🌟 NovaFace-DiT (512x512)

NovaFace-DiT is a Multimodal Diffusion Transformer (MM-DiT) model trained entirely from scratch for high-fidelity human face synthesis. It leverages the powerful Rectified Flow Matching technique and is deeply inspired by the Stable Diffusion 3 architecture.

Despite being trained on a highly constrained hardware setup (a single consumer-grade GPU) and a highly curated dataset (70,000 images from FFHQ), NovaFace-DiT demonstrates the incredible efficiency and scaling capability of the custom MM-DiT architecture.

High-fidelity samples generated by NovaFace-DiT using complex text prompts.

📊 Model Details

• Model Type: Text-to-Image Diffusion Transformer (MM-DiT)
• Parameters: ~260 Million
• Text Encoder: T5-Base (768-dim)
• Latent Space: Custom 8-channel VAE (f8)
• Training Dataset: FFHQ (Flickr-Faces-HQ)
• Resolution: 512x512
• License: Creative Commons BY-NC-SA 4.0 (Non-commercial)

⚡ Requirements & Custom VAE

NovaFace-DiT operates in an optimized 8-channel latent space and requires our custom-trained Autoencoder (VAE) to decode images properly. Standard SDXL or SD3 VAEs are not compatible.

👉 Download the Custom 8-Channel VAE here (Note: Please download this VAE to generate images)

🚀 How to Use (Code & UI)

This repository contains only the model weights (.safetensors). To actually generate images, inspect the architecture, or resume training, please visit our official GitHub repository which contains a full production-ready Gradio UI and training pipeline.

🔗 Official GitHub Repository: devbnamdar/MM-DiT-From-Scratch

Quick Setup:

1. Clone the GitHub repository.
2. Download the NovaFace-DiT.safetensors from this Hugging Face page and place it in your local checkpoints/ directory.
3. Download the Custom VAE from its separate repository and place it in your local vae_models/ directory.
4. Launch the Gradio app:

python gradio_ui/app.py

5. In the Gradio UI, go to the "⚙️ Settings" tab, enter the path to your downloaded model (e.g., checkpoints/NovaFace-DiT.safetensors) in the "Base Model Path" field, and click "Load Models to GPU".

⚠️ Limitations and Bias

• Domain Specific: This model was trained exclusively on the FFHQ dataset. It is highly specialized in generating human portraits (shoulders and above). It is not designed to generate landscapes, animals, or full-body shots.
• Text Rendering: The model does not generate legible text or complex typography.
• Bias: As the model is trained on FFHQ, it may inherit demographic or lighting biases present in the original dataset.

📄 Citation

If you use this model or the accompanying codebase in your research or projects, please cite:

@misc{namdar2026mmdit,
  author       = {Namdar, Bunyamin},
  title        = {MM-DiT From Scratch: High-Fidelity Diffusion Training on Limited Dataset},
  year         = {2026},
  publisher    = {GitHub},
  url          = {https://github.com/devbnamdar/MM-DiT-From-Scratch}
}