OpenAI GPT-Image 2 Now Available on PicMorph
OpenAI's state-of-the-art image model lands on PicMorph via Replicate.
PicMorph Editorial Desk
OpenAI GPT-Image 2 Now Available on PicMorph
OpenAI''s latest image model, GPT-Image 2, is now wired into PicMorph through Replicate. You can use it directly from the AI editor for both text-to-image generation and reference-image editing.
Primary source: openai/gpt-image-2 on Replicate
What changed
GPT-Image 2 is positioned by OpenAI as their state-of-the-art image generation and editing model. The published model card highlights:
- Strong instruction following on dense, multi-clause prompts
- Sharp text rendering inside images, including small type and complex layouts
- Identity-preserving edits when supplied with reference images
- Style consistency across batches
It supersedes GPT-Image 1.5 in PicMorph''s lineup but does not replace it; GPT-Image 1.5 remains available for workflows that depend on transparent backgrounds or the legacy parameter set.
How it differs from GPT-Image 1.5
If you have been using GPT-Image 1.5, the parameter surface is leaner and the limits have moved.
| Parameter | GPT-Image 1.5 | GPT-Image 2 |
|---|---|---|
aspect_ratio | many ratios | 1:1, 3:2, 2:3 |
quality | low / medium / high | low / medium / high / auto |
background | auto / transparent / opaque | auto / opaque (no transparent) |
moderation | auto / low | auto / low |
output_format | webp / png / jpeg | webp / png / jpeg |
number_of_images | up to 4 | up to 10 |
input_fidelity | low / high | removed |
output_compression | 0â100 | removed |
| Reference images field | image_input | input_images (array) |
If you need transparent backgrounds, stay on GPT-Image 1.5. If you need higher batch counts or cleaner editing semantics, GPT-Image 2 is the better default.
When to reach for it
- Posters, ads, and UI mocks with embedded text. Text rendering is the standout improvement; previous OpenAI image models often introduced subtle character corruption at small sizes. GPT-Image 2 is noticeably steadier here.
- Iterative edits on a real subject. Pass the original image (or several variants) via
input_imagesand describe the change. Identity is preserved more reliably than re-prompting from scratch. - Batch generation up to 10 variants. Useful for moodboards or A/B candidates without re-submitting the request.
When in doubt, start with quality: "auto" â the model will balance fidelity and cost without you having to guess.
Pricing and credits
On PicMorph, GPT-Image 2 is priced at 6 credits per generation, slightly above GPT-Image 1.5 to reflect the higher per-call cost upstream. Credit pricing on Replicate itself can change over time; refer to the Replicate model page for the authoritative cost.
Editorial note
This article is a source-based summary of OpenAI and Replicate''s public materials, not an independent benchmark. Real-world quality depends on prompt design, reference image selection, and Replicate''s rollout state. For the latest schema and capabilities, refer to the official model card.
Related Posts
Google Launches Gemini Embedding 2 Preview: A New Multimodal Retrieval Layer
An industry-news style brief on Google's Gemini Embedding 2 Preview, focused on model capabilities, architecture direction, and potential strategic advantages.
Seedance 2.0: Official Launch Brief and Practical Takeaways
A source-based summary of what ByteDance officially announced for Seedance 2.0 in February 2026, including capabilities, availability, and practical caveats.
Google Introduces Nano Banana 2: Official Announcement Summary
A source-based summary of Google's Nano Banana 2 announcement.