OminiControl

Overview

OminiControl is a versatile framework designed to unify control mechanisms in diffusion-based image generation models. It excels in both subject-driven and spatially aligned tasks, such as inpainting, colorization, and depth-aware synthesis. Built on the Diffusion Transformer (DiT) architecture, OminiControl is lightweight, adding only 0.1% additional parameters to the base model.

It introduces a unified mechanism for integrating image conditions directly into DiT models without additional encoders or heavy modifications. By using LoRA (Low-Rank Adaptation) weights for fine-tuning, it maintains computational efficiency while enhancing model capabilities. OminiControl's unique parameter reuse strategy allows it to inject control signals effectively, making it ideal for tasks requiring subject fidelity, realistic modifications, or detailed alignment.

One of its standout features is the "Subjects200K" dataset, containing over 200,000 images designed for identity-consistent generation. This supports robust training for personalized and identity-preserving outputs.

It requires a setup with significant GPU capacity, typically 30GB VRAM or more for local deployment.

Limitations

1. The model's subject-driven generation primarily works with objects rather than human subjects due to the absence of human data in training.
2. The subject-driven generation model may not work well with FLUX.1-dev.
3. The released model currently only supports the resolution of 512x512, but there are plans to release a 1024x1024 version