??大家好!欢迎来到破狼主页。感谢大家的支持与鼓励。在AIGC探索道路上,我将与你一路同行。喜欢就星标关注破狼或文末扫码加入交流群 !
今天文章将介绍一款由字节跳动最新开源的自适应一致性方案:HyperLoRA ,该框架目标是在通过参数高效的自适应生成方法,解决人物一致性肖像合成中的挑战。
基于模型微调的方法(如 LoRA 和 DreamBooth)可以生成逼真的输出,但需要对个体样本进行训练,耗时且资源密集,还存在不稳定的风险。基于适配器技术(如 IP-Adapter)通过冻结基础模型参数并采用插件架构实现零样本推理,但往往缺乏自然性和真实性。 HyperLoRA则是字节提出来的一种参数高效的自适应生成方法,它使用自适应插件网络生成 LoRA 权重,将 LoRA 的优越性能与适配器方案的零样本能力相结合。通过精心设计的网络结构和训练策略,HyperLoRA实现了零样本个性化肖像生成(支持单张和多张图像输入),具有高逼真度、保真度和可编辑性。
研发团队明确将 HyperLoRA 分解为 Hyper ID-LoRA 和 Hyper Base-LoRA。前者是在学习身份信息,而后者则用于拟合其他内容(如背景和服装)。这种设计有助于防止无关特征泄露到 ID-LoRA。在训练过程中,采用了固定预训练的 SDXL 基础模型和编码器的权重,仅通过反向传播更新 HyperLoRA 模块。在推理阶段,集成到 SDXL 中的 Hyper ID-LoRA 生成个性化图像,而 Hyper Base-LoRA 是可选的。
本文使用ComfyUI-HyperLoRA插件体验。模型文末网盘获取。
models/hyper_lora/clip_processor/clip_vit_large_14_processor
。下载:https://huggingface.co/openai/clip-vit-large-patch14/
models/hyper_lora/clip_vit/clip_vit_large_14
。下载:https://huggingface.co/openai/clip-vit-large-patch14
models/insightface/models
。下载:https://github.com/deepinsight/insightface/tree/master/python-package#model-zoo
models/hyper_lora/hyper_lora
。下载:https://huggingface.co/bytedance-research/HyperLoRA
models/
├── hyper_lora/
│ ├── clip_processor/ # CLIP 处理器文件目录
│ ├── clip_vit/ # CLIP ViT 模型文件目录
│ └── hyper_lora/ # HyperLoRA 模型文件目录
└── insightface/
└── models/ # InsightFace 模型文件目录
HyperLoRA一致性ComfyUI工作流下载地址:
beautiful woman, long flowing dark brown hair, silky texture, hair illuminated by warm orange backlight, radiant cinematic lighting, smooth skin, glowing complexion, elegant neckline, wearing a sleeveless fitted orange dress, deep neckline, modern fashion, soft shadows, studio photography, dramatic warm atmosphere, high contrast, meticulously styled hair, sophisticated style, high-definition image, fashion portrait, golden hues, artistic photography.
beautiful woman, long flowing dark brown hair, silky texture, hair illuminated by warm orange backlight, radiant cinematic lighting, smooth skin, glowing complexion, elegant neckline, wearing a sleeveless fitted orange dress, deep neckline, modern fashion, soft shadows, studio photography, dramatic warm atmosphere, high contrast, meticulously styled hair, sophisticated style, high-definition image, fashion portrait, golden hues, artistic photography.
chinese women, 1girl, solo, long_hair, breasts, looking_at_viewer, brown_hair, dress, cleavage, brown_eyes, collarbone, upper_body, small_breasts, sleeveless, lips, realistic, orange_dress
chinese women, A striking composition featuring a woman with flowing, dark brown hair cascading over her shoulders, illuminated by warm, golden backlighting that creates a halo-like effect. She is wearing a vivid orange sleeveless gown with a deep V-neckline, highlighting her collarbone and smooth, glowing skin. The overall color palette merges rich amber tones with soft textures, evoking a cinematic, elegant portrait feeling.
? 推荐不想本地自己折腾的同学一个可在线使用Runninghub平台可在线体验AI应用和工作流(注册即送1000积分可用)。主页更多精彩工作流可在线体验: https://www.runninghub.cn/user-center/1890418187312222210?utm_source=kol01-RH059 。阿里万相-最强开源图生视频AI应用:https://www.runninghub.cn/ai-detail/1894632237306937345?utm_source=kol01-RH059 。AI工作流:https://www.runninghub.cn/post/1894584540348743681/aiDetail?utm_source=kol01-RH059
? FLF2V首尾帧:迎来ComfyUI原生支持!98%首尾帧匹配率,视频抖动率降低37%
? [ComfyUI]UnionPro2.0:更强多合一CN控制模型重大更新!基于黑森林F.1文生图姿势&深度&线稿等控制感兴趣加入[AGI技术交流群]+V
如果觉得文章不错,就请赞、在看与转发三连