AdDiffusion reformulates diffusion model inference as a Markov Decision Process. A lightweight PPO-trained policy network observes the evolving latent state during denoising and selects from three actions at each step: continue, stop (early termination), or refine (selective region re-denoising). This enables prompt-dependent adaptive computation — simple prompts terminate early, complex prompts allocate additional effort to under-generated regions.

Standard diffusion inference uses a fixed number of steps regardless of prompt complexity. AdDiffusion wraps any pretrained diffusion model (without modifying its weights) and learns when to stop, when to continue, and when to selectively refine — all conditioned on the prompt and current generation quality.

Noise ──→ [Agent decides at each step] ──→ Final Image
              │
              ├─ continue: standard denoising step
              ├─ stop: early termination (save compute)
              └─ refine: re-denoise under-generated regions

• H1 (Efficiency): Equivalent quality to 50-step inference at ≤30 average NFE
• H2 (Quality): Higher quality than fixed-step methods at equal compute budget
• H3 (Adaptive): Prompt-dependent behavior — simple prompts stop early, complex prompts use more steps
• H4 (Refinement): Selective region refinement outperforms additional full-image steps

addiffusion/
├── src/
│   ├── diffusion/        # Pipeline wrapper, attention extraction, region refinement
│   ├── agent/            # Policy/value networks, state features, PPO, episode loop
│   ├── rewards/          # Composite reward with normalization
│   ├── evaluation/       # Metrics, VLM-as-Judge, significance testing
│   └── baselines/        # Fixed-step, SAG, Attend-and-Excite, oracle
├── tests/                # Unit + integration tests
├── configs/              # Hydra YAML configs
├── scripts/              # SLURM templates
├── discovery.md          # 38 analytical findings with fixes
├── experiment.md         # Step-by-step experiment execution guide
├── plan.md               # 14-week execution plan with decision gates
└── phase1.md             # Phase 1 completion report

# Install dependencies (requires Python 3.10, CUDA 12.4)
uv venv --python 3.10 && source .venv/bin/activate
uv sync

# Run CPU-only tests (no GPU needed)
uv run python tests/run_all.py --cpu-only

# Run full test suite (requires GPU + model weights)
uv run python tests/run_all.py

# Train the agent
uv run python src/train.py --config configs/default.yaml

# Or submit to SLURM
sbatch scripts/train.slurm

Benchmarks: COCO-30K, DrawBench, PartiPrompts, GenEval, T2I-CompBench

Metrics: FID, CLIP Score, ImageReward, HPS v2, Aesthetic Score, VLM-as-Judge (GPT-4o + Gemini)

Baselines: DDIM, DPM-Solver, LCM, SDXL-Turbo, SAG, Attend-and-Excite, Oracle-Stop

• Python 3.10, PyTorch 2.6.0, CUDA 12.4
• A100-40GB (SD 1.5) or A100-80GB (SDXL)
• ~800 GPU hours for full pipeline (training + baselines + evaluation)