search

jabir-zheng / TCD

Official Repository of the paper "Trajectory Consistency Distillation"
https://mhh0318.github.io/tcd
290 stars 14 forks source link

About Training Schedule #8

Open SpadeLiu opened 4 months ago

SpadeLiu commented 4 months ago

Hi,

Thanks for your work. I wonder in the training phase, which interval is used to sample 'n' in algorithm 1. 1 or 20 ? In other words, can 'n' be 0,1,2,...,978,979 as in CM or just 0,19,39...,959,979 as in LCM?

jabir-zheng commented 4 months ago

Both sampling schedules are applicable in TCD, and our experiment revealed that the latter yields slightly superior performance.

George0726 commented 4 months ago

Hi jianbin,

I found that TCM Lora cannot work well on Fine-tuned models, which is similar to LCM lora. Hence, I am trying to fine-tune a TCM LoRA on my SDXL models. May I ask if you used dynamic gamma(eta) during training or just the constant value like eta, as you refer to during inference? I got pretty weird results from my basic implementation of your algorithm :).

jabir-zheng commented 4 months ago

Hi @George0726 ,

In our experiments, we observed that TCD works well for various fine-tuned models. If you are willing to share more information about your model and test examples, we may be able to assist you further in identifying the cause.

Regarding gamma, it only appears during inference to control the destination in each sampling step. The training objective is to enable the model to access any point along the PF-ODE via our TCF, hence gamma does not exist during the training process. Please refer to Algorithms 1 and 2 in our paper for detailed implementation.

George0726 commented 4 months ago

Appreciating for your kindness! I will share some information with you privately