docs/inrflow_integration.md

# INRFlow integration

INRFlow is implemented as a self-contained built-in ADAM model plugin. It uses
the existing manifest-driven Training and Generations interfaces, job queue,
checkpoint catalog, progress reporting, pause/cancel behavior, live previews,
and asset registry. No separate workspace or UI redesign is required.

## What is implemented

- RGB coordinate-value fields in ambient image space; there is no VAE or other
  modality-specific compressor.
- A patch-local cross-attention encoder that turns coordinate/RGB pairs into
  spatial context latents.
- A time-conditioned transformer trunk for global context.
- A query decoder that predicts one RGB velocity per requested coordinate-value
  pair.
- Conditional independent continuous flow matching from Gaussian noise at
  `t=0` to image data at `t=1`, with uniform or logit-normal time sampling.
- Random point-query subsampling during training.
- EMA model weights, periodic and final checkpoints, optimizer-state resume,
  deterministic preview seeds, and training metadata.
- Euler and Heun generation, live ODE previews, query chunking, Smart
  Generation, and square resolution-flexible coordinate queries.

## Comparison with regular Flow Matching

For a useful architecture comparison, train both backends from the same dataset
at the same native resolution, use the same number of optimizer updates, and
generate with the same seed and Euler/Heun step count. Keep INRFlow's output at
Native for the primary comparison. Its alternative output resolutions test the
continuous coordinate representation, but they are not an apples-to-apples
quality comparison because they extrapolate away from the training grid.

ADAM's default INRFlow is deliberately much smaller than the reference research
models so it can be explored on a consumer GPU. It preserves the defining model
structure and loss, but it should not be treated as a reproduction of published
quality or scaling results.

## Boundaries

MaskGIT remains out of scope. PixelRow is a separate autoregressive-row plugin,
and Glow is not implemented. None of those architectures share INRFlow's
checkpoint format.

Method reference: [INRFlow: Flow Matching for INRs in Ambient
Space](https://arxiv.org/abs/2412.03791). Reference implementation:
[apple/ml-inrflow](https://github.com/apple/ml-inrflow).