prior-diffusion

Prior-guided diffusion research scaffold focused on controllability-quality trade-offs.

What this codebase optimizes for

• Thin integration with official Diffusers APIs
• Pluggable latent initialization operators
• Pluggable denoising step hooks
• Easy ablation across model, scheduler, init policy, and start timestep offset

Quickstart

1. Install dependencies:

pip install -e ".[dev]"

2. Run baseline (Gaussian init):

python main.py \
  --prompt "a dancer in a studio, full body" \
  --model-id stable-diffusion-v1-5/stable-diffusion-v1-5 \
  --controlnet-id lllyasviel/control_v11p_sd15_openpose \
  --condition-image examples/openpose.png \
  --init-operator gaussian \
  --output-dir outputs

3. Run low-frequency prior injection:

python main.py \
  --prompt "a dancer in a studio, full body" \
  --model-id stable-diffusion-v1-5/stable-diffusion-v1-5 \
  --controlnet-id lllyasviel/control_v11p_sd15_openpose \
  --condition-image examples/openpose.png \
  --init-operator lowfreq \
  --lowfreq-strength 0.4 \
  --lowfreq-cutoff 0.2 \
  --output-dir outputs

Directory layout

• src/prior_diffusion/adapters: Thin wrappers around Diffusers
• src/prior_diffusion/operators: Swappable methods (latent init, step hooks)
• src/prior_diffusion/experiments: Runtime composition and execution
• tests: Interface-focused unit tests
• docs/PROJECT.md: Current high-level architecture contract