Open ybu-lxd opened 1 month ago
CiaoHe
commented
1 month ago sure, we have basic results on XL-2
| ckpt(cfg=1.0) | FID-50K | sFID | IS | P | R |
|---|---|---|---|---|---|
| DiM-XL-2 (1.7M steps) | 11.16 | 5.99 | 102.5493 | 0.68228 | 0.6382 |
But to be honest, the mamba is quite unstable for large model parameter training. But for smaller size (S / B), the convergence speed is really sharp.
ybu-lxd
commented
1 month ago It's good to know that you've achieved basic results on the XL-2. Your conclusions are very valuable to me, especially regarding the values of FID-50K and IS. Thank you again for your selfless sharing
CiaoHe
commented
1 month ago It's good to know that you've achieved basic results on the XL-2. Your conclusions are very valuable to me, especially regarding the values of FID-50K and IS. Thank you again for your selfless sharing
If you want to explore further about Mamba in the diffusion area, you can contact me. Maybe there still is some space to exploration.
ybu-lxd
commented
1 month ago Thank you for your reply, I will conduct experiments on my data set immediately.
ybu-lxd
commented
1 month ago It's good to know that you've achieved basic results on the XL-2. Your conclusions are very valuable to me, especially regarding the values of FID-50K and IS. Thank you again for your selfless sharing
If you want to explore further about Mamba in the diffusion area, you can contact me. Maybe there still is some space to exploration.
May I ask which data set you did the training on? I did not achieve good results when training the XL structure on my own data set. Also, can you provide the weights? I would be very grateful.
CiaoHe
commented
1 month ago sry we cannot provide with ckpt. we trained on IN-1k
获取 Outlook for iOShttps://aka.ms/o0ukef
发件人: ybu-lxd @.> 发送时间: Monday, May 27, 2024 4:40:21 PM 收件人: CiaoHe/DiM @.> 抄送: He Cao @.>; Comment @.> 主题: Re: [CiaoHe/DiM] about Image sampling (Issue #1)
It's good to know that you've achieved basic results on the XL-2. Your conclusions are very valuable to me, especially regarding the values of FID-50K and IS. Thank you again for your selfless sharing
If you want to explore further about Mamba in the diffusion area, you can contact me. Maybe there still is some space to exploration.
May I ask which data set you did the training on? I did not achieve good results when training the XL structure on my own data set. Also, can you provide the weights? I would be very grateful.
― Reply to this email directly, view it on GitHubhttps://github.com/CiaoHe/DiM/issues/1#issuecomment-2132964567, or unsubscribehttps://github.com/notifications/unsubscribe-auth/AJJ3OE7RS4R37KJ4J3Q55H3ZELWPLAVCNFSM6AAAAABHRZD6ACVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDCMZSHE3DINJWG4. You are receiving this because you commented.Message ID: @.***>
Thanks for the Mamba based diffusion model code. Has your team successfully used this code to generate images? Thank you again for your selfless sharing and I look forward to your reply. Wish you the best at work!