mahadev-balla/isl-conditional-diffusion-cosine-64

ISL Conditional Diffusion — 64×64

A class-conditioned DDPM trained from scratch to generate 64×64 RGB images of Indian Sign Language hand gestures.

The model uses class conditioning on one of 35 ISL classes and supports classifier-free guidance during sampling.

The complete implementation and experiments are available in the GitHub repository.

Model

• DDPM with UNet2D architecture
• 64×64 resolution
• RGB images
• Class-conditional generation
• 35 ISL classes
• Classifier-free guidance (CFG)
• EMA weights

Training

• Dataset: 42,000 images (1,200 images per class × 35 classes)
• Noise schedule: cosine
• Batch size: 128
• Learning rate: 1e-4
• Mixed precision: fp16
• Training steps: 65,000
• EMA decay: 0.9999
• CFG label dropout: 0.15
• Data augmentation: enabled

Sampling

• Default sampler: DDIM
• Training diffusion timesteps: 1,000
• Routine inference steps: 100
• Evaluation inference steps: 50
• Default guidance scale: 1.0
• Random seed: 42

Results

At the FID-optimal guidance scale of 1.0:

• FID: 57.24
• Semantic accuracy: 99.2%

The model provides strong class control under the reported evaluation setting.