jabir-zheng / tcd Goto Github PK

Thank you for publishing this. This result is quite awesome and it's great to see how the details can get such X10 improvements!

Can you release your training code ?

Thanks for this amazing sampler, it's way better than LCM in my estimation.

The file size of the LoRA could be much smaller with no ill effect however.

Full rank file size: 375.6mb
Resizing to rank 4: File size 23.8mb
Average Frobenius norm retention: 91.92% | std: 0.101

Resizing to rank 2: File size 12.1mb
Average Frobenius norm retention: 88.57% | std: 0.140

this is the repo: https://github.com/JettHu/ComfyUI-TCD

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?

Hi, awesome work! I had a question with regards to the distillation algorithm for TCD (Algorithm 1/2 in the paper, particularly w.r.t. Eq. 21). In the original LCM paper (to my understanding), the skipping step $k$ denotes the size of the single-step ODE solve used by the teacher model to solve from $t_{n+k}$ to $t_n$ (e.g. solving from 950 to 930 using a single step that is sized $k=20$). However, in the paragraph before equation 21 it is noted that $\Phi^{k}$ denotes $k$ "discretization steps" of a one-step ODE solver. Thus, my question is: do you use multiple calls of the ODE solver (with the teacher model) to solve to some timestep between $t_{n+k}$ and $t_m$ (e.g. solving the integral with 2 single-step ODE solves thus two calls to the teacher model), or are you still only using a single ODE solver call across that interval (similar to LCM)? If so, how many? Thank you!

Hi, I find "eta=0" means determined sampling? Maybe it should swap?

Thanks for the impressive work. It would be great if supporting the sampler in SD-diffusion-webui

I used other turbo models to indicate that the generation time is actually longer than the original turbo model, or it cannot be generated
The config attributes {'interpolation_type': 'linear', 'skip_prk_steps': True} were passed to TCDScheduler, but are not expected and will be ignored. Please verify your scheduler_config.json configuration file.

Hi,
The text-to-image results looks great, but inpainting results are extremely blurry (using the default parameters)

I have faced the same effect when tried to train LCM for the task of inpainting

Do you have any intuition on why this happens? Or any suggestion to mitigate it?

Great job! Can you provide a goolge colab demo for using?Thanks!

Thank for your excellent contribution!

I am attempting to implement this for the ComfyUI community, however there are some differences in structure between Diffusers (more explicit time scheduling) and ComfyUI (array of sigmas, and abstracted scheduling).

The tricky part is how to do the adjusted timestep calculation using eta/gamma.
Something is incorrect so the image just goes black after the first step, when eta is > 0.

I wonder someone on your team would be able to look at my code please? (or anyone else here on GitHub who is good with diffusers and comfyui)
https://github.com/dfl/comfyui-tcd-scheduler

Can you criate a extension for A1111/forge to adjust gama ?

Is TCD like LCM with maximum CFG = 2 and ignore negative prompts ?

Hello, thanks a lot for your work, I was wondering if you're going to make a cascade lora since you've made a 2.1 ?

We sadly found out our Consistency Trajectory Models (CTM, ICLR24) was plagiarized by Trajectory Consistency Distillation (TCD)! See Twitter and Reddit.

We are deeply disappointed of TCD author's inappropriate reaction. Accordingly, we reported their plagiarism issue to their affiliated universities, hugging face team, and ICML. *Speaking on behalf of myself

Can an additional Lora model be used on top of TCD-Lora？