Key Takeaways: Chapter 18 — Image Generation — Midjourney, DALL·E, and Stable Diffusion

  • Diffusion models generate images by iteratively de-noising random visual noise, guided by your prompt and the model's learned understanding of text-image relationships. Each generation starts from different random noise, which is why the same prompt produces different images each time — and why generating multiple versions and selecting is the standard workflow.

  • The three major platforms make different tradeoffs: Midjourney optimizes for aesthetic quality and rewards learning its parameter system; DALL·E 3 in ChatGPT optimizes for convenience and natural language prompting; Stable Diffusion optimizes for maximum control and open-source flexibility. Match the platform to your actual use case.

  • Midjourney is the right choice for creative professionals who generate images regularly and want consistently high aesthetic output. The Discord-based workflow is unusual but becomes natural quickly. The parameter system is learnable and produces much better results than default prompting.

  • DALL·E 3 in ChatGPT is the right choice for business professionals who need images occasionally and do not want to invest in learning a new system. Natural language prompting and conversational refinement make it accessible without training. Its text-in-images capability is the best among major platforms.

  • Stable Diffusion is the right choice for technical users who want maximum control, need self-hosted deployment for privacy, generate high volumes, or need ControlNet's precision composition capabilities. The ecosystem of fine-tuned models is vast and powerful.

  • Effective prompts describe visual qualities, not just subjects. The most powerful prompt elements: style and medium (photography, illustration, painting), lighting description, composition details, mood and atmosphere, color palette. Vague subject descriptions produce generic results; specific visual quality descriptions produce precise results.

  • Photography and art terminology produces better photorealistic results because these terms appeared consistently with specific visual qualities in training data. "Shot on Fujifilm X-T5 with 35mm lens, f/1.8, shallow depth of field, morning golden hour" specifies a visual quality package the model has learned.

  • Negative prompts remove unwanted elements. Adding --no hands in Midjourney or filling the negative prompt field in Stable Diffusion significantly reduces the frequency of common failure modes. Common negative prompt content: deformed hands, extra fingers, bad anatomy, blurry, watermark, text.

  • Hands and fingers are the most reliable AI-generated image tell, because hands appear in enormous variety in training data, creating a complex distribution that generates anatomically incorrect results frequently. Design around them: crop tightly, position people so hands are not prominent, use inpainting to fix specific hand regions.

  • Iteration is the workflow, not the failure. Professional users iterate 5-20 times for important images. Attempting to write a perfect prompt on the first try is less efficient than generating, evaluating, and refining.

  • Midjourney's --chaos parameter controls variation between the four images generated in one prompt. Use high chaos (50-100) early in exploration to find unexpected directions. Use low chaos (0-20) when refining a found direction.

  • Midjourney's /describe command reverse-engineers visual prompts from reference images. Upload an image you like and receive vocabulary describing its visual qualities. This is one of the fastest ways to build prompt vocabulary for specific aesthetic targets.

  • Reference images (--sref for style, --cref for character in Midjourney) dramatically improve consistency and let you communicate visual intent more precisely than text alone. Using a reference image as a style guide produces more targeted results than text description alone.

  • Maintaining visual consistency across a series of images is genuinely difficult without advanced techniques. Each generation is statistically independent. Use --sref for style reference, --cref for character reference, consistent seed values, and post-production design work to unify series.

  • DALL·E 3's text rendering is significantly better than other platforms. For images that need to include readable words — signage, labels, mockups — DALL·E 3 is the better choice.

  • ControlNet (Stable Diffusion) enables precise compositional control through reference images: extract a human pose, provide a sketch, or use an edge-detected image to constrain the spatial arrangement of generated output. This capability does not exist in Midjourney or DALL·E 3.

  • Commercial image rights are governed by platform terms of service, which vary between platforms and change over time. Read current terms (not summaries) before commercial use. The legal landscape for AI-generated image copyright is unsettled in most jurisdictions as of early 2026.

  • Generating photorealistic images of real, identifiable people raises significant legal and ethical concerns — right of publicity, potential defamation, implied false endorsement. This is a clear "do not" case regardless of platform.

  • Aspect ratio affects composition. Set your target aspect ratio from the first generation prompt. Generating in 1:1 and wishing you had 16:9 wastes iteration rounds.

  • When not to use AI image generation: when photographic accuracy of real events or real people matters, when brand consistency across a series is critical without post-production investment, when audiences expect real photographs and receiving AI-generated images would be misleading, when the workflow overhead exceeds the value of the image.

  • The 90-minute mood board is realistic, not aspirational. A skilled user who knows their platform can go from a written brief to 8-12 mood board images in 60-90 minutes. Early in the learning curve, plan for more time.

  • AI image generation generates more value in concepting and communication than in final production. Its highest-value uses — mood boards, early-stage concept visualization, presentation illustration — are in the parts of creative workflow that previously required expensive time or compromised quality. Highest-volume final campaign deliverables remain better served by traditional production.

  • The process of generating images forces conceptual clarity. Writing a precise image prompt requires you to articulate what the image should communicate. This forcing function has secondary value for the quality of the underlying ideas being illustrated.