New OpenAI model GPT Image 2

GPT Image 2 — New OpenAI Model with 16-Reference Image-to-Image

GPT Image 2 is OpenAI's next-generation image model, the successor to GPT Image 1.5. The headline upgrade is multi-reference image-to-image: the model accepts up to 16 input images in a single request and composes the output while respecting every one of them. It also ships a dedicated edit endpoint for in-place changes — edits are applied to the source image without regenerating untouched pixels. GPT Image 2 keeps the family's signature strengths: legible text rendering, photorealism, and long structured-prompt adherence.

What's new in GPT Image 2 vs. 1.5

GPT Image 2 builds on the same multimodal backbone as GPT Image 1.5 but adds four capabilities that no prior OpenAI image model supported end-to-end:

• Multi-reference (up to 16 images). You can attach a character reference, a style sample, a color palette, a background, and a composition reference — and the model fuses them into one coherent output. Previously this required chaining multiple generations or external LoRAs.
• Native edit endpoint. Upload a finished image and describe the change; the model treats the request like a mask, leaving the rest of the pixels alone. Ideal for poster tweaks, product shots, and banner iteration without losing the source.
• Quality tiers (low / medium / high / auto). Pick fast, cheap iteration for exploration or maximum fidelity for final assets. Price and latency scale with the tier.
• Better long-prompt adherence. GPT Image 2 handles multi-step instructions over 1,500+ characters more reliably — lists of requirements, technical parameters, and composition constraints are respected more often.

Multi-reference: what 16 references are actually for

The 16-reference capability is the main reason to move from GPT Image 1.5 to GPT Image 2. Practical workflows:

• Marketplace product cards. Product shot + 3 backgrounds + lighting reference = a consistent series of listings.
• Character consistency. 4–8 portraits from different angles + clothing reference + scene = new frames of the same person, no LoRA training required.
• Brand-style transfer. Logo + brand colors + examples of existing creative = new assets that land inside your brand guidelines.
• Mood boards and collages. Assemble a visual idea from 10–16 sources and get a single image that unifies them stylistically.
• Style transfer. Source photo + artistic style reference = the same subject rendered in the chosen manner with original details preserved.

GPT Image 2 without ChatGPT Plus — how it works on Hubery.ai

On Hubery.ai, GPT Image 2 is available through a clean web UI — no ChatGPT Plus subscription ($20/mo), no VPN. Attach 1 to 16 reference images, describe the task, pick a quality tier, and you get a result in 8–25 seconds. Pricing ranges from about $0.07 at low quality to $0.40 at high quality. Free tokens are granted on sign-up.

Hubery.ai also gives you access to 50+ AI models inside a single account: GPT Image 2, GPT Image 1.5, DALL-E 3, Midjourney, Stable Diffusion, Flux, Seedream, NanoBanana, Ideogram, plus dozens of text, code, and audio models. No need to pay each vendor separately — everything runs off one balance.

GPT Image 2 vs GPT Image 1.5

GPT Image 1.5 is still a great choice for one-shot text-to-image generations: it's cheaper, faster, and for a typical prompt the output is comparable. GPT Image 1.5 remains available on Hubery.ai. Move to GPT Image 2 when you need image-to-image with multiple references, in-place editing without regenerating, explicit quality control, or reliable execution of a long brief. For commercial assets with character or brand consistency, GPT Image 2 saves several iteration cycles compared to 1.5.

GPT Image 2 vs DALL-E 3 and Midjourney

DALL-E 3 is built on a prior-generation diffusion pipeline and supports neither multi-reference nor a native edit endpoint. For collages, product photos with character consistency, and branded series, GPT Image 2 stays noticeably closer to the brief.

Midjourney v7 still leads on artistic atmosphere, but it runs through Discord, has no multi-image API, is weaker with on-image text, and can't edit a specific region of a finished image. GPT Image 2 wins where precision, text, edit, and predictable iteration matter.

Both DALL-E 3 and Midjourney are also available on Hubery.ai — pick the tool that fits the task without switching platforms.

Prompt examples for GPT Image 2

Ready-to-use prompts you can paste into Hubery.ai. For multi-reference prompts, attach the images first and then paste the text:

[Multi-ref: 1=product photo, 2=studio backdrop, 3=lighting reference]
Marketplace card: product from ref1 on backdrop from ref2,
match the lighting and shadows of ref3, add a "NEW" badge
in the top-right corner, 1024×1024, high quality.

[Multi-ref: 1=portrait of the person, 2=outfit sample, 3=scene]
Same person as ref1 wearing the outfit from ref2, in the
location shown in ref3. Natural light, full-length portrait,
photorealistic.

[Edit endpoint: a finished movie poster uploaded]
Replace the tagline with "Premieres May 1" in the same font,
leave everything else untouched. Preserve composition pixel-for-pixel.

[Text-to-image, high quality]
Coffee shop logo "North" — minimal, black on white,
a geometric pine tree built from two triangles, wordmark below
in a monospaced font. No gradients.

Professional use cases

GPT Image 2 covers jobs that previously required a designer or multiple specialist models:

• E-commerce. Series of product cards with consistent background, lighting, and style — built from one product shot and a handful of references.
• Marketing and advertising. Social creatives, banners with accurate copy and pricing, A/B variants in low quality, finals in high.
• Branding. Logo concepts, identity mockups, and creative that stays inside an existing brand guideline via multi-reference.
• Product prototyping. UI mockups and landing-page illustrations that match the visual language of the product.
• Editorial and social content. Original covers and illustrations with clean, legible copy — instead of stock imagery.
• Post-production. The edit endpoint replaces routine Photoshop chores: swap a background, remove an object, update a caption, all without re-rendering the whole image.