dreamyou070

Asking for image layer in Unet training

1

Hi, I want to follow training AnimateDiff. In the tutorial (https://github.com/guoyww/AnimateDiff/blob/main/__assets__/docs/animatediff.md) first I finetune with configs/training/v1/image_finetune.yaml config. However in Paper, there is no unet finetuning process only Lora training for...

why training ...?

1

in the paper, you found the unimportant SD block/layer. In that case, you may not have to retrain the model (because if you erase unimportant block/layer, the performance is almost...

asking about LCM lora

1

in the paper, you make image consistency model with lora, is that the lcm_lora ? 

Asking about SVDSolver

1

when you train LCM_svd, you set svd_solver like, svd_solver = SVDSolver(args.N, noise_scheduler.config.sigma_min, noise_scheduler.config.sigma_max, 7,0.7, 1.6) why you change training timestep to args.N (for example 40), can I get advice about...

why not public LCM Stable Diffusion T2I model?

1

In the paper, T2V or I2V LCM model is distilled from T2I LCM model. But I thinks there is no training code or model... Can you check once ?

In the paper, you project motion magnitude to the position embedding. I am little confusing... motion magnitude means hoe the region (from the gt dataset video) is movable. is that...

size issue with AnimateLCM I2V model

2

when inference AnimateLCM I2V, the recommended size is (768,512) However, it is not possible to inference on A100 GPU. In the paper, you trained on A800 GPU.. is there any...