How to Use It

1. Download the required LoRA model: Download the TileXL LoRA model from https://civitai.com/models/1455220/tile-xl and upload it into your project via the Model Manager.
2. Download and import the JSON workflow: Download the .JSON file of this workflow. On your Workflows tab, click "Load Workflow from File" and select the downloaded .JSON file.
3. Configure your White Template Tile (download below) and your ControlNet Tile (download below).
4. Select your base model: Choose a model compatible with SDXL, such as juggernautXL. This provides the foundational aesthetic and generation capabilities.‍
5. Enter your prompt: In the User Prompt field, describe what you want to generate (e.g., “volcano++, magma, mobile game style, illustrative”). This will be dynamically framed by a predefined tile art style prefix and isometric suffix for consistency.‍
6. Adjust style strength: Use the Tile Style Strength slider to set how strongly the LoRA (Low-Rank Adaptation) influences the style. Suggested range: 0.3 to 0.9.‍‍
7. Fine-tune ControlNet influence: If using structural guidance (like edge maps or pose inputs), you can tweak ControlNet Weight from 0 to 1 to control how much the structure input affects the output.‍‍
8. Invoke the workflow: Click ‘Invoke’ to generate your image. The workflow automatically combines prompt conditioning, noise generation, style mixing, and structural guidance behind the scenes.

How It Works

1. Model Initialization: The workflow starts by loading the selected SDXL base model (e.g., juggernautXL), initializing it with UNet, CLIP, and VAE components for text-to-image generation.
2. Prompt Construction: A final prompt is built by joining three parts: a fixed style prefix (“T1L3 Style”), the user’s custom input (e.g., “volcano++, magma”), and a fixed suffix (“2.5d, isometric”) to ensure consistent visual framing.
3. Prompt Conditioning: The constructed prompt is processed using the sdxl_compel_prompt node to generate positive prompt embeddings. A corresponding negative prompt suppresses undesired traits (e.g., “flat, blurry, sketch”).
4. Style Injection (LoRA): A LoRA adapter trained for tile-style aesthetics is selected and applied with adjustable strength. The adapter modifies the base model’s UNet and CLIP layers to infuse the image with a specific stylistic character.
5. Structure Guidance (optional): If enabled, a ControlNet layer processes a structural image input (such as a sketch or edge map). The user-defined control weight determines how much this structure influences the final composition.
6. Latent Noise Generation: A random seed initializes a latent tensor at 1024×1024 resolution. This noise acts as the starting point for image generation.
7. Latent Denoising: The denoising process begins, guided by prompt embeddings, LoRA-adapted model weights, structural controls, and latent noise. It runs for 30 steps using the DPM++ 3M K scheduler, producing a coherent latent representation.‍
8. Image Decoding & Output: The denoised latents are decoded into a final image using the VAE. The output is rendered to the canvas and can be saved to the gallery for further use.

‍

White Template Tile image:

Controlnet Tile image: