Sharing Learning Tip - prompt iterating with nano-banana just reduces image quality after a few cycles. Instead, always take the 1st high quality image output, start a NEW chat, ask it to generate a new image based on attached IMAGE 1 and referencing body pose/facial expression of characters in IMAGE 2 (which should be latest scene). Wasted a lot of time going thru pictures, only to realize later upscaling the image won't improve quality. - This quality degradation is a known phenomenon in iterative AI editing often called "Latent Drift" or "VAEDecay." - In 2026, Nano Banana (and models like it) works by converting your image into a mathematical "noise" space (latent space) and then back into pixels. Every time you ask for an edit on an existing image, it goes through this conversion again. After 3–4 rounds, the "translation" starts to lose data, making the image look "crunchy," overly smoothed, or "plastic-y."

ByteDance officially launched Seedance 2.0 just yesterday (February 12, 2026), and it is being positioned as a "Sora-killer" specifically for creators who need precision over randomness. You'll find the Seedance vs. Sora 2 comparison particularly interesting because it represents two different philosophies: Production Control (Seedance) vs. Physical Realism (Sora 2). https://seedance2ai.online/ Actually that doesn't have the full set of input features - which are only on its Chinese full version: https://jimeng.jianying.com/ but requires a Douyin or Mainland TikTok (+86) account.