Boost Your Design Workflow: Mastering AI Image Generation with Stable Diffusion, Midjourney, and Vega

Recently there has been a flood of articles about AI creation, causing anxiety. Like many designers, I adopted AI to assist design production and share my applications and thoughts.

AI can generate impressive images, but still has limitations: it cannot fully understand aesthetic or cultural differences, leading to style, tone, or anatomical defects (e.g., elongated necks, extra fingers). Designers must use AI cautiously, relying on judgment and aesthetic ability.

Pros and Cons of Three Image‑Generation AIs

1. Stable Diffusion – advantage: can be deployed locally, customizable models, no copyright risk; disadvantage: requires powerful hardware, many Mac models cannot run it.

2. Midjourney – advantage: user‑friendly, high‑quality output, no copyright risk; disadvantage: not customizable, controllability is weaker, struggles with some personified IP.

3. Vega – a Chinese version of SD, easy to use, customizable, stable output with large material feed, but overall quality lags behind the first two and still has some copyright concerns.

By leveraging the strengths of each platform while avoiding their weaknesses, we can improve production.

The key points in AI‑assisted creation are improving controllability and output quality .

Controllability represents the efficiency of the creation process, while quality reflects the satisfaction of the final result.

01. How to improve controllability – break down into:

Master basic commands

Apply complex commands

Select appropriate training images

Accumulate commands

Iterate and optimize

02. How to improve output quality – break down into:

Build training and generation libraries

Combine platforms

Extract elements locally

Fine‑tune adjustments

Consolidate the generation library

Improving Controllability

1· Master basic commands – focusing on Midjourney, learn common prompts (/imagine, /settings, /describe, /blend), description word usage, and parameter values.

2· Apply complex commands

1) Leverage existing community prompts to discover style patterns and extract keywords for expansion.

2) Use “prompt generators” that describe quality, style, author, lighting, composition, etc., to assist when a style is hard to achieve.

3· Choose suitable training images – recommend several resource sites for gathering material.

4· Accumulate commands – organize collected prompts in a personal “second brain” (e.g., Notion) for future reuse.

5· Multiple attempts – because AI cannot be fully controlled, invest time to tweak keywords, replace or add reference images, and find a personal workflow.

Improving Output Quality

1· Build training and generation libraries – manage collected reference images and generated outputs in a lightweight system such as Egael.

2· Combine platforms & extract elements – use the strengths of each AI: Chinese SD for customizable style, Midjourney for high‑quality rendering, then merge results.

I illustrate the workflow with a Mother’s Day illustration case.

Step 1: Define scene – “mother holding child, Joy offering flowers”.

Step 2: Use Chinese SD to generate the mother‑child composition, then feed that image into Midjourney for refinement.

Step 3: Train Joy’s IP in Chinese SD to create the flower‑offering element.

Step 4: Upscale elements with Upscayl for higher pixel density.

Step 5: Assemble and adjust composition into a draft.

Step 6: Fine‑tune color tone, correct anatomical errors, and adjust lighting and overall harmony.

Step 7: After adjustments, consolidate the final image and add it back to the generation library for future cycles.

Overall, AI drawing dramatically impacts design work by boosting efficiency and expanding creative possibilities, while also introducing professional pressure. Maintaining a positive attitude, continuously learning AI techniques, and adapting to new challenges are essential.

Final encouragement: “God, grant me serenity to accept what I cannot change, courage to change what I can, and wisdom to discern the difference.”