How to Use Stable Diffusion for High‑Quality Promotional Images

Platform

Current AI drawing mainly uses Stable Diffusion (SD), a diffusion model derived from DDPMs, capable of text‑to‑image, image‑to‑image, and post‑processing, often replacing Photoshop.

Three Ways to Use SD

Build your own web UI. The popular project AUTOMATIC1111/stable-diffusion-webui on GitHub offers high flexibility and the latest updates, but requires environment setup, a VPN for Chinese users, and may be unstable.

Use web‑based services built on stable‑diffusion‑webui. Examples include LiblibAi (https://www.liblib.art/) and MJ (https://mj.wxcbh.cn/home/). These platforms handle maintenance and provide stable connections; they usually charge for usage but do not require a VPN.

Install desktop software. Draw Things, available for free on the Mac App Store, runs locally without a VPN and provides a graphical interface for SD models.

Draw Things Interface

The interface includes sections for model selection, sampler, steps, random seed, and other parameters. Users start by choosing the "text‑to‑image" mode.

Model (Checkpoint)

Checkpoints are the core SD models, typically sized 2‑7 GB. Different versions (V1.0, V1.5, V2.0, V2.1) produce different styles. Models can be downloaded directly in Draw Things or imported from sites such as LiblibAi.

Lora

Lora models are lightweight fine‑tuned extensions (tens to hundreds of MB) that add specific styles or subjects. Multiple Loras can be applied on top of a base model. Their files also use the safetensors suffix and must be placed in the designated folder.

Hypernetwork

Similar to Lora but generally less effective; it must be used together with a base model.

Sampler

Samplers affect image style and quality. Common choices include DDIM, PLMS, DPM, DPM++, Euler‑a, and the Karras series. Euler‑a works well for icons and simple scenes, while DPM/DPM++ excel at realistic portraits. Karras samplers improve overall quality.

Steps

Steps refer to the number of denoising iterations. For quick tests, 10‑15 steps are sufficient; 20‑30 steps give balanced quality; 40+ steps increase time with diminishing returns, useful for detailed textures like fur or skin.

Random Seed

The seed controls the initial noise. Seed = ‑1 generates a new random seed each run. Specifying a seed reproduces similar images, useful for consistency.

Prompt

Prompts are the key to image content, style, and composition. Positive prompts describe desired elements, while negative prompts exclude unwanted features (e.g., low quality, NSFW, bad eyes, extra fingers).

Example prompt for a "618 Watermelon Promotion" image:

Many people eat watermelons by the flowing water in the lake, high‑quality miniature photography

. Weighting can be applied by enclosing a term in parentheses and adding a factor, e.g., (watermelon:1.5) to emphasize watermelons.

Other Parameters

Additional settings include image resolution and aspect ratio, text guidance strength, and batch size.

Post‑Processing (Secondary Editing)

After generation, the "image‑to‑image" mode allows refinement. Adjust the strength: 0 keeps the original image, < 30 % makes subtle tweaks, > 70 % produces major changes, and 1 creates a completely new image.

Redraw Elements

Use the eraser tool to remove unwanted parts, then let the model redraw them, often achieving better results than traditional Photoshop erasing.

Image Expansion

Increase canvas size, move the original image to one side, erase the edge where expansion is desired, and let the model generate new content to fill the expanded area.

Upscaling

Start with a low resolution for speed, then increase resolution and re‑render to improve clarity. Existing images can also be upscaled using the same workflow.

By following these steps, you can create high‑quality AI‑generated promotional graphics for events such as the 618 sales campaign.