>>105936356
ROUWEI-GEMMA

URL: https://civitai.com/models/1782437/rouwei-gemma

New text encoder adapter for SDXL image generation models that replaces the traditional CLIP encoders with a Gemma-3-1b language model.

Improve SDXL's ability to understand complex prompts by using an LLM instead of CLIP text encoders, allowing for better prompt adherence and support for longer prompts (up to 512 tokens).

>Focus:
Specifically designed for anime-style image generation models (Rouwei) without censorship restrictions.

>Current Capabilities:
Processes booru tags, natural language prompts, and structured formats (markdown, XML, JSON)
Better understanding of long and complex prompts
Reduces "tag bleeding" between different prompt elements
Current Limitations:

>Still experimental/proof of concept
Inconsistent character/style recognition
Cannot generate quality text in images
No support for prompt weights or emphasis yet
Technical Details:

>Requires ComfyUI with custom nodes
Works with Rouwei checkpoints
Installation involves downloading the adapter and Gemma-3-1b model
Future Plans: Further training of both the LLM and UNet components, improved custom node functionality, and release of training code.