Architecture memo
This page exists for investors, research partners, and buyers who already understand diffusion models and want to know what we built around them.
Candidate generation
FLUX provides the pixel engine. Each run is a structured sweep through a constrained latent neighborhood rather than a single guess. Variance is budgeted. Identity drift is capped before ranking runs.
Identity preservation
The embedding is a contract, not a soft prompt. Landmarks, micro-texture, and lighting response must remain coherent across the candidate set or the frame is discarded upstream.
Ranking and selection
The composite score is explicit: aesthetic coherence, predicted engagement under calibrated assumptions, identity lock, and diversity across the final basket. Weights encode policy you set per channel or per launch.
Training and compute
Training and fine runs execute on Lambda Labs H100 clusters with deterministic job graphs. Private throughput routing is available when latency and custody requirements demand it.
Deployment posture
Default delivery is off-chain generation with signed URLs and lineage metadata. Optional chain attestations exist for teams that require an external witness. Neither path changes the core claim: reconstructed photography, ranked, identity-bound.