omnichar/OmniChar

What it solves

Omnichar Studio provides a unified environment for creating and maintaining consistent AI characters across different generative models. It eliminates the need to manually re-wire reference photos or train separate LoRAs for every single shot, allowing users to define a character once and reuse them consistently across various images and videos.

How it works

The system uses a portable .char file that stores a character's references and an identity fingerprint. Depending on the model being used, the studio handles the likeness differently:

• FLUX.2: Uses references directly in the prompt's token sequence without training.
• Krea 2: Trains a small, reusable adapter for the character.
• MiniMax H3: Supports both direct reference-to-video and trained adapters.
• Hosted Models: Sends references and locked descriptions via API.

Additionally, the studio includes a built-in LoRA trainer that allows users to train on images or video clips locally on their own GPU, with the resulting .safetensors files integrated directly into the generation workflow.

Who it’s for

It is designed for AI artists and creators who need high visual continuity for characters across multiple frames, shots, or different AI models (image and video), and those who prefer to run their generation and training pipelines locally on their own hardware.

Highlights

• Portable Characters: Define a character once in a .char file and use them across multiple models.
• Integrated LoRA Training: Train on stills or video clips locally on your GPU with integrated monitoring and loss curves.
• Multi-Model Support: Compatible with FLUX.2, FLUX.1, and Krea 2 for images, and MiniMax H3 and LTX-2.5 for video and sound.
• Continuity Scoring: Provides a numerical continuity score for video takes to track identity drift.
• Low-VRAM Options: Includes Z-Image Turbo for low-memory generation and specific profiles for tight-VRAM hardware.