# TODO:
# - Generate embeddings (1, 1024) without context length
# - DDPM Linear Warmup + Cosine Annealing - LR scheduler
# - DDPM ancestral sampler

# - CFG rescale (SDXL trick)
# σ = std(uncond)
# guided = uncond + s * (cond - uncond)
# guided = guided * (σ / std(guided))

# - Progressive distillation

# | Thing                    | Why it exists                   |
# | ------------------------ | ------------------------------- |
# | DDPM ancestral sampler   | More randomness & diversity     |
# | Latent → image decode    | Turn tensors into actual images |
# | CFG rescale              | Fix CFG artifacts               |
# | Progressive distillation | Go from slow → real-time        |