Shared decision cues for choosing between commercial services and open models, then understanding what actually makes a credible conversational video experience.
What the numbers say
Advertised latency or FPS describes one part of rendering, not presence by itself. Test audio-video synchronisation, recovery after a speech turn and stability as network conditions change.
What to combine in the table
Read real-time alongside transport, conversation, body, performance direction and sovereignty. A highly realistic face may suit a demo without providing the direction, listening or scene format you require.
Stay alert to
The avatar is only one system layer: STT, TTS, LLM, WebRTC and orchestration all affect the experience. Also verify likeness rights, moderation, data residency, parallel streams, session cost and a fallback plan.
Reading method: start with the scene to sustain and the required degree of direction; then compare presence, latency and control on the same lines, before validating load, rights and fallback over a full session.
Four layers not to conflate
These cues avoid mistaking a strong voice model for an avatar, or signal detection for performance direction. Each card separates what is documented from what still needs Où est Ava ? or Plastic Dilemma measurement.
Video identity and turn-taking
Creates a character from image or video, declares French on the PAL and steers patience, interruption and re-engagement. Test directed French voice, perception consent/policy and full-scene latency.
Open Tavus profile →Decoupled video rendering
Keeps avatar and video transport with HeyGen while Où est Ava ? retains STT, LLM, RAG, French voice and performance intent. Test voice switching, rights and lip-sync; no documented deterministic live gesture.
Open HeyGen profile →How the line is delivered
Carries free-form direction, non-verbals, cross-turn audio context and Expressive/Balanced/Stable modes. It is a voice layer to pair with a renderer; test French directions, sensitive scenes and repeatability.
Open Inworld profile →French, precision and residency
Targets structured terms, timestamps and stable timing. EU residency can be activated and verified on a paid plan; it is neither Swiss residency nor explicit emotion direction.
Open Gradium profile →Two paths to compare
Commercial services accelerate staging and live-flow testing. Open models require building the capacity but enable discussion of sovereignty, licences and infrastructure cost. Both paths use the same questions of presence, latency, direction and fallback.
Live comparison: transport, rendering, performance direction, pricing, concurrency and provider dependency.
Browse commercial services →Code, weights & GPUOù est Ava ?/Plastic Dilemma reading: licences, VRAM, planning streams, Exoscale rental, CapEx and observed load results.
Browse open-source models →For Où est Ava ?, the question is not how to rank generic clips: it is how to produce a short shot consistent with what was just said — a memory, place, clue or world shift — and insert it at the right moment. A second path explores visual environments that react live without being avatars.
These models produce a standalone shot. They suit reactive editing; they are not a continuous video stream.
ByteDance Seed
Memory, location or clue shot triggered between two exchanges
Official source ↗Versioned insert factory from dialogue context
Official source ↗xAI
Preview or modify an insert while retaining the source shot
Official source ↗These world models generate images as an action, navigation or event unfolds. They open a path to reactive settings or worlds, with their own costs, rights and action limits.
Runway · Early access / to qualify
Real-time frame-by-frame generation
Camera pose, actions and audio; explorable environment
Explored scene or reactive setting, separate from the avatar
Official source ↗Google DeepMind · Research prototype / limited Project Genie access
720p at 20–24 fps; a few minutes of interaction
Navigation and prompt-controlled world events
Research path for a reactive world, not a production building block
Official source ↗World Labs · Research preview
Interactive frame rate targeted on one H100
Interactive frames and spatial persistence from images
Spatial path to cost before any service promise
Official source ↗Decart · Product access to qualify
Lucy: announced live transformation at 30 fps
Live stream transformation or interactive world depending on product
Test a consented live transformation, not a generated narrative insert
Official source ↗Two recent reference points to read separately: market definitions are not additive.
Digital humans · global
Interactive avatars represented 58.63% of the segment in 2025.
Mordor Intelligence ↗AI avatars · United States
Avatar-based customer service represented 32.9% of this market in 2025.
Grand View Research ↗USD billions · published checkpoints, not annual interpolation
USD millions · published checkpoints, not annual interpolation
Digital humans · 2025
Digital humans · 2025
Digital humans · 2025
Reading for GamiWays: segment growth supports exploration, but does not establish the feasibility of Où est Ava ? or Plastic Dilemma. The decision criterion remains evidence of a credible, controllable and economically sustainable field experience.
GamiWays strategic positioning
GamiWays combines AI conversation, video avatar, sequencing and narrative or learning control. Value does not come from an isolated video model: it depends on integration, the chosen sovereignty posture and measurements made in real experiences.