Work / Case study
Custom models that keep a character the same in every image, call and video
Overview
Unscripti.ai is a mobile app for AI character chat, voice calls and content. Off-the-shelf models could not keep a character's face, voice and style the same from one output to the next. tMinus1 built the model stack and the content pipeline behind the product.
General image models changed a character’s face between generations and cost too much at consumer scale.
A character app only works if the character stays the same person. General image models changed faces between generations, ignored poses and outfits, and cost too much per image at consumer scale. Social content for each character was made by hand, one video at a time.
A custom model stack on serverless GPUs: identity control, pose and depth ControlNets, staged repair and an AI content engine.
A custom image model stack served on serverless GPUs: an in-house SDXL-based checkpoint with a 4-step distilled LoRA for speed, identity control with InstantID and IP-Adapter, and ControlNets for pose, depth and detail. A second pipeline runs pose, outfit, identity, refinement and repair as separate stages. Language models were fine-tuned to keep each character’s voice consistent in chat. A content engine writes chat-story scripts with Claude, voices them with ElevenLabs or OpenAI and renders vertical 1080x1920 video with Remotion.
We started with the hardest requirement, identity, and measured it on every change. Models are versioned on Hugging Face and shipped as Docker workers on RunPod serverless, with a cold-start target under 5 seconds and 30–60 seconds per generation. Each pipeline stage has its own settings, so one weak stage can be tuned without retraining the rest.
Consistent characters across chat, art and video, from one model stack.
One model stack serves chat images, character art and short-form video, with each character’s identity held across all three. Content that was made by hand, one video at a time, now comes from an automated pipeline.
Content SafetyAdult content platformAge assurance and an automated safety gate for generated media.
AI EngineeringIropeakFrom AI idea to a working product. Scoped, built and handed over.
AI TrainingLimitless ConsultingHands-on AI training on real client work, not toy examples.
Cloud ArchitectureAutomatum80+ ISVs live. $120M+ transacted. Built and operated by tMinus1.
AWS Cloud SecurityAssignarZero-downtime AWS migration with confirmed tenant isolation and 28% cost reduction.
AWS Cloud SecurityInduction HealthcareEight-week security uplift that cleared NHS enterprise procurement on the first submission.
Cloud ArchitectureCSIROElastic AWS Batch environment that cut queue waits 60% and freed $1.2M in annual compute budget.
Cloud ArchitectureEnergy Exemplar90-day cloud migration for a global energy simulation platform — zero incidents, 40% faster solvers.
AWS Cloud SecurityAirtasker$340K in annual AWS spend recovered in four weeks — no re-architecture, all as Terraform.
Unscripti.ai A tMinus1 case study
Request a technical call