Aulafy
Courses/Generative AI: image, voice and video/Control, references, and LoRA

Control, references, and LoRA

A good course doesn't need spectacular images: it needs consistent, readable images that stay true to the concept. Control matters more than visual impact.

  • Separate style, composition, and content.
  • Use references without copying brands, people, or third-party material.
  • Document LoRA adapters and editing changes.

Visual brief before generating

TerminalCode
{
  "objetivo": "explicar búsqueda híbrida en RAG",
  "formato": "16:9 para portada de lección",
  "estilo": "editorial claro, técnico, sin estética futurista exagerada",
  "elementos_obligatorios": ["documentos", "índice vectorial", "filtro de permisos"],
  "elementos_prohibidos": ["logos reales", "caras reconocibles", "texto privado"],
  "paleta": ["azul petróleo", "naranja suave", "gris claro"],
  "lectura": "de izquierda a derecha"
}

Control tools

  • Fixed seed: compare changes without shifting everything at once.
  • Composition reference: maintains framing and layout.
  • Inpainting: fixes specific areas without redoing the entire image.
  • LoRA: maintains a recurring style or visual domain.
  • Upscale: improves final output after content is approved.

Edit log

TerminalCode
imagen: rag-hibrido-portada.png
modelo_base: FLUX.1-schnell
lora:
  nombre: estilo-editorial-aulafy
  peso: 0.55
referencias:
  composicion: boceto-propio-001.png
ediciones:
  - zona: "texto ilegible"
    tecnica: "inpainting"
  - zona: "icono de documento"
    tecnica: "regeneracion parcial"
aprobado_para: "curso gratuito"