Revisiting what a diffusion model predicts
A preprint submitted to arXiv on October 16, 2026 examines design choices in transformers that generate directly in pixel space. The researchers propose MFlow, which learns a mixture between clean-image and velocity prediction. They observe that the preferred parameterization changes with model capacity and data representation.
The tested vanilla pixel-transformer family also shows a deficit in high-frequency information associated with fine surface texture. ConvRefiner is a small convolutional module trained after the main model while keeping that backbone frozen. The paper and project page report that frequency characteristics can improve while FID stays similar, illustrating why one aggregate score does not capture every visual property.
A practical interpretation for creators
Unconvincing fabric or leaf texture in a crop may reflect architecture and training choices rather than an insufficiently detailed prompt. That does not make this a universal explanation for all generated-image defects, nor does it prove that scaling alone fixes them. The manuscript is under review, its results are author-reported, and it is not an Adobe product-feature announcement. Production decisions should still combine actual image crops with human inspection of material detail and structure.