02 / 05 · Forge
Best-of-three on every hero shot.
For each hero asset we fire three models in parallel. The judge layer (next page) picks the winner. You only see the survivor — but the rejected attempts stay in the lineage so we can train the LoRA on what passed.
Multi-model
Why ensembles beat single-model pipelines
Every model has a failure mode. Flux is great at lighting; sometimes loses fine product detail. Recraft is great at composition; sometimes mangles text. Ideogram is great at text; sometimes flatter lighting. Running all three on the same brief and letting the judge pick lifts hero-shot accept rate from ~40% to ~75% on our benchmarks.
Brand LoRA
Per-workspace, compounding moat
Every workspace gets its own LoRA, trained on your reference photos via Vertex AI Custom Training. Vertex re-trains automatically every 50 approvals, which means a six-month-old workspace produces measurably better outputs than a brand-new one.
On top of LoRA: IP-Adapter / Redux pinning lets you anchor a specific reference photo for outputs that need to match a specific lookbook frame.
The stack
What's actually firing
| Asset kind | Models | Where |
|---|---|---|
| Image — hero | Flux 1.1 Pro Ultra · Recraft V3 · Ideogram 3.0 | fal.ai · Replicate |
| Image — batch | Flux Schnell · Recraft V3 fast | RunPod self-hosted A100 |
| Video — hero | Veo 3 · Kling 2.6 Pro · Runway Gen-4 | Vertex AI · fal.ai |
| Video — batch | Wan 2.2 5B · Wan 2.2 A14B | RunPod self-hosted A100 |
| Talking head | HeyGen Rachel · Hedra Character-3 | HeyGen · Hedra |
| Voice clone | ElevenLabs Pro v2 (per-brand) | ElevenLabs |
Cost discipline
Spend caps, not surprise bills
Daily and per-run spend caps configured in Secret Manager, enforced before the first inference call. If your team's running a campaign and crosses the cap, the dispatcher pauses and pings the workspace owner — no $5K Friday-night bills.