Skip to main content

Amanda Co-Pilot — Generative Motion Synthesis

In training — coming post-launch

Amanda Co-Pilot is actively training. The Co-Pilot interface (left slide panel inside the Kinetiq Editor) is live as a placeholder while the underlying generative motion model finishes training. Demo-ready releases will roll out post-launch through 2026.

Amanda 2.0 multi-engine architecture diagram

Architecture diagram of the Amanda 2.0 multi-engine system. The Co-Pilot panel inside Kinetiq is the motion-generation extension of this architecture.

What it is​

Amanda Co-Pilot is the in-editor generative motion partner. It lives as a left slide panel inside the Kinetiq Editor at app.thirdrez.com/sandbox and app.thirdrez.com/kinetiq. When the user describes the motion they want in plain language — "a frustrated wave goodbye", "a confident strut walk cycle", "a falling-asleep idle" — Amanda Co-Pilot will generate the keyframes directly into the active timeline, ready to be polished by Mirror, Loop Fixer, Foot Locking, and the rest of the Kinetiq pipeline.

The Co-Pilot is not the same product as the standalone Amanda 2.0 chatbot at thirdrez.com/amanda. The chatbot is a general-purpose AI assistant (programmer, secretary, RP, multiple personas). The Co-Pilot is purpose-built for motion synthesis inside the editor.

Status: in training​

Two parallel training tracks are underway:

1. The MDM/LoRA generative motion model​

Built on Motion Diffusion Models (MDM) with Low-Rank Adaptation (LoRA) style adapters. Trained in-house to keep iteration costs low while we refine the architecture and the ecological dataset. The base model handles general human locomotion and gesture; LoRA adapters specialize for Casual, Combat, Dance, Romance and other domain styles (see Style LoRAs Vision for the full list).

2. The Co-Pilot interface​

The left-slide panel UI inside the Kinetiq Editor is already deployed as a placeholder showing "Amanda 2.0 is in training". This lets early users see where the feature will live and starts establishing the muscle memory of "describe motion → motion appears in timeline" before the inference backend goes live. The panel visible today is the literal interface that will host the Co-Pilot once the model graduates from training.

This dual-track approach — interface-first, model-second — means launch day for Co-Pilot will not be a UI change. It will be the moment the prompt input goes from disabled to enabled, and motion starts streaming into the timeline.

What you can do today​

While the generative motion model finishes training, the rest of the Kinetiq Engine is fully operational:

  • Vision Capture (BETA) — webcam or video file → mediapipe pose extraction → 3D skeleton. The closest thing to "AI motion" available today, and it works right now.
  • Retarget Studio — pull animations from Mixamo, mocap libraries, or other rigs and retarget onto your avatar.
  • Marketplace — browse and download production-ready animations curated by the Thirdrez team.
  • Standalone Amanda 2.0 chatbot — multi-engine AI for general assistance (programming, RP, secretary, etc.). Free tier includes 10 messages/day.

Why we chose to ship interface before backend​

GPU compute for inference is expensive. Running real-time prompt-to-motion serving for thousands of users while the model is still iterating its training would burn budget on inference of suboptimal weights. By landing the interface first and being honest about the "in training" state, we:

  1. Preserve runway for proper training cycles (we don't ship half-baked weights to production)
  2. Establish the UX before users experience the model — early users are seeing the panel, the placeholder, and the menu position, building muscle memory
  3. Stay coherent with the architecture — the diagram above shows the multi-engine structure that's already live for the standalone Amanda chatbot; the Co-Pilot is the "motion generation" engine in that same family, queued behind training completion

This is honest, not vaporware. The Co-Pilot panel exists. The MDM/LoRA model exists and is training. The integration plumbing is queued. The thing missing is time, not technology.

Roadmap timeline​

PhaseTargetDescription
Co-Pilot UI placeholder✅ LiveLeft slide panel deployed in /sandbox and /kinetiq
Style LoRA pipeline finalizationQ2 2026Training infrastructure, eval harness, adapter library
Closed beta (selected creators)Q2-Q3 2026Limited inference availability for trusted Discord testers
Public BETAH2 2026Backend goes live, prompt input enabled for paying tiers
GA + style LoRA marketplaceTBDUser-trained LoRAs, style sharing, advanced controls

Subscribe to release notes by following @ThirdrezHQ or joining Discord.