tianweiy DMD2 issues - Githubissues

tianweiy / DMD2

(NeurIPS 2024 Oral 🔥) Improved Distribution Matching Distillation for Fast Image Synthesis

Other

524 stars 28 forks source link

issues

Newest

Newest Most commented Recently updated Oldest Least commented Least recently updated

don't support both options fsdp and gradient_checkpointing . why?

#52 yanmj0601 opened 1 week ago
2
any loss curves for the first phase of training

#51 SummerTrains opened 3 weeks ago
5
Why the ablation studie of No Regress. is different from the result reported in DMD?

#50 YuChen-Liang closed 1 month ago
1
The DMD loss increases during the first stage of training.

#49 CS123n opened 2 months ago
2
About Loss

#48 joelulu opened 2 months ago
0
train a 4step SDXL got CUDA error: no kernel image is available for execution on the device

#47 yuanzhi-zhu closed 2 months ago
6
why divide by the mean of torch.abs(p_real)

#46 PeiqinSun closed 2 months ago
2
Why use the intermediate estimated x_0|t as the distribution matching input， but not the final image of generator?

#45 vv12kant opened 3 months ago
2
Support SD3？

#44 mliand opened 3 months ago
1
DMD2 for Flux Dev

#43 pandayummy opened 3 months ago
2
Unable to replicate results for sdv1.5

#42 omrastogi closed 3 months ago
4
train a 4step SDXL got AssertionError: Invalid: ``_post_backward_hook_state`

#41 joelulu opened 3 months ago
8
FSDP is too slow, why not use deepspeed?

#40 sunhaha123 opened 3 months ago
1
is this exepcted when you are training sdv1.5 in stage2: gen_cls_loss (g loss) keeps increasing while guidance_cls_loss (d loss) deacreases.

#39 fire2323 opened 3 months ago
9
When I was training with cls_on_clean_image, I got an error: Given groups=1, weight of size [320, 4, 3, 3], expected input[1, 3, 512, 512] to have 4 channels, but got 3 channels instead

#38 koking0 closed 4 months ago
1
DMD2 and Kandinsky, or Playground 2.5?

#37 Quick-Eyed-Sky opened 4 months ago
0
muti-step model for sdv1.5

#36 kmpartner opened 4 months ago
5
LoRA training

#35 SYZhang0805 opened 4 months ago
4
how long is the extra regression loss pre-training in 1-step SDXL distillation trained?

#34 grovessss opened 4 months ago
1
When i applyed the dmd loss to velocity prediction , the loss varies a lot。

#33 icelighting opened 4 months ago
1
fsdp with gradient checkpointing

#32 LifuWang-66 closed 4 months ago
1
After training, there are two pytorch_model files. Which one should I use?

#31 koking0 opened 4 months ago
5
The arg "cls_on_clean_image"

#30 SYZhang0805 closed 4 months ago
10
resuming from the best checkpoint and training for an additional 2,000 iterations.

#29 SYZhang0805 closed 4 months ago
1
prompt input to unet

#28 LifuWang-66 closed 4 months ago
3
Is this method suits for velocity prediction?

#27 icelighting closed 4 months ago
2
Question about gen_cls_loss and guidance_cls_loss

#26 weleen closed 4 months ago
2
When i train sdxl model, saving fsdp model case training stop

#25 icelighting closed 4 months ago
3
Distillation questions

#24 Bigfield77 closed 5 months ago
4
is there plan for releasing code for grad accumulation?

#23 grovessss opened 5 months ago
4
Generating VAE latents for a different architecture

#22 sborse3 closed 5 months ago
4
set model to train mode

#21 yuanzhi-zhu closed 5 months ago
8
Convert 4-step unet into a checkpoint file

#20 coastbluet opened 5 months ago
6
When i train SDXL-4steps model, i get OOM error, i use a single A100.

#19 bengen-y closed 5 months ago
2
sdv15 train with torchrun successful, but sdxl train with fsdp failed.

#18 conquer-pan opened 6 months ago
0
How about DMD2 on img2img task?

#17 dawei03896 opened 6 months ago
2
How does it compare to ByteDance/Hyper-SD or sd-community/sdxl-flash ?

#16 wodsoe closed 4 months ago
2
Please tell me where is the data files: sdxl_vae_latents_laion_500k_lmdb? I can not find it.

#15 conquer-pan closed 5 months ago
1
Current CUDA Device does not support bfloat16. Please switch dtype to float16.

#14 conquer-pan opened 6 months ago
0
Update README.md

#13 lmxyy closed 6 months ago
0
Is it possible to integrate with diffusers?

#12 dawei03896 closed 4 months ago
1
request default -1 seed and negative prompt

#11 nogold4554 closed 5 months ago
2
Is it possible to integrate with HiDiffusion?

#10 self-destruction opened 6 months ago
6
Why is uncorrelated noise used in backward_simulation?

#9 XiaXin-Aloys closed 6 months ago
1
Request for test_folder.py in ImageNet-64x64 Experiments

#8 NikkoLKR closed 6 months ago
2
Why add the discriminator header on the fake denoiser and not on the real denoiser where the parameters are fixed?

#7 IDKiro closed 6 months ago
3
Possibility for 30fps image generation?

#6 beyarkay opened 6 months ago
6
a Lora version to add to any Community Models?

#5 xiaoleexiang closed 4 months ago
6
Add Float16 Precision Handling, Tiny VAE Option and Batch Generation

#4 voodoohop closed 6 months ago
1
demo dead 💀

#3 Mike-FreeAI closed 6 months ago
1