Section 120 · Chapter 15, How Models Work
How Image Generation Models Work
Image generation is usually a denoising process guided by text, seed, model, and safety constraints.
What to do
- Start with prompt adherence.
- Do not let a high aesthetic score hide a broken aspect ratio, missing alpha channel, unsafe style, or changed product identity.
- Use masks, perceptual diffs, OCR, face or product similarity checks when appropriate, and human review for things automated metrics miss.
- Ask for a sign, label, UI screenshot, prescription bottle, chart title, or package front with exact text.
- Use a small prompt suite that covers the product's real use.
Evidence to preserve
- Include prompts with exact counts, relative position, occlusion, reflections, transparent objects, hands, small text, dark scenes, low contrast, unusual aspect ratios, and non-English text.
- Preserve the inputs, versions, configurations, raw outcomes, and results for image generation, image generation models work needed to reproduce work on How Image Generation Models Work.
- Report results for image generation, image generation models work by relevant slice, separate blocker failures from averages, state uncertainty and blind spots, and connect the result to a release decision.
Expert note
Evaluate image models with a mix of human review, vision-language judges, OCR, perceptual metrics, prompt adherence rubrics, safety classifiers, similarity checks for preserved regions, and slice tests for demographics, languages, styles, and sensitive domains. Always keep prompts, negative prompts, seeds, model versions, sampler settings, aspect ratio, masks, safety settings, generated artifacts, rejected artifacts, and reviewer notes with the eval record.
Continue the conversation
Apply this to your context.
Save your product context once, then open a focused conversation that combines it with this concept.
Cite this page
Jason Arbon. "How Image Generation Models Work." Testing AI Knowledge Edition, section 120.
https://jarbon.ai/testing-ai/knowledge/ch120-image-generation-models-work.html