pljj315

how to train? is there any code for training instantID?

6

want to re-train this InstantID, but I can't find the ready-made code. is there any code for training instantID ?

File "/mnt/nfs/file_server/public/lijiahui/code/study_consistent/CustomNet/main/app.py", line 194, in main model.load_state_dict(ckpt) File "/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py", line 2153, in load_state_dict raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format( RuntimeError: Error(s) in loading state_dict for CustomNet: Unexpected key(s) in...

用diffusers实现"Tencent-Hunyuan/HunyuanDiT-Diffusers"模型，只能简单prompt，否则会出图错误

3

简单prompt "a girl"能正常出图， 换成下面图片中的较长一点的prompt就有问题？ 代码见下图   

about training： how to train the envhdr_MLP?

1

按照论文的方法进行训练（8c_unet与envhdr_MLP的参数一起进行训练）最终得到的envhdr_HDR的作用微乎其微，输入envhdr并不能进行打光的控制。所以我很好奇，在这一部分的训练过程中是有什么trick么？怎样训练才能让envhdr_MLP真的提取到环境贴图特征并很好的影响到生成中去呢？

Questions about image_based_lighting: 1. In the supplementary material, the description of creating shadow_map with normal is not clear. Are the terms' shadow map 'and' shading map 'problematic in this paragraph?...

qwen-image-edit lora训练 是否有显存优化手段

4

使用单机8卡49GA6000，按照repo中的lora去训练，已经启动了 --use_gradient_checkpointing --use_gradient_checkpointing_offload --enable_fp8_training，降低了lora_rank，zero_stage2/3都试过，全都在**accelerator.prepare步骤爆显存** 求各位指导： 1. 是否是deepspeed配置没起作用？配置方式见下图（--config_file） 2. 是否有其他方法能够优化显存？