Luo Cong

Training on vg dataset

1

Thanks for your excellent work! According to your paper, pairnet performs quite well on vg dataset for sgg task. Could you provide some instructions on how to train the model...

I find in the "stage1.yaml" file `max_train_steps` is set to 300,000, and I want to ask is it really necessary to train such a long schedule? In animate anyone paper,...

In this repo `beta_end` is set to 0.020 for sdxl noise_scheduler: https://github.com/guoyww/AnimateDiff/blob/8d191074a10a708d9a8fb4701ffd55a1e1c2f689/configs/inference/inference.yaml#L24 but in original sdxl noise_scheduler setting, `beta_end` is set to 0.012, refer to [this](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/462165984030d82259a11f4367a4eed129e94a7b/scheduler/scheduler_config.json#L4). So is this because...

tensor shape mismatch error when running `null_inversion.invert()`

3

when I do `null_inversion.invert()`, following error occurs: ``` Traceback (most recent call last): File "/home/luoc/code/UniEdit/uniedit_a.py", line 157, in (image_gt, image_enc), x_t, uncond_embeddings = null_inversion.invert(None, prompt=config.inverse_prompt, offsets=(0,0,0,0), verbose=True) File "/home/luoc/code/UniEdit/nti.py", line...

Bad reconstruction of ddim inversion

1

Thanks for your great work! I found the reconstruction quality of ddim inversion is really bad(without null text optimization), is it normal?(the middle one is reconstructed video) 

Thanks for your great work! Can you release the evaluation dataset?

About training video length

3

Hello, I noticed that you're able to train on more than 300 frames using an A100 GPU. I'm curious about your training process - are you only training the `to_q`...