world applications.
The abstract doesn't specify limitations, but it focuses on DiT-based super-resolution models and tests on specific datasets. It may not generalize to other types of models or tasks beyond super-resolution. The approach only modulates a small number of channels, so its effectiveness in other contexts is unexplored.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
SPARK: Input-Conditioned Sparse Activation Modulation for Frozen DiT-based Super-Resolution — Federico Putamorsi, Leonardo Zini, Marcella Cornia et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.