DALL-E 3

DALL-E 3 features

DALL-E 3 (OpenAI, October 2023) is OpenAI's image generation model. Its distinctive feature is automatic prompt enhancement: before generating an image, DALL-E 3 uses a GPT-4-class model to rewrite and expand your prompt — adding details, fixing ambiguities, and optimising for image generation quality. This means brief prompts like "a dog in a park" generate much more detailed, coherent images than previous models where you had to craft detailed prompts manually. The tradeoff: you lose precise control over the exact image — the prompt rewriting can add details you didn't want.

Generating images via API

import openai, base64, os
from pathlib import Path

client = openai.OpenAI()

# Generate a new image
response = client.images.generate(
    model="dall-e-3",
    prompt="A photorealistic illustration of a robot learning to paint in the style of Monet",
    size="1024x1024",        # "1024x1024", "1792x1024", "1024x1792"
    quality="standard",      # "standard" or "hd" (2x cost, higher detail)
    style="vivid",           # "vivid" (bright, dramatic) or "natural" (muted, realistic)
    n=1,                     # DALL-E 3 only supports n=1 per request
    response_format="url",   # "url" (1hr expiry) or "b64_json"
)

image_url = response.data[0].url
revised_prompt = response.data[0].revised_prompt  # prompt after GPT rewriting
print(f"Revised prompt: {revised_prompt}")
print(f"Image URL: {image_url}")

# Download and save
import requests
img_data = requests.get(image_url).content
Path("output.png").write_bytes(img_data)

# Or get base64 directly
response_b64 = client.images.generate(
    model="dall-e-3",
    prompt="Abstract geometric art",
    size="1024x1024",
    response_format="b64_json",
)
img_bytes = base64.b64decode(response_b64.data[0].b64_json)
Path("output_b64.png").write_bytes(img_bytes)

Prompt enhancement behaviour

DALL-E 3 rewrites every prompt before generation. The revised prompt is returned in response.data[0].revised_prompt. You can partially suppress this by explicitly stating requirements: "I NEED to generate this exact prompt with no modifications. DO NOT add details. Prompt: [your prompt]". However, OpenAI recommends working with the rewriting rather than against it — the model is trained to produce better images from its revised prompts.

The rewriting typically: adds lighting descriptions, specifies art style explicitly, adds details about composition, clarifies ambiguous subjects, and removes content that might trigger safety filters.

Quality and style options

• quality="standard": Faster generation (~6s), lower cost ($0.040/image at 1024×1024)
• quality="hd": More detail in larger compositions ($0.080/image at 1024×1024)
• style="vivid": Hyper-real, saturated, cinematic — good for dramatic scenes
• style="natural": More muted, photographic, realistic — good for product images and portraits
• size="1792x1024": Wide landscape format ($0.120/image HD)
• size="1024x1792": Tall portrait format ($0.120/image HD)

Editing images

from PIL import Image
import io

# DALL-E 2 supports inpainting (edit with mask) — DALL-E 3 does NOT currently
# Use DALL-E 2 for edits:
with open("original.png", "rb") as img_file, open("mask.png", "rb") as mask_file:
    response = client.images.edit(
        model="dall-e-2",
        image=img_file,
        mask=mask_file,        # transparent areas = regions to fill
        prompt="A wooden table with a vase of flowers",
        n=1,
        size="1024x1024",
    )

# For DALL-E 3 "editing": use img2img style with a detailed description
# or use GPT-4V to describe the image, then regenerate with modifications in the prompt

Comparing DALL-E 3 vs Flux vs Stable Diffusion

• DALL-E 3: Best at following complex text descriptions and rendering accurate text in images. Closed, API-only. $0.04–$0.12 per image. No fine-tuning.
• Flux (Black Forest Labs): State-of-the-art open-source model. FLUX.1[dev] rivals DALL-E 3 quality. Runnable locally. Fine-tunable. No per-image cost beyond compute.
• Stable Diffusion XL: Mature ecosystem, enormous community, many fine-tuned variants. Slightly below DALL-E 3 quality on prompt following but more flexible and customisable.

Gotchas

• n=1 only: DALL-E 3 only generates one image per API call. To get multiple variations, make multiple API calls — not n=4 like DALL-E 2.
• URL expiry: The returned image URL expires after ~1 hour. Download and store the image immediately if you need it later — don't save just the URL.
• Content policy: DALL-E 3 has strict content filtering. Real people's faces, copyrighted characters, and adult content are blocked. The revised prompt sometimes changes your subject to avoid policy violations without telling you explicitly.
• Text rendering: DALL-E 3 is one of the few models that can render short text accurately in images. For longer text, accuracy drops. Use it for logos and signs with short phrases.