important: Based on testing, when using the Qwen 2.1 LoRA, the CFG should not be lower than 3.0, otherwise the quality will be very poor. CFG 4.0 is recommended. Do not use acceleration LoRAs, or it will not work.
经测试使用Qwen 2.1 版本的Lora，CFG不能低于3.0，否则会非常劣质。推荐CFG4.0.不能用加速lora，否则会失效，
V1: Trained on face data from the debut trailer with region-focused enhancements on the face. Note that the character's facial model has been slightly updated in recent builds—if you prefer the latest look, go with V2 instead.
important：The V2 dataset uses face captures from the latest in-game gameplay reveal. Like the previous version, it relies on regional training with a dedicated focus on facial details. Adding keywords like "GTA style," "3D game rendering," or "Unreal Engine 5" to your prompt will get you much closer to the in-game aesthetic. Testing shows that no trigger words are required—natural language prompts work best. For optimal results, avoid overly detailed character feature descriptions.
Description
The V2 dataset uses face captures from the latest in-game gameplay reveal. Like the previous version, it relies on regional training with a dedicated focus on facial details. Adding keywords like "GTA style," "3D game rendering," or "Unreal Engine 5" to your prompt will get you much closer to the in-game aesthetic. Testing shows that no trigger words are required—natural language prompts work best. For optimal results, avoid overly detailed character feature descriptions.
FAQ
Comments (2)
Good Thanks
Great Lora , big Thanks