Linearizing Vision Transformer with Test-Time Training
For practitioners deploying large vision Transformers, this work provides a practical method to reduce inference cost while leveraging pretrained weights, though the approach is domain-specific to architectures structurally aligned with TTT.
The authors address the challenge of converting pretrained Softmax attention Transformers to linear-complexity attention without retraining from scratch. By aligning architectural and representational properties with Test-Time Training (TTT), they achieve comparable text-to-image quality to fine-tuned Softmax models in Stable Diffusion 3.5, with 1.32× and 1.47× inference speedups at 1K and 2K resolutions after only 1 hour of fine-tuning.
While linear-complexity attention mechanisms offer a promising alternative to Softmax attention for overcoming the quadratic bottleneck, training such models from scratch remains prohibitively expensive. Inheriting weights from pretrained Transformers provides an appealing shortcut, yet the fundamental representational gap between Softmax and linear attention prevents effective weight transfer. In this work, we address this conversion challenge from two perspectives: architectural alignment and representational alignment. We identify Test-Time Training (TTT) as a linear-complexity architecture whose two-layer dynamic formulation is structurally aligned with Softmax attention, enabling direct inheritance of pretrained attention weights. To further align representational properties, including key shift-invariance and locality, we introduce key instance normalization and a lightweight locality enhancement module. We validate our approach by linearizing Stable Diffusion 3.5 and introduce SD3.5-T$^5$ (Transformer To Test Time Training). With only 1 hour of fine-tuning on 4$\times$H20 GPUs, SD3.5-T$^5$ achieves comparable text-to-image quality to the fine-tuned Softmax model, while accelerating inference by 1.32$\times$ and 1.47$\times$ at 1K and 2K resolutions.