OpenSenseNova released Looped-DiT, a PyTorch implementation of a text-to-image diffusion transformer. It scales computation by looping shared transformer blocks within each denoising step. The model uses Deep Supervision and Self-Modulating Attention to stabilize training. It builds on the MMDiT backbone from MiniT2I and a frozen FLAN-T5-Large text encoder. The repository includes training, inference, and evaluation code for B/32, B/16, and L/16 variants. Pretrained checkpoints are available on Hugging Face for B/32 and B/16. Evaluation covers GenEval, DPG-Bench, PRISM-Bench, T2I-CoReBench, SpatialGenEval, and TIIF-Short. Training data includes CC12M, FLUX-Reason-6M, and fine-tuning datasets like Std-120K and Fine-T2I. The code supports inference at any loop depth without retraining. Default settings use 100 Euler steps, guidance scale 6.0, and loop depth 4.
What it unlocks
Running a shared group of transformer blocks multiple times per denoising step to scale computation without increasing model size.
- 4 loop depth
- 512 x 512 resolution
- 100 Euler steps
- 6.0 classifier-free guidance scale
What you need to act on it
- PyTorch 2.1 or newer with CUDA
- 8 GPUs for evaluation
- separate environments for mmdet and vLLM
Sources