IT之家 10 月 21 日消息,视觉人工智能实验室 Black Forest Labs 当地时间本月 15 日以开放权重的形式发布了 FLUX 3 Image Klein 图像生成模型。
FLUX 3 Image Klein 具备 5.6B 的参数,针对消费级 GPU 开展了优化。其 BF16 权重需要 12.2GB 内存、FP8 权重则需要 5.6GB 内存。这款模型的 INT8 版本可借助 ComfyUI 或 Diffusers 在仅配备 8GB 显存的 NVIDIA GeForce RTX GPU 上运行。
依据测试,精简后的 INT8 版本在 GeForce RTX 5090 上生成 1MP 的图像大约需要 4s,在 GeForce RTX 4070 上借助 ComfyUI 生成 1MP 图像不到 9s。
FLUX 3 Image Klein 兼容文本转图像、图像编辑、多参考生成,原生兼容 4K 分辨率,可达成像素级精确编辑和稳定的图案限位表现。Black Forest Labs 称在多个情形下优于尺寸更大的开源模型,在速度-质量平衡上具备更好的表现。
参考
本文来源:https://www.ithome.com/1/016/062.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
IT House October 21st news, visual artificial intelligence laboratory Black Forest Labs local time on the 15th of this month released the flux 3 Image Klein image generation model in the form of open weights.
Flux 3 Image Klein is optimized for consumer GPUs with 5.6B parameters. The BF16 weight requires 12.2GB of RAM, and the FP8 weight requires 5.6GB of RAM. The INT8 version of this model runs on NVIDIA GeForce RTX GPUs with only 8GB of graphics memory with ComfyUI or Diffusers.

According to the test, the reduced INT8 version takes about 4s to generate a 1MP image on the GeForce RTX 5090, and less than 9s to generate a 1MP image on the GeForce RTX 4070 with ComfyUI.
Flux 3 Image Klein is compatible with text-to-image, image editing, multi-reference generation, natively compatible with 4K resolution, and can achieve pixel-level accurate editing and stable pattern limiting performance. Black Forest Labs says it outperforms larger open source models in multiple scenarios and performs better on speed-mass balance.
This article is from: https://www.ithome.com/1/016/062.htm For copyright issues, please write to: 13951431913#139.com or submit a review on the right.
