STABLE DIFFUSION-

图像生成训练

STABLE DIFFUSION-

IMAGE GENERATION TRAINING

Stable diffusion是一种基于潜在扩散模型（Latent Diffusion Models）的文本到图像生成模型，能够根据任意文本输入生成高质量、高分辨率、高逼真的图像。 面对概念建筑的一定风格元素需求，我们训练了特定的Lora模型。通过收集大量目标图片将LoRA模型的参数注入（inject）SD模型，从而改变SD模型的生成风格，以下是lora模型公式： W=W0+BA 我们使用带GUI的工程kohya_ss，为图片打上标签（caption），输入图片的位置，选择Basic Caption，在Prefix项加上元素关键词，保证描述和元素的对应。 通过特定意向图片的选取融合，配合关键词描述得到了相对有指向型的建筑形制。
Stable diffusion is a text-to-image generation model based on Latent Diffusion Models, capable of generating high-quality, high-resolution, and highly realistic images based on arbitrary text inputs. To meet the specific stylistic elements required for conceptual architecture, we trained a specific Lora model. By collecting a large number of target images, we injected the parameters of the Lora model into the SD model, thereby altering the generation style of the SD model. The formula for the Lora model is as follows: W = W0 + BA We utilized the kohya_ss engineering project with a GUI to label the images by adding captions. We provided the image location, selected the Basic Caption option, and added relevant keywords in the Prefix field to ensure correspondence between the description and the architectural elements. By merging selected images with specific intentions and incorporating keyword descriptions, we obtained relatively directed architectural forms.