Kandinsky 3.1 is a text-to-image diffusion model designed for high-quality, realistic image generation with enhanced features. It targets researchers and power users seeking advanced control and efficiency in AI-driven visual content creation, offering improvements over its predecessor, Kandinsky 3.0.

How It Works

Kandinsky 3.1 builds upon a latent diffusion architecture, incorporating a Flan-UL2 text encoder and a large U-Net. A key innovation is Kandinsky Flash, a distilled model using Adversarial Diffusion Distillation on latents for significantly faster inference (4 steps) without quality degradation. It also features prompt beautification via an LLM (Intel's neural-chat-7b-v3-1) and integrates IP-Adapter and ControlNet for image-conditional generation.

Quick Start & Requirements

• Install via pip install -r requirements.txt after creating a conda environment.
• Requires CUDA 11.1+ and PyTorch 1.10.1.
• Example usage provided in ./examples Jupyter notebooks.
• Official HuggingFace repository and project page links are available.

Highlighted Details

• Kandinsky Flash offers 4-step inference, 3x faster than the base model.
• Integrates prompt beautification using an LLM for improved text-to-image results.
• Supports inpainting, image fusion, and image variations.
• IP-Adapter and ControlNet enable image-conditional generation.

Maintenance & Community

The project is actively developed by a team including Vladimir Arkhipkin, Anastasia Maltseva, Andrei Filatov, and Igor Pavlov. Links to HuggingFace and a Telegram bot are provided for community engagement.

Licensing & Compatibility

The model weights are released under a permissive license, allowing for commercial use and integration into closed-source applications.

Limitations & Caveats

The initial installation instructions specify PyTorch 1.10.1+cu111, which is an older version and may require careful dependency management for compatibility with newer CUDA toolkits.