3 ms·
I imagine the technique of having AI recreate the image from scratch based on a very detailed description might work.
by ZeWaka 5mo ago
I imagine the technique of having AI recreate the image from scratch based on a very detailed description might work.
- raincole 5mo agoThat'd not work with today's technology. No open model's prompt adherence is anywhere remotely close to ChatGPT/NanoBanana. 'remotely' here is a funny understatement, as I don't have a strong enough word in my vocabulary to describe how far the open models are behind the closed ones. Writing a more detailed description does not make the models stick to it more.
- vunderba 5mo agoDefinitely. I run an entire site built around a series of benchmarks that focus on prompts of increasingly difficult complexity with a focus on adherence, and even the state-of-the-art local models are probably only about thirty percent as good as proprietary models like Gemini 3.1 Flash Image and GPT Image 2. Comparing Qwen-Image, Flux.2, ZiT, NB2, and gpt-image-2 https://genai-showdown.specr.net/?models=qi,nbp3,f2d,g2,zt https://genai-showdown.specr.net/?models=qi,nbp3,f2d,g2,zt