---
license: mit
---
[![Discord](https://img.shields.io/discord/232596713892872193?logo=discord)](https://discord.gg/2JhHVh7CGu)

A semi custom network trained from scratch for 799 epochs based on [Simpler Diffusion (SiD2)](https://arxiv.org/abs/2410.19324v1)

[Modeling](https://huggingface.co/Blackroot/SimpleDiffusion-TensorProductAttentionRope/blob/main/models/uvit.py) || [Training](https://huggingface.co/Blackroot/SimpleDiffusion-TensorProductAttentionRope/blob/main/train.py)

This network uses the optimal transport flow matching objective outlined [Flow Matching for Generative Modeling](https://arxiv.org/abs/2210.02747)

A modified tensor product attention with rope is used instead of regular MHA [Tensor Product Attention is All You Need](https://arxiv.org/abs/2501.06425)

xATGLU Layers are used in some places [Expanded Gating Ranges Improve Activation Functions](https://arxiv.org/pdf/2405.20768)

This network was optimized via [Distributed Shampoo Github](https://github.com/facebookresearch/optimizers/blob/main/distributed_shampoo/README.md) || [Distributed Shampoo Paper](https://arxiv.org/abs/2309.06497)

```python train.py``` will train a new image network on the provided dataset (Currently the dataset is being fully rammed into GPU and is defined in the preload_dataset function)

```python test_sample.py step_799.safetensors``` Where step_799.safetensors is the desired model to test inference on. This will always generate a sample grid of 16x16 images.

| | |
|:---:|:---:|
| ![samples](./epoch_39.png) | ![samples](./epoch_159.png) |
| ![samples](./epoch_459.png) | ![samples](./epoch_799.png) |

![stats](./stats.png)