{"title":"Exercise: Text-to-Image Generation","description":null,"content":"# Exercise 1: Generate an Image and a Variation\n\nGenerate an image from a text prompt, then generate a second version with one detail changed, so you can see how prompt edits steer the output.\n\nImage generation is prompt engineering for pixels. The fastest way to learn what the model responds to is to generate one image, change a single attribute in the prompt (style, color, setting), and generate again. Holding everything else fixed shows you which words actually move the result.\n\n### Requirements\n\n- Generate an image from a base prompt with `client.images.generate`.\n- Change one attribute in the prompt and generate a second image.\n- Print confirmation and the size of each returned image payload (you cannot display the image inline here).\n\n\n**main.py**\n\n```python\nfrom openai import OpenAI\nimport base64\n\nclient = OpenAI() # Uses OPENAI_API_KEY from environment\n\nbase_prompt = \"A red bicycle parked outside a cafe, photorealistic\"\nvariation_prompt = \"A blue bicycle parked outside a cafe, photorealistic\"\n\n# TODO: generate an image for base_prompt with client.images.generate(model=\"gpt-image-2\", ...)\n\n# TODO: generate an image for variation_prompt\n\n# TODO: print a confirmation and the byte size of each decoded image\n```\n\n\n### Expected Output\n\nBoth calls return image data, and the only intended difference between them is the changed attribute.\n\n\n```plaintext\nBase image generated: 1843200 bytes\nVariation generated: 1850112 bytes\n```\n\n\n### Solution\n\n\n**main.py**\n\n```python\nfrom openai import OpenAI\nimport base64\n\nclient = OpenAI()\n\nbase_prompt = \"A red bicycle parked outside a cafe, photorealistic\"\nvariation_prompt = \"A blue bicycle parked outside a cafe, photorealistic\"\n\ndef generate(prompt: str) -> bytes:\n response = client.images.generate(model=\"gpt-image-2\", prompt=prompt)\n return base64.b64decode(response.data[0].b64_json)\n\nbase_image = generate(base_prompt)\nprint(f\"Base image generated: {len(base_image)} bytes\")\n\nvariation_image = generate(variation_prompt)\nprint(f\"Variation generated: {len(variation_image)} bytes\")\n```\n\n\nThe two calls are identical except for one word in the prompt, which isolates that attribute as the cause of any difference, exactly the controlled-change approach you used with temperature earlier. The model returns the image as base64, which you decode into bytes (and would normally write to a file or display). Building intuition this way, one attribute at a time, is how you learn to write image prompts that produce what you intend rather than something close to it.\n\n---\n\n# Exercise 2: Generate the Same Subject in Several Styles\n\nHolding the subject fixed while varying the style isolates style as a prompt dimension. Generate the same scene in several styles and confirm each renders. This is how you build a style picker or A/B different looks for the same content.\n\n### Requirements\n\n- Keep one subject and a list of styles.\n- Generate an image for each `subject, style` combination.\n- Print the style and the byte size of each generated image.\n\n\n**main.py**\n\n```python\nfrom openai import OpenAI\nclient = OpenAI() # Uses OPENAI_API_KEY from environment\nimport base64\n\nsubject = \"a bicycle parked outside a cafe\"\nstyles = [\"photorealistic\", \"watercolor painting\", \"pixel art\"]\n\n# TODO: for each style, generate an image of the subject in that style\n# TODO: print the style and the byte size of each generated image\n```\n\n\n### Expected Output\n\nThree images are generated, one per style, each returning image data.\n\n\n```plaintext\nphotorealistic: 1843200 bytes\nwatercolor painting: 1798144 bytes\npixel art: 1655296 bytes\n```\n\n\n### Solution\n\n\n**main.py**\n\n```python\nfrom openai import OpenAI\nclient = OpenAI() # Uses OPENAI_API_KEY from environment\nimport base64\n\nsubject = \"a bicycle parked outside a cafe\"\nstyles = [\"photorealistic\", \"watercolor painting\", \"pixel art\"]\n\nfor style in styles:\n response = client.images.generate(model=\"gpt-image-2\", prompt=f\"{subject}, {style}\")\n data = base64.b64decode(response.data[0].b64_json)\n print(f\"{style}: {len(data)} bytes\")\n```\n\n\nAppending the style to a fixed subject keeps everything else constant, so the only variable across the three images is the rendering style, which is what lets you compare them fairly. Looping over a list of styles is the same approach you would use to offer a user several looks or to test which style performs best for a use case. Each call returns independent image data, decoded here just to confirm it arrived.","pageType":"ai-engineering"}

Get Premium