code/docs/actor-3.md

# Actor-3 — Better consistency

Actor-3 is the consistency-oriented MR-IQA-2 Actor. It was initialized from
Qwen3.5-4B and trained for five epochs (1,455 optimizer updates). The language
model was fully trainable; the vision encoder and visual aligner were frozen.

The training objective combines pairwise source/edited-image consistency with
retained edit utility and loss-side KL (`beta=0.02`). Its preservation-first v9
prompt discourages semantic changes to subjects, objects, layout, text,
composition, and unaffected regions.

The exact prompt contracts are published as:

- [training prompt](../configs/prompts/actor_3_training_prompt.json)
- [test prompt](../configs/prompts/actor_3_test_prompt.json)

Run deterministic single-image inference with:

```bash
python examples/actor_3_inference.py /absolute/path/to/input.jpg
```

To use an already downloaded checkpoint:

```bash
python examples/actor_3_inference.py /absolute/path/to/input.jpg \
  --model /absolute/path/to/actor-3 \
  --local-files-only
```

The script prints and optionally saves the raw completion plus the parsed
`reasoning.evidence`, `reasoning.solution`, and `rating` fields.