Downloads · 30 days
12
100% of all-time downloads
Diffusion-cuneiform
hetanshishah/Diffusion-cuneiform
Diffusion-cuneiform is a machine learning model from hetanshishah. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for diffusers.
A diffusion model trained to generate novel Cuneiform symbols, one of the world's earliest known writing systems.
All-time downloads
12
Public
Parameters
77.4M
309 MB on disk
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors309 MB · 100%
At a glance
- Library
- diffusers
- Access
- Public
- Created
- Sep 6, 2026
- Updated
- Sep 6, 2026
- SHA
- 128a21aa
Model card
From the Hugging Face model README
🏺 Cuneiform Symbol Generation with Diffusion Models
A diffusion model trained to generate novel Cuneiform symbols, one of the world's earliest known writing systems.
Model
This model uses a UNet2DModel from Hugging Face Diffusers and generates 128×128 grayscale images of Cuneiform-like symbols.
Architecture
- Model: UNet2DModel
- Input/Output: 1-channel grayscale
- Image size: 128 × 128
- Channel widths: 128, 128, 256, 256, 512, 512
- Attention: Included at the 512-channel level
- Diffusers version: 0.37.1
Dataset
The base dataset was created from the Cuneiform block of Unicode Standard Version 17.0 (U+12000–U+1239F).
- 922 unique Cuneiform glyphs
- Geometric augmentations were applied to increase variation
- Final dataset: 10,142 images
- Images were cropped, centered, converted to grayscale, and resized to 128×128
Training
The model was trained locally on an RTX 3050 Laptop GPU.
- Epochs: 30
- Batch size: 16
- Optimizer: AdamW
- Learning rate: 1e-4
- Precision: FP16 mixed precision
- Loss: MSE
- Noise schedule: Linear
- Final dataset: 10,142 images
Results
Three dataset sizes were compared: 922, 5,000, and 10,142 images, using both DDPM and DDIM sampling.
The 10,142-image model produced the strongest results. DDPM sampling generated mostly coherent and visually meaningful Cuneiform-like symbols, capturing features such as wedge-shaped stroke tips and structured stroke layouts.
DDPM produced higher-quality results than DDIM, particularly in fine stroke separation and wedge-tip detail, although it was substantially slower.
The experiments also showed that increasing the dataset from 922 to 5,000 images produced the largest improvement in generation quality.
Limitations
Some generated symbols contain:
- Overly thick or merged strokes
- Faint or fragmented strokes
- Unusual arrangements of otherwise Cuneiform-like structures
The experiments suggest that dilation and erosion augmentations contributed to some of these artifacts.
Files
config.json
diffusion_pytorch_model.safetensors
- Library
- diffusers
- Created
- Sep 6, 2026
- Updated
- Sep 6, 2026