AI LLM image models are getting impressive! In a couple of years, we’ll go from basic drawings to convincing photorealistic synthetic images!!
First of all we need to set our individual key, in this case for OpenAI on the OS Linux and MAC such as:
export OPENAI_API_KEY=zzzzzzzzzzzzzzzzzzzzzz
Or it can be set by GUI menu: Edit, Preferences, AI Settings or/and Services for AI.
GPT Image 2 (gpt-image-2) is OpenAI’s state-of-the-art image generation and editing model, released on April 21, 2026, as the successor to previous GPT Image versions (and integrated into ChatGPT Images 2.0).
It excels at fast, high-quality text-to-image and image-to-image tasks, delivering superior photo-realism, pixel-perfect text rendering (including multilingual support), better instruction-following, improved layouts, and stronger editing capabilities compared to earlier models.
Available via the OpenAI API and ChatGPT, it supports flexible image sizes (up to near-4K resolutions with constraints), high-fidelity image inputs, and complex visual tasks like product photography, posters, and precise designs, making it particularly useful for both creative and commercial applications.
OpenAI runs gpt-image-2 on its own massive datacenter infrastructure, which heavily relies on advanced NVIDIA GPUs (part of their broad partnership with NVIDIA for training and inference at scale). Exact details on the specific boards or cluster configurations for this model are not publicly disclosed, but OpenAI’s workloads generally use high-end NVIDIA accelerators such as:
- H100/H200: Common for high-throughput inference.
Blackwell-series (B200, GB200 NVL72, etc.): For newer, more efficient large-scale deployments.
These enterprise GPUs provide the massive VRAM (80–141+ GB HBM), high-bandwidth memory, and Tensor Core performance needed for fast diffusion-style or transformer-based image generation at high resolutions and batch sizes. Consumer boards like game PC RTX 4090s are irrelevant here, OpenAI uses datacenter-grade hardware in large clusters for production serving.
In short, we cannot run the real gpt-image-2 kinds of models on our own hardware (local or self-hosted); we must use OpenAI’s cloud service. For fully local image generation, alternatives like open-source models (Stable Diffusion 3, Flux, Aurora, etc.) running on consumer or datacenter NVIDIA GPUs are the way to go.
We can set in Wolfram Language the model to use as follows, simple examples:

Example: with a cat, as all the Internet are crazy about cats

ImageSynthesize["A white cat with a red hat and blue glasses in a forest with a green background, photorealistic",LLMEvaluator->LLMConfiguration["Model"-><|"Service"->"OpenAI","Name"->"gpt-image-2"|>]]
Out[]=

Example: now with a dog

ImageSynthesize["Dog with red hat at Niagara Falls",LLMEvaluator->LLMConfiguration["Model"-><|"Service"->"OpenAI","Name"->"gpt-image-2"|>]]
Out[]=

Example: Lets create a fake UFO typical noisy image

In[]:=
ImageSynthesize["Create a blury fake UFO image over a downtown at night",LLMEvaluator->LLMConfiguration["Model"-><|"Service"->"OpenAI","Name"->"gpt-image-2"|>]]
Out[]=
GPT Image 2 (gpt-image-2) is a natively integrated multimodal image generation model from OpenAI, deeply coupled with the underlying GPT architecture (not a standalone diffusion model like earlier DALL·E versions). It processes text prompts (and optional image inputs) through advanced language understanding, often employing a “thinking” or reasoning mode that plans, composes, self-verifies, and can even incorporate web search before generating the final output in a more efficient single-pass architecture. This enables superior instruction-following, precise layouts, near-perfect text rendering (including multilingual), photo-realism, and complex edits.
It was trained on massive multimodal datasets combining internet-scale image-text pairs, high-quality curated visuals, and synthetic data, using a combination of supervised fine-tuning, reinforcement learning from human/AI feedback (RLHF/RLAIF), and iterative optimization focused on fidelity, instruction adherence, and safety—building on the foundation of prior GPT multimodal training at OpenAI.
More advanced elaborated prompt engineering:

Example: Product Photography (Great for commercial/brand use)

In[]:=
ImageSynthesize["A premium product photography of a sleek black wireless earbuds case floating in mid-air with subtle water droplets, dramatic studio lighting with soft blue and purple rim light, clean white background with minimal shadow, ultra-realistic, 8k detail, shot on Canon EOS R5, commercial advertising style",​​LLMEvaluator->LLMConfiguration["Model"-><|"Service"->"OpenAI","Name"->"gpt-image-2"|>]]
Out[]=

Example: Text-Heavy Illustration (Showcases text rendering strength)

In[]:=
ImageSynthesize["A minimalist motivational poster with bold sans-serif typography that reads 'The only limit is your imagination' in elegant white text with subtle gold outline, dark navy blue gradient background with faint constellation patterns, modern luxury aesthetic, high contrast, perfect text alignment, cinematic composition, 2K resolution",​​LLMEvaluator->LLMConfiguration["Model"-><|"Service"->"OpenAI","Name"->"gpt-image-2"|>]]
Out[]=

Example: Fantasy Scene with Complex Composition

Common use cases

1. Marketing and advertising

Generate:
◼
  • product hero shots
    • ◼
  • social media creatives
    • ◼
  • banner ads
    • ◼
  • concept visuals for campaigns
    • ◼
  • packaging mockups
    • ◼
  • posters and flyers
    • Example:

    2. Rapid concept art

    Useful for:
    ◼
  • game environments
    • ◼
  • character ideation
    • ◼
  • cinematic mood boards
    • ◼
  • architecture concepts
    • ◼
  • creature design
    • ◼
  • fantasy and sci-fi world-building
    • This is often one of the biggest productivity wins: generating many directions quickly before committing to manual production.

    3. Storyboarding and pre-visualization

    For film, animation, games, and presentations:
    ◼
  • scene roughs
    • ◼
  • camera framing ideas
    • ◼
  • lighting mood exploration
    • ◼
  • costume and prop concepts
    • ◼
  • location visualization

    • 4. Graphic design drafts

    Generate first-pass designs for:
    ◼
  • posters
    • ◼
  • album covers
    • ◼
  • book covers
    • ◼
  • invitation cards
    • ◼
  • presentation backgrounds
    • ◼
  • editorial illustrations
    • Text rendering has historically been weak in many systems, but newer models are much better for layout-heavy work.

    5. E-commerce and product visualization

    Create:
    ◼
  • clean catalog images
    • ◼
  • lifestyle product scenes
    • ◼
  • alternate colorways
    • ◼
  • seasonal backgrounds
    • ◼
  • staged product compositions
    • ◼
  • packaging previews
    • This is especially useful when photography is expensive or slow.

    6. Educational content

    Produce:
    ◼
  • historical scene illustrations
    • ◼
  • scientific concept art
    • ◼
  • visual explanations
    • ◼
  • classroom posters
    • ◼
  • geography and culture visuals
    • ◼
  • children’s educational material

    • 7. Publishing and media

    Generate images for:
    ◼
  • blog posts
    • ◼
  • magazine concepts
    • ◼
  • article headers
    • ◼
  • newsletters
    • ◼
  • thumbnails
    • ◼
  • infographics backgrounds

    • 8. UI/UX and app prototyping

    Designers can create:
    ◼
  • onboarding illustrations
    • ◼
  • empty-state artwork
    • ◼
  • visual themes
    • ◼
  • conceptual dashboard art
    • ◼
  • branded imagery for prototypes

    • 9. Personalized creative content

    Generate custom:
    ◼
  • greeting cards
    • ◼
  • wallpapers
    • ◼
  • avatars
    • ◼
  • profile images
    • ◼
  • family storybook illustrations
    • ◼
  • gifts and novelty prints

    • 10. Architectural and interior visualization

    Use prompts for:
    ◼
  • room redesign ideas
    • ◼
  • facade concepts
    • ◼
  • landscape mood studies
    • ◼
  • furniture staging
    • ◼
  • lighting variations
    • ◼
  • material/finish exploration

    • 11. Fashion and beauty

    Generate:
    ◼
  • editorial looks
    • ◼
  • apparel concept designs
    • ◼
  • style boards
    • ◼
  • cosmetic campaign imagery
    • ◼
  • accessory mockups

    • 12. Synthetic data and simulation

    Potentially useful for:
    ◼
  • creating varied image sets
    • ◼
  • scenario simulation
    • ◼
  • visual stress testing
    • ◼
  • generating examples for downstream workflows
    • Though for serious ML dataset creation, quality control and bias review are very important.

    Beyond pure text-to-image

    ◼
  • text-only generation
    • ◼
  • image variation
    • ◼
  • inpainting transparent regions
    • ◼
  • masked editing of a selected region
    • That makes image models useful not only for creating images, but also for editing them.

    Best-fit business uses

    In practice, text-to-image is strongest when you need:
    ◼
  • speed over traditional production
    • ◼
  • many variations
    • ◼
  • low-cost ideation
    • ◼
  • custom visuals on demand
    • ◼
  • niche imagery that stock photos do not cover

    • Limitations and cautions

    Important caveats:
    ◼
  • outputs may be inconsistent across runs
    • ◼
  • prompt wording strongly affects results
    • ◼
  • fine details may still fail
    • ◼
  • brand/logo accuracy may be imperfect
    • ◼
  • legal and copyright policies matter
    • ◼
  • bias and safety issues must be reviewed
    • ◼
  • photorealistic fake imagery can be misused

    • In one sentence

    Text-to-image models are most useful for fast visual ideation, custom marketing graphics, product visualization, concept art, design drafts, and image editing workflows.

    CITE THIS NOTEBOOK

    AI LLM text to image models​
    by Daniel Carvalho​
    Wolfram Community, STAFF PICKS, July 5, 2026
    ​https://community.wolfram.com/groups/-/m/t/3745213