TA Toolbox Home / Tool Tutorial
AI Painting Online Text-to-Image Tool Tutorial
Choosing the Right Online AI Painting Tool
Currently, mainstream online AI painting text-to-image tools include:
- **Midjourney**: Known for artistic styles, requires Discord, high generation quality.
- **Stable Diffusion WebUI**: Open-source and free, can be deployed locally or used on hosted platforms, highly customizable.
- **DALL·E 3**: Developed by OpenAI, integrated into ChatGPT, good at understanding natural language.
- **Wenxin Yige, Tongyi Wanxiang**: Chinese tools supporting Chinese prompts, easy access.
For beginners, it is recommended to start with **online platforms** (such as Jimeng, LiblibAI) that require no environment setup, just registration.
Writing Effective Prompts
Prompts are the key to text-to-image. The basic structure is: **Subject description + Style/Artist + Details/Lighting + Quality tags**.
Example:
`A girl wearing Hanfu, standing under a cherry blossom tree, beautiful ancient style, soft lighting, high quality, rich details, 8k`
Tips:
- Use specific nouns and adjectives, avoid vague expressions.
- Specify styles, such as "cyberpunk", "watercolor", "oil painting".
- Add negative prompts to exclude unwanted content, like "low quality, blurry, deformed".
- Learn from prompts of excellent works.
Setting Generation Parameters
Parameters vary slightly among tools, common ones include:
- **Image size**: Affects composition and detail; common sizes are 512x512, 768x1024, etc.
- **Sampling steps**: Generally 20-30 steps are enough; too high may overfit.
- **CFG Scale**: Controls prompt adherence; 7-10 is commonly used.
- **Seed**: Fixing the seed can reproduce the same image.
- **Model selection**: Different models excel in different styles; can switch and try.
Steps for Online Text-to-Image
- Open the website of the chosen tool (e.g., Jimeng: https://jimeng.jianying.com).
- Register/Log in.
- Find the "Text to Image" feature entry.
- Enter prompts in the text box; negative prompts can be added.
- Set image size, quantity, style, and other parameters.
- Click "Generate" and wait from a few seconds to a few minutes.
- Preview results and download/save the satisfactory image.
- If results are unsatisfactory, adjust prompts or parameters and regenerate.
- **Iterative optimization**: Quickly generate low-step drafts to lock composition, then increase steps and quality.
- **Reference image feature**: Some tools support uploading reference images to achieve style transfer or character consistency.
- **Inpainting**: Redraw unsatisfactory areas to improve details.
- **Batch generation**: Generate multiple images at once and pick the best.
- **Community learning**: Browse tool communities to learn prompts and parameters from others.
Advanced Tips and Tricks
FAQ
Why are the images I generate blurry?
Possible reasons: image size too small, sampling steps too low, or model quality not high. It is recommended to increase size, increase steps to around 30, and select high-quality models.
Should I use Chinese or English for prompts?
Most foreign tools support English better, but Chinese tools like Wenxin Yige and Tongyi Wanxiang support Chinese. It is recommended to use English prompts mainly, with the help of translation tools if needed.
How can I fix the same character in different poses?
Use the character reference image feature, or use the same seed value and fine-tune prompts. Training a dedicated LoRA model is also possible.
Do generated images have copyright?
Different tools have different regulations. Generally, users have the right to use generated images, but commercial use requires attention to platform terms; some open-source models allow commercial use.
Is there a big difference between free and paid online tools?
Paid tools usually offer higher generation quality, faster speed, and more features. Free tools may have usage limits or watermarks. Choose according to your needs.