Tilly Norwood is an AI-generated actress created by Eline Van der Velden and her company, Xicoia, sparking significant debate in Hollywood regarding the future of acting and AI’s role in the industry.
Tilly Norwood is a synthetic, AI-generated performer introduced in 2025 by Xicoia (the AI/talent arm of Particle6). Marketed like a rising screen star — with an Instagram, publicity shots, a satirical sketch called AI Commissioner, and a Zurich Film Festival debut.
Particle6/Xicoia have not released an engineering paper with every component, but reporting and the nature of current generative tools let us reconstruct a plausible production pipeline.
Base imagery: generative image models (reporting + inference)
What you see in Tilly’s headshots and short scenes likely began with high-quality image generation using diffusion or GAN-style models trained to produce photoreal human faces (the same family of models behind Stable Diffusion, Imagen, Midjourney, etc.). These models produce the “still” appearance and a range of poses and expressions. Reporting states multiple AI tools were used to create visuals.
(Inference): to achieve a consistent identity across many images, teams typically fine-tune a diffusion model on a curated set of outputs (or use a “tokenized identity” method) so that prompts reliably produce the same face and hair, with style/lighting variations.
Motion & video: neural rendering + temporal coherence (inference)
Short videos showing Tilly walking red carpets or delivering lines require generating temporally coherent frames. That can be done by:
- audio-driven facial animation systems (e.g., Wav2Lip, audio-to-expression models) to sync mouth shapes to dialogue, and
- 3D-aware neural rendering (NeRF variants or 3D-conditioned generative models) to keep head geometry consistent across frames, plus optical-flow or temporal diffusion refinements to avoid flicker.
Reports mention criticisms of uncanny valley artifacts (odd mouth movement, blurring) — a tell-tale sign of heavy generative postprocessing rather than pure motion capture.
Voice: cloning or synthetic TTS (inference + common practice)
If Tilly speaks, that voice can be produced by neural TTS or a voice-cloning stack (Tacotron/VITS style systems or bespoke commercial TTS). Because Xicoia presented a British-accented persona, they likely selected or synthesized a voice to match the visual persona. (News coverage also references scripted satire rather than real-time conversational autonomy.)
Compositing, quality fixes, and human oversight (reported)
Journalists repeatedly note that Tilly’s content was produced and curated by humans and that the project involved many tools and edits — not an autonomous AI “actor” making choices on set. That human-in-the-loop stage includes manual cleanup (teeth/mouth fixes, eye artifacts), color grading, and attention to realism.
Xicoia/Particle6 presented Tilly as a product of a suite of AI tools; reporting specifically mentions a sketch created with around ten tools and a script written with ChatGPT
Software used to create: Particle 6 sketch
ChatGPT: Used to generate, iterate, and polish the script
Veo3: Character + Voice creation – text-to-video & frames-to-video generations
Seedance : Image-to-video generations
Imagen 4: Character generations (stills)
Runway + Flux Character variations/copies + environment variations
Elevenlabs: Voice copies + fixes
Adobe: Podcast Audio fixes
Premiere: Pro Editing + grade
Topaz Bloom: Image
Topaz Video AI: Video