ShellD (Shell Diffusion)
Small DiT-based Text-to-Image Latent Diffusion Model
ShellD is a lightweight text-to-image model that generates 256×256 images from natural language prompts. It uses a Diffusion Transformer (DiT) backbone operating in a compact VAE latent space, making it feasible to train and run on consumer GPUs.
Model Architecture
Text Prompt → [MiniLM-L6-v2 (frozen)] → Text Embedding (384-d)
↓
Random Noise → [VAE Encoder] → Latent (16ch) → [DiT (12 blocks)] → Denoised Latent → [VAE Decoder] → 256×256 Image
| Component | Details | Params |
|---|---|---|
| Text Encoder | sentence-transformers/all-MiniLM-L6-v2 (frozen) | 22.71M |
| VAE | Encoder + Decoder with residual blocks, 3 down/up stages, latent dim=16 | 23.43M |
| DiT | 12-layer Transformer with self-attention, cross-attention (text), and adaptive timestep conditioning. Patch size=4, hidden dim=256, 8 heads | 20.80M |
| Total | 66.95M (trainable: 44.23M) |
VAE (Autoencoder)
The VAE compresses 256×256 RGB images into a 16-channel latent with spatial size 32×32 (downsampled by 8×). It uses residual blocks with GroupNorm and SiLU activations. During training, a KL penalty (β=0.1) keeps latents close to a standard normal distribution.
DiT (Diffusion Transformer)
The DiT operates on patched latents (patch size 4 → 8×8 = 64 patches). Each block includes:
- Self-attention for spatial relationships
- Cross-attention conditioned on text embeddings
- Adaptive timestep conditioning via an MLP-projected sinusoidal embedding
- Dropout (0.1) in attention and MLP for regularization
Diffusion Process
Standard DDPM (Denoising Diffusion Probabilistic Model) with 1000 timesteps and a linear beta schedule (β₁=1e‑4, βᵀ=0.02). The model is trained to predict the added noise ε. Classifier-free guidance (CFG) is used during training with a text-conditioning dropout probability of 15%.
Training
Dataset
jackyhate/text-to-image-2M — ~2M high-quality text-image pairs in webdataset format. Loaded via the datasets library with streaming to avoid materializing the full 2TB+ dataset into memory. A rotating 2000-image in-memory buffer (~400 MB RAM) is refreshed each epoch from a fresh random stream to provide shuffle diversity without disk I/O bottlenecks.
Data Augmentation
Random horizontal flip (p=0.5), color jitter (brightness/contrast/saturation ±0.2, hue ±0.05), and random affine transforms (rotation ±10°, translation ±5%, scale 0.9–1.1) via torchvision.
Hyperparameters
| Parameter | Value |
|---|---|
| Image size | 256×256 |
| Batch size | 8 |
| Optimizer | AdamW (β₁=0.9, β₂=0.999, lr=1e‑4, weight decay=0.01) |
| LR schedule | Linear warmup (500 steps) + Cosine annealing |
| Gradient clipping | 1.0 (norm) |
| Mixed precision | FP16 via torch.cuda.amp.GradScaler |
| Dropout | 0.1 (DiT attention + MLP) |
| EMA | Exponential moving average (decay=0.999) applied at every step; EMA weights used for validation and final checkpoint |
| Early stopping | Patience of 8 epochs on validation loss (10% held-out split) |
| KL weight | 0.1 (β-VAE style) |
VAE Pretraining
Before diffusion training, the VAE is pretrained for 10 epochs on reconstruction + KL loss with a higher learning rate (lr=1e‑3) to establish a meaningful latent space. During this phase the DiT and text encoder are frozen.
Training Phases
- VAE Pretraining (10 epochs) — Train encoder + decoder on image reconstruction to establish a meaningful latent space. DiT and text encoder are frozen.
- DiT Diffusion Training (up to 30 epochs, early-stopped) — Freeze VAE, train DiT to denoise latents conditioned on text embeddings. CFG dropout randomly replaces text embeddings with a learned null embedding to enable classifier-free guidance at inference time.
Usage
Requirements
pip install torch safetensors sentence-transformers pillow numpy huggingface-hub
For training, also install:
pip install torchvision datasets
Inference (standalone — loads from Hugging Face)
from inference import ShellDInference
pipe = ShellDInference("FlameF0X/ShellD")
image = pipe.generate("a serene lake surrounded by mountains")
image.save("output.png")
ShellDInference automatically downloads weights from Hugging Face via huggingface_hub on first use and caches them locally.
Streaming Generation
View the diffusion process unfold step-by-step:
for img, step_info in pipe.generate_stream(
prompt="a futuristic city at night",
num_steps=250,
cfg_scale=3.0,
display_every=25, # emit an image every 25 steps
):
print(f"Step {step_info['step']}/{step_info['total']}")
img.save(f"progress_{step_info['step']:04d}.png")
Intended Use
- Educational exploration of diffusion transformers
- Lightweight text-to-image generation on consumer hardware
- Starting point for fine-tuning on custom datasets
Limitations
- 256×256 resolution only (no upscaling built in)
- Limited prompt understanding due to small DiT and frozen lightweight text encoder
- Quality depends on training data distribution — may not match large-scale models like SDXL or Flux