LTH14/mar

Issues

• Clarification of speed

  #45 opened 2 months ago by zehongs
  6

• About VAE channels

  #56 opened a month ago by pokameng
  10

• Questions about causal methods

  #27 opened 2 months ago by Tom-zgt
  6

• About the positional encodings of diffusion MLP.

  #54 opened a month ago by whwjdqls
  2

• Scaling diffusion MLP for learning higher dimensional VAE feature

  #55 opened a month ago by zythenoob
  5

• The influence of VAE feature dim

  #53 opened a month ago by Tom-zgt
  1

• About training with cached vae latents

  #52 opened a month ago by RohollahHS
  2

• Challenges in Memorizing Single or Few Images

  #49 opened a month ago by kifarid
  3

• About Encoder and Decoder in MAE

  #42 opened a month ago by RohollahHS
  2

• About Train

  #36 opened 2 months ago by pokameng
  31

• Faster training with fp16 or bf16

  #29 opened 2 months ago by shaochenze
  5

• Request for Causal AR Version Release

  #39 opened 2 months ago by aengusng8
  3

• Diffusion MLP initialization and training problem

  #46 opened a month ago by Kevin-thu
  1

• Reproducing the BASE model.

  #48 opened a month ago by cxxgtxy
  18

• model and training code for the AR variant

  #34 opened 2 months ago by MikeWangWZHL
  5

• Dataset used for training...

  #50 opened a month ago by Nobody-Zhang
  2

• Reconstruction loss in ELBO

  #51 opened a month ago by Paulmzr
  2

• Small Issue/error in the code

  #47 opened 2 months ago by niklasbubeck
  1

• Add HF integration to MAR

  #32 opened 2 months ago by jadechoghari
  14

• About masking ratio

  #43 opened 2 months ago by RohollahHS
  1

• Training Problems

  #44 opened 2 months ago by drx-code
  1

• Unable to calculate FID

  #41 opened 2 months ago by xbyym
  2

• Doesn't work well in speech generation task.

  #35 opened 2 months ago by FacePoluke
  8

• Difference Between MAR and MAGE

  #30 opened 2 months ago by JeremyCJM
  5

• CFG for cross-entropy

  #38 opened 2 months ago by shaochenze
  2

• About Training Loss

  #37 opened 2 months ago by Ferry1231
  4

• Is Autoencoder ok?

  #33 opened 2 months ago by Ferry1231
  1

• The CFG strategy - linear. vs constant

  #31 opened 2 months ago by yuhuUSTC
  9

• Question on the Value of Training Loss for DiffuLoss with MAR and Causal Methods

  #20 opened 3 months ago by bugWholesaler
  18

• why the inputs of diffusion are multiplied by diffusion_batch_mul?

  #14 opened 3 months ago by Erisura
  3

• generate images with arbitrary resolutions,

  #28 opened 2 months ago by Leiii-Cao
  4

• Inference details of an ablation experiment.

  #26 opened 2 months ago by tgxs002
  1

• VAE decoded as NaN in early stages of training

  #25 opened 2 months ago by xiazhi1
  2

• Train Code for VAE Used in Paper

  #19 opened 3 months ago by Ferry1231
  16

• Buffer Size for Class Condition

  #21 opened 2 months ago by zhuole1025
  9

• FID evaluation reference data

  #8 opened 3 months ago by MArSha1147
  18

• Why main_cache do not use flip augment?

  #24 opened 2 months ago by xiazhi1
  2

• How should inference be performed when using VQ-16 (discrete)? During decoding, should we use the AR output for VQ and then decode?

  #23 opened 2 months ago by Tom-zgt
  2

• About the mask schedule during training

  #22 opened 2 months ago by zythenoob
  4

• MAR for Image-to-Image Generation

  #18 opened 3 months ago by Bili-Sakura
  3

• Training settings for MAR series

  #16 opened 3 months ago by HuangOwen
  2

• Latent Dimensions of VAE

  #15 opened 3 months ago by Vinnieassaulter
  4

• Loss for training KL-VAE

  #13 opened 3 months ago by Vinnieassaulter
  1

• About the "per-token" distribution modeled by diffusion model

  #11 opened 3 months ago by chrisway613
  4

• Training epochs

  #12 opened 3 months ago by sihyun-yu
  2

• Class-Conditional Free Text-Guided Generation

  #4 opened 3 months ago by Jake-wei
  4

• Why not whole DIT block?

  #10 opened 3 months ago by WeitaoLu
  2

• Is diffusion position embedding necessary?

  #9 opened 3 months ago by chrisway613
  2

• Generation FID is much lower than Reconstruction FID for models using VQ-16 (discrete) provided by LDM codebase

  #6 opened 3 months ago by ShiFengyuan1999
  4

• The Impact of MLP Depth

  #5 opened 3 months ago by Robertwyq
  2