返回知识库
模块 11

Advanced Specialization

Character consistency, LoRA training, and hybrid character systems

The course presents a structured approach to prompt engineering for character consistency, primarily centered around the following framework:

[Identity Anchor] + [Scene Description] + [Technical Specs] + [References]

  • Identity Anchor: This is the core part of the prompt that defines the character's unchangeable features. It can be a detailed physical description, a named reference (e.g., 'Our hero character Alex'), or a callback phrase (e.g., 'The same young woman from before').
  • Scene Description: This part of the prompt describes the environment, action, and mood without redefining the character's core features. It should focus on what the character is doing and where they are.
  • Technical Specs: This includes parameters that control the image generation process, such as aspect ratio (--ar), model version (--v), and style (--style).
  • References: This involves using image references (--cref) and style references (--sref) to guide the AI. The character reference (--cref) is used to maintain facial and physical identity, while the style reference (--sref) is used to control the artistic style of the image.

Another prompt structure mentioned is the use of clear separation for different elements to avoid 'context corruption':

CHARACTER: [character description] SCENE: [scene description]

For LoRA models, the prompt structure is:

a photo of lora:character_v1:0.8 [blocked] in a park

Where <lora:character_v1:0.8> is the LoRA identifier with a specific weight.