Ostris introduced a new content generation page within the AI-Toolkit interface, combining training and inference workflows in one place. The update lets users train loras and generate media without transferring checkpoints to external tools.

Supported models and media types

The inference engine supports many of the same open models used for training, covering image, video and music generation families. Available visual models include variants such as Krea 2, Flux versions, Qwen Image 2.1, Qwen Image 2512, SDXL and Z-Image Turbo.

• Video models in the interface include MiniMax H3, LTXV 2.3 and later revisions, plus Wan 2.2.
• Music models listed for generation include Ace-Step 1.5XL, MiniMax Music 3 (under testing) and Ye2.

Quantisation and model fusion

The tool provides several quantisation options for transformer and text encoder components, including bf16, mixed integer and floating formats such as fp8 and other vendor formats. It also supports merging a trained lora into a base model to simplify deployment and inference workflows.

Performance and current limitations

At present the feature set is alpha with minimal tuning and known speed limitations compared with existing tools like Comfy. For example, an fp8 Krea 2 run in AI-Toolkit generates 1024x1024 images in 9 steps taking about 13 sec on a 4090 GPU.

By contrast, the same task in Comfy using bf16 completes in 8 steps and roughly 5 sec, while 2048x2048 renders may exceed one minute or fail to complete. Developers note ongoing work to improve stability and runtime performance.

Deployment and tooling

AI-Toolkit can be installed and updated conveniently through community installers and helper scripts, streamlining local setup for experimentation. The team positions the interface as complementary to more mature ecosystems, acknowledging Comfy’s broader flexibility and integrations remain unmatched.