大老： 测试中的一些问题 / Some issues during testing

[gg22mm]

Thank you for the project, it's definitely worth learning from~~ / 感谢大佬的项目，非常值得学习 ~~!

大老： 测试中发现这些问题，其它的都可以跑通没问题 / Da Lao: These issues were found during testing, but everything else can run smoothly without any problems ~~

1 Super-resolution(超分辨率)

我发现： Super-resolution超分辨率 少了数据集调有代码 / Old man: Super resolution is missing, but there is code for adjusting the dataset:
https://github.com/CreamyLong/stable-diffusion/blob/master/configs/retrieval-augmented-diffusion/768x768.yaml

768x768.yaml no data and lightning

2 Layout-to-Image(布局到图像):

数据集用的是 / What is the dataset used for： https://www.kaggle.com/datasets/wangrui071/mscoco-2017-trainval-annotations
python mail.py --base configs/latent-diffusion/layout2img/coco_sg2im_ldm_Layout2I_vqgan_f8.yaml --train True

'''# 这里有问题
self.transformer_blocks = nn.ModuleList(
[BasicTransformerBlock(inner_dim, n_heads, d_head, dropout=dropout, context_dim=context_dim[d],
disable_self_attn=disable_self_attn, checkpoint=use_checkpoint)
for d in range(depth)]
)'''

    #wll 修改  context_dim=context_dim[d] 改成 context_dim=context_dim[0]，这样是没报错了，但是不知道有没有别的问题,也有可能是我的数据集有问题，只否提供一个测试数据集呢？
    self.transformer_blocks = nn.ModuleList(
        [BasicTransformerBlock(inner_dim, n_heads, d_head, dropout=dropout, context_dim=context_dim[0],
                               disable_self_attn=disable_self_attn, checkpoint=use_checkpoint)
            for d in range(depth)]
    )

3 Semantic-to-Image(语义到图像):

python main.py --base configs/latent-diffusion/mask2img/config.yaml --train True

少了：flickr 类 / Missing: Flickr class

4 Depth-to-Image(图像深度 , 未开发 Image depth, undeveloped)

[CreamyLong]

Super-resolution

[CreamyLong]

Depth-to-Image. Currently, I may not have time to implement it

[Tonsty]

@CreamyLong, have you reproduce layout results as in the original stable diffusion paper?
I used your layout config and have trained for 10 epochs （COCO dataset，batch size=16）, but the log images obtained in training phrase are as follows,
which I think are not correct training results. The training loss have awalys keep around 0.27 not decreased during the whole training process.

input bbox

input image

decoded recontruction directly from first-stage embeded latent

sample image from latent diffusion model (ddim_step=200, eta=0.)