Introducing ChatGPT Images 2.0

TL;DR

OpenAI's ChatGPT Images 2.0 launches with dramatically improved text rendering, thinking mode, photorealism, and interactive image editing capabilities.

Key Points

• 1.ChatGPT Images 2.0 is available to all users immediately. It launched live during this stream in ChatGPT and the API, with a benchmark score of 1512 versus Gemini Flash's 1270 — a massive leap.
• 2.The model introduces a 'thinking mode' for image generation. It can perform web searches, maintain coherence across multiple images, generate QR codes, and check its own work before outputting a final result.
• 3.Text rendering is dramatically improved across all languages. Asian languages like Hindi, Chinese, Korean, and Japanese — which have thousands of characters — can now be generated without errors, including full pages of dense text.
• 4.The model supports interactive, conversational image editing. Rather than one-shot prompting, users can iteratively refine images through dialogue, adjusting details like replacing a French fry with a ballpoint pen.
• 5.Photorealism is a major upgrade, triggered by keywords like 'photorealistic' or 'shot on iPhone.' The model replicates grain, lighting imperfections, and realistic textures, including a faithful recreation of a 2015 lecture hall.
• 6.New aspect ratio support allows images up to 3:1 and 1:3. This enables panoramic and ultra-tall images; one demo produced a fully consistent 360-degree panorama of the moon landing with accurate sun and shadow direction.
• 7.The model can write coherent, contextually accurate long-form text on images. A demo generated a realistic old-timey newspaper about Tim Cook leaving Apple with correct layout, headlines, and plausible body copy.
• 8.Transparent PNG background generation is now supported. Users can generate images with true transparency, useful for dropping assets directly into Photoshop or design tools — though it broke inconsistently during the live demo.
• 9.A 4K API experiment demonstrated extreme detail precision. The model wrote 'GPT Image' on a single grain of rice within a large pile, visible only when zoomed in significantly.
• 10.The model still has notable limitations exposed during live testing. It failed to generate a wine glass filled to the brim, got clock times wrong, and handled the Where's Waldo prompt by making Waldo transparent rather than hidden in a crowd.
• 11.The thinking mode allowed the model to find and quote real social media reactions to the beta 'duct tape' model. It synthesized posts from Threads, LinkedIn, and Reddit into a single image alongside a working QR code linking to ChatGPT.