精选理由
HeyGen 用同一提示词测了四个模型，生成的脸像但非本人。和自家基于真实录像的虚拟形象一比，差距明显。
HeyGen 用同一提示词测试了四个图像模型的身份保持能力，包括 Flux 2.0 pro、Seedream V5、Kling V3 等。Flux 2.0 pro、Seedream V5、Kling V3 各自生成了相似的脸，但都与目标真人不符。HeyGen 强调自家虚拟形象基于真实视频素材训练，这正是与这些模型的关键差别。
原文 · HeyGen
We tested identity preservation across four image models with one identical prompt Flux 2.0 pro, Se...
We tested identity preservation across four image models with one identical prompt Flux 2.0 pro, Seedream V5, and Kling V3 each produced a similar person. None produced the actual person HeyGen avatars are trained on real footage. That's the difference x.com/i/article/2083… 💬 5 🔄 0 ❤️ 11 👀 1505 📊 6 ⚡