GOAT Intern (🎮,🐐)

Very excited to announce Mighty Bear Labs, our brand new Dev Diaries initiative at the studio[mightynet.tv/log11a]! It’s your backstage pass to our latest experiments and breakthroughs. In our first edition, we dive into some txt2img model evaluations, Nano Banana vs Qwen for image editing, and share an update on Nexus, our powerful in-house creative production agent. We’re testing the format, so expect more to come! 1. Nano Banana, Qwen, Kontext NanoBanana by @GoogleDeepMind wins against Qwen image edit by @Alibaba_Qwen . It even performs better than Kontext from @bfl_ml , overall, for our use cases. The main feature Kontext has an edge over NanoBanana is managing image ratio, zooming in/out, cropping, etc... Qwen image edit is great but, for our needs, we haven't found a a particular area where it did better than NB/Kontext. It's only the first release of Qwen Image Edit so, we'll revisit, as always. 2. Txt2Img models quality Dreamina 3.1 by @BytePlusGlobal / ByteDance and Qwen image quality and prompt adherence are beast! It was a limited eval, but they excelled in both pixel art and 3d style in that test. Surprisingly, Qwen's quality is superior on @huggingface , than on @FAL So, if you're doing visual development, starting with text, these are definitely SOTA models. 3. Nexus(our Production Agent) new update "Zoom into his head, crop at his shoulders and make aspect ratio 16:9" We've had Kontext support in Nexus for while, but just got our mind blown by the ability to direct very specifically where to crop and what format/ratio to output, using natural language! It just works 🤯