erwold/mv-split

MVSplit-DiT is the research code and released model weights behind an arXiv paper called Mean Variance Split Residuals for 1000 Layer Diffusion Transformers. Diffusion Transformers are a type of AI model that generates images by starting with random noise and gradually cleaning it up into a picture. Most of these models stack a modest number of layers, so the paper's main contribution is a new way of building the residual connections, the shortcuts that carry information between layers, called MVSplit. That change is what lets this model stack 1000 layers instead of the usual handful, and the released checkpoint uses exactly that depth. To actually use it, you run a sampling script that takes a text prompt and produces an image. Three pieces work together: the 1000 layer DiT model itself, a Qwen3 text encoder that turns your prompt into numbers the model can use, and a FLUX.2 autoencoder that turns the model's internal output into real pixels. You need to download all three separately, since only the DiT weights ship from this repository's own Hugging Face page, the other two come from their own sources. Generation happens through a flow matching process, a step by step method for turning noise into a finished image over a set number of steps. For speed, the heaviest math runs on custom Triton GPU kernels, borrowed and adapted from the Unsloth project, with slower plain PyTorch versions available as a fallback on machines without Triton. Settings you can adjust include image size, how many denoising steps to run, how closely the output follows your prompt, and batch size, though several architecture flags must match the released checkpoint exactly or generation will fail. The project is written in Python and requires a CUDA GPU for practical use, CPU only works through the slow fallback path.