Why GPT‑4o’s Image Generation Is Overwhelming Users—and What It Means for AI

GPT‑4o native image generation

OpenAI launched image generation in GPT‑4o less than 72 hours ago. The feature is currently limited to Plus, Pro, and Team subscribers; free users have been promised three daily generations in the future.

Performance and usage limits

High demand quickly saturated GPU resources, prompting OpenAI to impose temporary usage caps.

Users report generation times up to 30 minutes per image.

OpenAI is working on efficiency improvements while maintaining the temporary limits.

Image quality and underlying architecture

Generated images demonstrate strong prompt understanding, consistent style, and detailed knowledge, indicating that visual output is driven by the LLM rather than a separate diffusion model.

Technical upgrades described by OpenAI:

Convert the user prompt into a detailed English instruction.

Leverage the model’s internal knowledge base and conversation context.

Pre‑process any reference images before generation.

This chain‑of‑thought pipeline yields higher fidelity and style consistency.

Limitations and copyright handling

After generation, some outputs are blocked due to copyright detection, raising questions about pre‑generation filtering.

Requests for specific copyrighted styles (e.g., “Studio Ghibli”) are sometimes rejected, while similar non‑copyrighted styles are allowed, leading to perceived double standards.

Derivative applications, filter apps, and meme‑based legal threats have emerged around the generated images.

Recent updates

OpenAI released an update that makes the image generation feature available to all paid users; free users are slated to receive the same access in the coming weeks.

The update does not alter the underlying technical pipeline but aims to stabilize resource usage.

Code example

收
藏
，
分
享
、
在
看
，
给
个
三
连
击呗！