blacksnowskill 

Civitai recently added a new beta toggle for LoRA training called "AI Toolkit Training," claiming it offers "improved quality and flexibility."

As someone who deeply respects Ostris and the actual open-source AI Toolkit repository, I need to clear up some heavy misconceptions. Ostris created a fantastic tool, but Civitai’s implementation of it is half-baked, misleading to the average user, and currently acts as a trap that wastes your Buzz if you are training SDXL models.

Here are the facts about what this toggle actually does, and what Civitai isn't telling you:

If you are training an SDXL model, do not use the AI Toolkit toggle. Civitai’s "Standard" training method (which is based on the heavily community-tested Kohya-ss scripts) is perfectly optimized for SDXL. It has mature optimizer support, proven presets, and predictable outputs.

Using the Ostris AI Toolkit for SDXL on Civitai currently gives you zero new benefits, and in many cases, yields significantly worse and less coherent results. It was not implemented to make your SDXL LoRAs better.

The tooltip promises "flexibility." But if you toggle it on, take a look at the UI. Are there any new advanced settings? No.

The real Ostris AI Toolkit is incredibly flexible, but Civitai has hidden all of its actual killer features. Where is the UI support for:

Concept Sliders: The absolute biggest feature of Ostris. The ability to train directional LoRAs (sliders) is completely missing.

Differential Output Preservation: Crucial for preserving base model knowledge while training specific concepts.

Blank Prompt Preservation & Differential Guidance: Advanced techniques that actually provide the flexibility Ostris is known for.