Tweet
After generating probably 10,000+ images, here's what actually separates good AI art from generic output.
Prompt structure that works:
Most people type "cool dragon" and wonder why it's boring. Here's the formula: subject + style + lighting + composition + quality markers.
Bad: "dragon" Good: "close-up of ancient dragon's eye, scales reflecting golden light, fantasy art style, dramatic side lighting, intricate detail, trending on ArtStation, 8K"
You're giving style direction, lighting control, composition framing, and quality indicators. The AI needs this specificity.
Which tool for what:
Midjourney – when you want it to look artistic and polished. It makes everything prettier. Responds well to photography terms like "shot on Hasselblad" or "35mm film" even for non-photo subjects. Best for concept art, characters, landscapes.
DALL-E 3 – when you need text in the image or conversational prompting. "Create a coffee shop logo called 'Morning Brew' with sunrise and warm colors, text in elegant script" – it just does it. Best for logos, infographics, anything with text.
Stable Diffusion – when you need control. Open-source, runs locally, train custom models, fine-tune everything. Requires technical setup with Automatic1111 or ComfyUI. Best for consistent characters, custom styles, production pipelines.
Flux – newer model, scary good at photorealism. Best for realistic product shots, portraits, anything that should look like a photograph.
Artist reference shortcut:
Want a specific aesthetic fast? Reference artists. "In the style of Greg Rutkowski" = epic fantasy. "James Gurney style" = realistic illustration. "Alphonse Mucha" = Art Nouveau.
Browse ArtStation to discover artists, reference them in prompts. Yes, this is controversial (training data ethics), but it's how most people use these tools. At least be aware of the questions.
Lighting is everything:
"Golden hour lighting" vs "harsh fluorescent" = completely different mood with same subject. Learn terms: Rembrandt lighting, rim lighting, backlighting, soft diffused light. This lighting guide teaches terms AI understands.
Character consistency:
Getting the same character across images is hard. What works:
Method 1: Extreme detail. "25-year-old woman, 5'6", shoulder-length auburn hair with bangs, green eyes, small scar above left eyebrow, heart-shaped face" – more specific = more consistent.
Method 2 (Midjourney): Use --cref [image URL] to reference previous character image.
Method 3 (Stable Diffusion): Train a LoRA on 20-30 images. Technical but perfect consistency. Civitai has guides.
Quality workflow:
Generate at max resolution → Upscale with Topaz Gigapixel ($99) or Upscayl (free) → Touch up in Photoshop → Export.
Professional reality: generate 20-50 variations, pick best 2-3, upscale, refine. First attempt is rarely your best. Midjourney generates 4 at a time – use the variation feature.
Prompt resources:
PromptHero – 10M+ prompts with results. Lexica.art – Stable Diffusion specific. Midjourney Community Feed – see what's being made with prompts shown.
Don't reinvent. Find prompts for your desired style, adapt them.
What still breaks:
Hands – fingers multiply or bend wrong. Plan to fix in Photoshop. Complex multi-character scenes – AI interprets creatively (wrong). Exact brand colors – color correct in post. Text (except DALL-E) – most models garble it.
Midjourney parameters:
--ar 16:9 = aspect ratio --stylize 500 = artistic interpretation (0-1000) --chaos 50 = variation amount (0-100) --v 6 = use version 6
Full list: Midjourney docs
Stable Diffusion tips:
Negative prompts are critical: "ugly, distorted, low quality, blurry, bad anatomy, watermark"
Emphasis syntax: (keyword:1.3) increases weight
Custom models on Civitai for specific styles – anime, photorealism, architecture, whatever.
Commercial use:
Midjourney: commercial rights with paid plan. DALL-E: commercial use allowed. Stable Diffusion: check specific model license. Copyright protection: legally unclear, courts are figuring it out.
Share your images with full prompts. Post what worked and what didn't. Ask why something isn't working. When you discover a technique, share it. We're learning together.
Prompt structure that works:
Most people type "cool dragon" and wonder why it's boring. Here's the formula: subject + style + lighting + composition + quality markers.
Bad: "dragon" Good: "close-up of ancient dragon's eye, scales reflecting golden light, fantasy art style, dramatic side lighting, intricate detail, trending on ArtStation, 8K"
You're giving style direction, lighting control, composition framing, and quality indicators. The AI needs this specificity.
Which tool for what:
Midjourney – when you want it to look artistic and polished. It makes everything prettier. Responds well to photography terms like "shot on Hasselblad" or "35mm film" even for non-photo subjects. Best for concept art, characters, landscapes.
DALL-E 3 – when you need text in the image or conversational prompting. "Create a coffee shop logo called 'Morning Brew' with sunrise and warm colors, text in elegant script" – it just does it. Best for logos, infographics, anything with text.
Stable Diffusion – when you need control. Open-source, runs locally, train custom models, fine-tune everything. Requires technical setup with Automatic1111 or ComfyUI. Best for consistent characters, custom styles, production pipelines.
Flux – newer model, scary good at photorealism. Best for realistic product shots, portraits, anything that should look like a photograph.
Artist reference shortcut:
Want a specific aesthetic fast? Reference artists. "In the style of Greg Rutkowski" = epic fantasy. "James Gurney style" = realistic illustration. "Alphonse Mucha" = Art Nouveau.
Browse ArtStation to discover artists, reference them in prompts. Yes, this is controversial (training data ethics), but it's how most people use these tools. At least be aware of the questions.
Lighting is everything:
"Golden hour lighting" vs "harsh fluorescent" = completely different mood with same subject. Learn terms: Rembrandt lighting, rim lighting, backlighting, soft diffused light. This lighting guide teaches terms AI understands.
Character consistency:
Getting the same character across images is hard. What works:
Method 1: Extreme detail. "25-year-old woman, 5'6", shoulder-length auburn hair with bangs, green eyes, small scar above left eyebrow, heart-shaped face" – more specific = more consistent.
Method 2 (Midjourney): Use --cref [image URL] to reference previous character image.
Method 3 (Stable Diffusion): Train a LoRA on 20-30 images. Technical but perfect consistency. Civitai has guides.
Quality workflow:
Generate at max resolution → Upscale with Topaz Gigapixel ($99) or Upscayl (free) → Touch up in Photoshop → Export.
Professional reality: generate 20-50 variations, pick best 2-3, upscale, refine. First attempt is rarely your best. Midjourney generates 4 at a time – use the variation feature.
Prompt resources:
PromptHero – 10M+ prompts with results. Lexica.art – Stable Diffusion specific. Midjourney Community Feed – see what's being made with prompts shown.
Don't reinvent. Find prompts for your desired style, adapt them.
What still breaks:
Hands – fingers multiply or bend wrong. Plan to fix in Photoshop. Complex multi-character scenes – AI interprets creatively (wrong). Exact brand colors – color correct in post. Text (except DALL-E) – most models garble it.
Midjourney parameters:
--ar 16:9 = aspect ratio --stylize 500 = artistic interpretation (0-1000) --chaos 50 = variation amount (0-100) --v 6 = use version 6
Full list: Midjourney docs
Stable Diffusion tips:
Negative prompts are critical: "ugly, distorted, low quality, blurry, bad anatomy, watermark"
Emphasis syntax: (keyword:1.3) increases weight
Custom models on Civitai for specific styles – anime, photorealism, architecture, whatever.
Commercial use:
Midjourney: commercial rights with paid plan. DALL-E: commercial use allowed. Stable Diffusion: check specific model license. Copyright protection: legally unclear, courts are figuring it out.
Share your images with full prompts. Post what worked and what didn't. Ask why something isn't working. When you discover a technique, share it. We're learning together.