{"id":14464,"date":"2025-08-14T14:21:32","date_gmt":"2025-08-14T13:21:32","guid":{"rendered":"https:\/\/www.keris-studio.fr\/blog\/?p=14464"},"modified":"2025-08-22T10:33:47","modified_gmt":"2025-08-22T09:33:47","slug":"comfyui-ai-for-architecture-case-study-01-render-from-basic-3d","status":"publish","type":"post","link":"https:\/\/www.keris-studio.fr\/blog\/?p=14464","title":{"rendered":"COMFYUI – AI\/Archi01\u00a0: Render from 3D"},"content":{"rendered":"

Task<\/h1>\n

ComfyUI Tutorial for Architecture: From Sketch to Realistic Render with ControlNet Canny<\/strong><\/p>\n

If you want to transform a simple sketch or plan into a detailed and realistic architectural render, ComfyUI, with its modular structure, is the perfect tool for this. In this tutorial, we will explore a simple yet powerful workflow using ControlNet Canny to turn a basic drawing into a high-quality image.<\/p>\n

This tutorial is designed for beginners with ComfyUI. We will walk through each step, from importing models to achieving the final result.<\/p>\n

<\/p>\n

The Step-by-Step Workflow<\/h1>\n

Here is a general overview of the workflow we will use. Each group of nodes represents a logical step in the process.<\/p>\n

Step 1: Load Models<\/h2>\n

The first step is to load all the necessary models and tools for the process to work.<\/p>\n

    \n
  • CheckpointLoaderSimple<\/strong>: This is the core of our workflow. This node loads a Stable Diffusion model. This model understands our text \u00ab\u00a0prompt\u00a0\u00bb and generates the image. For architecture and renders, models like \u00ab\u00a0DreamShaper<\/strong>\u00a0\u00bb or \u00ab\u00a0realisticVision<\/strong>\u00a0\u00bb are excellent choices. This node provides three outputs: the<\/li>\n<\/ul>\n

    MODEL<\/strong> (the base model), the CLIP<\/strong> (the text encoder that interprets our instructions), and the VAE<\/strong> (the decoder that transforms the latent image into the final image).<\/p>\n

      \n
    • VAELoader<\/strong>: Although CheckpointLoaderSimple can load a default VAE, it is often better to load a separate one for higher quality. The VAE (Variational AutoEncoder) is essential for the decoding phase, as it converts the abstract latent space into a readable image.<\/li>\n
    • ControlNetLoader<\/strong>: This node is the key component of our process. It loads a ControlNet model, which will guide the image generation based on an input image. In this case, the \u00ab\u00a0Canny\u00a0\u00bb model is ideal for preserving the contours and lines of our architectural drawing.<\/li>\n<\/ul>\n

      <\/p>\n

      Step 2: Prepare the Input Image<\/h2>\n

      This section handles importing your sketch and processing it so it can be understood by ControlNet.<\/p>\n

        \n
      • LoadImage<\/strong>: This is where you upload your sketch or plan. Choose your file, such as a facade sketch or an interior drawing.<\/li>\n<\/ul>\n

        <\/p>\n

          \n
        • Canny<\/strong>: This node is a \u00ab\u00a0preprocessor\u00a0\u00bb. It takes your input image and extracts its contours and edges. The Canny model is excellent for accurately capturing architectural lines. You can adjust the thresholds (<\/li>\n<\/ul>\n

          low_threshold and high_threshold) to control the fineness of the detected edges. The other two nodes connected to it,<\/p>\n

          <\/p>\n

          GetImageSize+ and JWImageResizeByLongerSide, automatically resize and define the image dimensions, ensuring your base image is suitable for the model.<\/p>\n

            \n
          • PreviewImage<\/strong>: Although not essential for the final result, this node is very useful for visualizing the output of the Canny preprocessor in real time. This way, you can see what the contours that ControlNet will use as a guide look like.<\/li>\n<\/ul>\n

            <\/p>\n

            Step 3: Enter Instructions (Prompt)<\/h2>\n

            This step is crucial for telling the model what you want to generate.<\/p>\n

              \n
            • CLIPTextEncode (positive)<\/strong>: This is your \u00ab\u00a0positive prompt\u00a0\u00bb. Here, you describe in detail the image you want to obtain. Feel free to be very specific about the style (modern, minimalist), materials (wood, concrete), lighting (natural, soft), and atmosphere. Use keywords that evoke well-known references like \u00ab\u00a0architectural digest\u00a0\u00bb or \u00ab\u00a0professional photography\u00a0\u00bb.<\/li>\n
            • CLIPTextEncode (negative)<\/strong>: This is your \u00ab\u00a0negative prompt\u00a0\u00bb. It is used to tell the model what you do not want to see in the result. For example, rendering flaws (\u00ab\u00a0blurry\u00a0\u00bb, \u00ab\u00a0low quality\u00a0\u00bb), undesirable colors (\u00ab\u00a0oversaturated colors\u00a0\u00bb), or non-architectural elements (\u00ab\u00a0cartoon\u00a0\u00bb, \u00ab\u00a0anime\u00a0\u00bb).<\/li>\n<\/ul>\n

              <\/p>\n

              Step 4: The Generation Process (KSampler & Final Output)<\/h2>\n

              This last part assembles all the elements to create the final image.<\/p>\n

                \n
              • ControlNetApplyAdvanced<\/strong>: This node receives the positive and negative prompts, as well as the contours of your image (output of the Canny node). It applies the ControlNet model to combine the influence of your prompts with the information from your initial drawing. The outputs of this node are the \u00ab\u00a0conditions\u00a0\u00bb for our process, which tell the KSampler how to behave.<\/li>\n<\/ul>\n

                <\/p>\n

                  \n
                • EmptyLatentImage<\/strong>: This node creates an \u00ab\u00a0empty\u00a0\u00bb image in the latent space, which will serve as the starting point for generation. The dimensions of this image are automatically defined by the<\/li>\n<\/ul>\n

                  In this example, we resize the image according to the input image.<\/p>\n

                  GetImageSize+ node to match your input image.<\/p>\n

                    \n
                  • KSampler<\/strong>: This is the generation engine. The KSampler takes the model, the conditions (prompts), and the empty latent image to generate a new latent image, following the given instructions. You can adjust key parameters here such as the seed (for reproducibility), the steps (quality), and the cfg (the level of prompt adherence).<\/li>\n<\/ul>\n

                    <\/p>\n