复现妙鸭相机证件照生成的pipeline
结论: V2的效果最好,V3的速度最快,但是面临妆容和ID不够匹配的问题,如果能训练一个 天真蓝/海马体/古装照 半身lora作为基底模型,效果会更好
此后开源的一些工作:
- 2023.8 facechain 工具链,解决一些工程问题,在训练lora前处理增加了landmark对齐旋转
- 2023.9 sd-webui-EasyPhoto 易用性改进,增加通过landmark对齐+openpose controlnet监督inpainting过程
Head/face inpainting based on sd-webui and lora models from civitai:
origin | face inpainting | face&neck inpainting |
---|---|---|
![]() |
![]() |
![]() |
- using chilloutmix as base model (realistic asian women)
- using tagger(wd14) for prompts reverse
- using textual inversion negative prompts (ng_deepnegative_v1_75t,badhandv4)
- optim mask blur for cheek edge >15
- DPM++SDE Karras sample step 30 for speed up
- controlnet reference 0.5
- controlnet softedge_pidinet 0.75
- controlnet openpose face 0.7
Fix face color different from background, more realistic
face&neck inpainting | img2img depth | img2img depth&softedge |
---|---|---|
![]() |
![]() |
![]() |
- have tried more realistic/style/portrait lora(not shown here)
- 由于没有风格lora约束,人物lora会带入风格(假设该lora训练时拟合了五官以外信息)
Only do one time img2img and can be faster.
origin | img2img | faceswap + img2img |
---|---|---|
![]() |
![]() |
![]() |
- using inswapper for faceswap
- private face swap model (not shown here)
- reference/softedge_pidinet/depth to control details
using 20 upper body images to train a face lora model
20张半身像(参考妙鸭)稳定训练出一个比较好的lora还是有一点困难,目前一些tips:
- 产品逻辑上,参考妙鸭,四组参数训练四个模型供用户选择
- 有人在用证件照训练lora底模 知乎
- 不做任何优化的前提下,20张图片A100大约20分钟收敛(batch4 不考虑显存和多路复用),还可以更快(降低rank等)
- 专注训练脸部lora(加入五官提示词)
- 一些更新的社区工作(Lion SDXL LyCORIS等等),lora在同类型图片(半身像)上预训练后可能会提升效果/减少训练时间
- base model from chilloutmix
- lora from on iu
- origin photo from 小红书
- pipeline extract from sd-webui based on sd-api
- lora training deploy%speedup
- 妙鸭数据爬取和对应的lora风格训练