Getting Started with AI Image Generation Using Stable Diffusion for Promotional Posters

With the rise of large AI models such as ChatGPT and Sora, AI‑driven creative tools have become increasingly accessible, allowing anyone to generate high‑quality images for marketing and promotional purposes.

The core model used for AI drawing is Stable Diffusion (SD), an evolution of Denoising Diffusion Probabilistic Models (DDPMs). Modern SD variants support text‑to‑image, image‑to‑image, and post‑processing functions that can replace many tasks traditionally done in Photoshop.

There are three common ways to use Stable Diffusion:

Set up your own SD‑based WebUI (e.g., https://github.com/AUTOMATIC1111/stable-diffusion-webui ), which offers maximum flexibility but requires environment configuration and, for users in China, a VPN.

Use third‑party web platforms that host SD models, such as LiblibAi ( https://www.liblib.art/ ), MJ ( https://mj.wxcbh.cn/ ), or PlaygroundAI ( https://playground.com/ ), which provide stable connections and do not need a VPN.

Run desktop software that packages SD locally; these programs depend on your computer’s GPU performance but give high customizability and are often free.

This article focuses on the desktop application Draw Things , a free macOS app available from the App Store that requires no VPN. Its interface includes a parameter panel for selecting models, samplers, steps, seeds, and other settings.

Model types include:

Checkpoint : the full SD model (typically 2‑7 GB) that determines overall style; files use the .safetensors extension.

Lora : lightweight fine‑tuned adapters (tens to hundreds of MB) that can be stacked on a base model to add specific styles.

Hypernetwork : similar to Lora but generally less effective and must be used together with a base model.

Samplers affect the diffusion process and image style. Common choices are DDIM, PLMS, DPM/DPM++, Euler‑a, and Karras variants. Selecting the right sampler for a given model and step count is crucial for quality.

The step count (sampling iterations) controls how many denoising passes are performed. For quick tests, 10‑15 steps are sufficient; 20‑30 steps give balanced quality and speed; 40+ steps improve detail at the cost of longer generation time.

The random seed determines the initial noise. Using Seed=-1 generates a new random seed each run, while a fixed seed reproduces the same image.

Prompts are the most important factor: positive prompts describe desired content, while negative prompts (e.g., "low quality", "NSFW", "bad eyes", "extra fingers") tell the model what to avoid. Prompt weighting can be applied with parentheses, e.g., (watermelon:1.5) to emphasize a term.

Additional parameters include image resolution, aspect ratio, text guidance scale, and batch size.

After generating an image, secondary processing can be performed using the "image‑to‑image" mode:

Inpainting : erase unwanted elements with an eraser tool and let the model redraw them, often achieving better results than Photoshop.

Outpainting : expand the canvas, erase the edge to be extended, and generate new content that blends with the original.

Upscaling : increase resolution and clarity by re‑generating the image at higher settings.

By adjusting these settings and iterating, users can produce high‑quality promotional graphics for events such as the 618 sales campaign.