Generate mallards.

Training data comes from verified mallard photos and high-quality segmented web crops. Hard negatives helped build the detector and reward filters, but the generator release is trained to model mallard-like positive examples.

Images are encoded into a 24 x 32 x 32 latent tensor. Training in this space keeps the model small enough for local Apple Silicon experiments while preserving more structure than the earlier tiny latent attempts.

A reward-weighted LatentUNetFlowV5 prior learns a continuous velocity field from noise toward plausible mallard latents. We sample it with Heun integration and a cosine time schedule.

The prior output is projected toward nearby entries in a top-256 mallard condition bank. This is the main global-structure aid: it reduces drifting shapes and keeps poses closer to real examples.

A LatentUNetFlowV6 refiner maps the projected condition into the final high-resolution latent. This stage is where feet, body outline, and water/feather transitions improved most.

The final latent tensor is decoded back to a 128 px RGB mallard image grid. The released sampler returns this PNG grid as base64 in the inference response.

The web app writes requests to S3, invokes SageMaker async inference, then polls for a JSON result containing a generated PNG grid. The endpoint is normally kept offline to avoid idle GPU cost.