多模態模型／生成式影像
GPT‑Image‑2.5 Splits Into Flare and Sunburst, Separating Fast Generation and High-Precision Editing Into Two API Paths
OpenAI has launched ChatGPT Images 2.5 and introduced two API options: Flare, optimized for throughput, and Sunburst, designed for fine-grained control. The new release emphasizes localized editing, multi-turn consistency, and latency reductions of up to 50%, although its performance and quality figures still come primarily from the publisher.
OpenAI launched ChatGPT Images 2.5 on September 8, while adding `GPT‑Image‑2.5 Flare` and `GPT‑Image‑2.5 Sunburst` for developers. Flare is positioned as the default choice for most applications, providing the new release’s image generation, reference-image transformation, and editing capabilities. OpenAI says it can reduce latency by up to 50% compared with GPT‑Image‑2 while delivering higher quality. Sunburst, by contrast, trades longer generation times for more granular control and targets finished advertising assets, product imagery, and high-value workflows that require repeated fine-tuning. This dual-model configuration allows engineering teams to route workloads according to interactive latency, batch throughput, and final quality instead of forcing every request to use the same inference budget.
The core improvement is not limited to higher-quality text-to-image generation in a single pass. OpenAI says version 2.5 more reliably preserves people and objects from reference photos, can modify only specified regions within complex backgrounds, and is less likely to let earlier changes drift or accumulate distortions across multiple editing turns. The ChatGPT interface now lets users comment directly on an image to identify regions for modification, constrain compositions with hand-drawn sketches, apply templates, and share prompts. For API-based products, the corresponding benefit is a reduction in the need to regenerate entire assets, create masks manually, or perform compositing in post-production. The model is also claimed to improve the accuracy of infographics, layouts, transparent backgrounds, and real-world content, but OpenAI has not disclosed the hardware, resolution, batch size, or sampling settings needed to reproduce its latency figures.
The expanded capabilities also increase the risk of convincing impersonations of real people. The system card says the new version retains multilayer classifiers for both text prompts and generated images while adding C2PA metadata and invisible watermarks. These mechanisms, however, cannot prevent screenshots, cropping, re-encoding, or downstream models from removing provenance signals. Axios’s early testing found tangible improvements in preserving the appearance of people and pets and in localized editing, but it does not constitute a systematic benchmark. Deployers should therefore measure p50/p95 latency at fixed resolutions, identity and brand consistency after successive edits, performance on non-Latin text, and the false-rejection rate caused by safety filters for legitimate edits involving real people before setting routing thresholds between Flare and Sunburst.