@METAFILM_Ai That's not a valid reason to use a .pt file. Nothing prevents you from loading a safetensor file in a standalone environment.

Kijai uploaded a safetensor, https://huggingface.co/Kijai/WanVideo_comfy/blob/main/Wan2_1-T2V-14B_CausVid_fp8_e4m3fn.safetensors

For CausVid, I believe there isn't ComfyUI support yet.

We’ll need to use the official Python-based inference codes

1) Clone https://github.com/tianweiy/CausVid and follow instructions to install requirements

2) Clone https://huggingface.co/Wan-AI/Wan2.1-T2V-1.3B into wan_models/Wan2.1-T2V-1.3B

3) Put the pt file inside checkpoint_folder/model.pt

4) Run inference code, python minimal_inference/autoregressive_inference.py --config_path configs/wan_causal_dmd.yaml --checkpoint_folder XXX --output_folder XXX --prompt_file_path XXX

EDITED: I haven’t tested this myself yet, just gathered the information from the GitHub repo and Reddit post.

Reddit posts about CausVid: https://www.reddit.com/r/StableDiffusion/comments/1khjy4o/causvid_generate_videos_in_seconds_not_minutes/

https://www.reddit.com/r/StableDiffusion/comments/1k0gxer/causvid_from_slow_bidirectional_to_fast/

@METAFILM_Ai where was it uploaded, this explanation is confusing as all hell

@rocky533 I guess he meant this? https://github.com/kijai/ComfyUI-WanVideoWrapper/tree/main/example_workflows Indeed it is confusing -_-

@METAFILM_Ai Aighty. I'll wait for it to be more consumer friendly to use. Thanks for the reply

可以試試這個fork

https://huggingface.co/Kaoru8/T5XXL-Unchained

之前也有解決過類似的問題，印象中還要覆蓋ComfyUI/comfy/text_encoders/t5_config_xxl.json的參數，有可能記錯，請先備份後再試。

@lym0 谢谢前辈回答 我把t5xxl-unchained-f16.safetensors放进了text_encoders资料夹里面 也是一样

不好意思再追加一个问题 Unet加载器有什麽特别需求吗? 我点执行还会显示UNETLoader'conv_in.weight'

@METAFILM_Ai 感謝糾正，確實搞混了。也試過RedCraft沒破解的T5一樣能生成nsfw内容。

My 4090 is out of memory, is my settings wrong or just the model needs more

@zczcg 该模型理论上只需要16Gb的显卡即可运行，建议检查，是否是其他模型被保留在显存内，占用了过多显存？

I second that question. Also, does this work I2V? I've tried on default or custom I2V workflows but it doesn't seem to work this way.

@METAFILM_Ai It's the same model? The model I download from there is named 'onTHEFLYWanAIWan21VideoModel_causvid14bVACE', which is not the model named in the workflow.

@ralphtandy is this the model RED CausVid -Wan2-1 -14B- DMD2-FP8」? Can anyone tell it?

This lora makes everything almost perfect now! Dropped from 190 seconds of generation to 90 s (720p upscaled). Thanks for the tip!

@METAFILM_Ai Thank you, i advise adding this to the description. No one is going to think to download training data for the workflow. Thanks for the response.

@METAFILM_Ai is not my torch or cuda version wrong?I try both cu12.8 and cu12.6, but in vain,btw,i can understand chinese,thks for reply,I will try it tomorrow,

@METAFILM_Ai i have no change with the sample workflow,but can't work

I find the reason,Virtual memory is set too small. I have set up 7000mb,is really Virtual memory important?

@METAFILM_Ai  我找到原因了，剛剛我把虛擬記憶體從7000mb拉到70000就解決問題了，虛擬記憶體有這麼重要嗎?不太懂。4090不就有24gb可以用了，而且我去看資源監控器，頂多專屬gpu記憶體也只有吃18/24GB，根本沒吃滿，是設定問題嗎,換句話說，他根本使用虛擬記憶體去補足的

@zczcg to be honest it is not. Problem is resolution. If you go higher resolution, VRAM fill be filled and it would like to use more ram where it is virtual ram. If you set your settings optimal, you won't be needing virtual ram whatsoever.

@Reelai Thanks for replying!! And how to set settings optimal?

In the training data folder. Took me a while to find as well lol

@METAFILM_Ai I didn't know that, it's hard to find if you're not used to it. But thanks for the work you guys put into these models, it's very much appreciated

For me there is just a download button, no training data....

@anno314 Right side. Below the "details" section. Above the "about this version" section

it's on the right, below the "details" section

@notme123123 isn't that just the model link? I am also too blind to see it

@NoMoreRice 

https://civitai.com/models/1295569?modelVersionId=1781040

you can find Training Data (8.9 MB)

@NoMoreRice First, click on the CausVid14B model at the top, right below the main title of this page. Then, look below the "details" section on the right side of the page, there are four files there. One is called "training data", that's where the workflow is

Nothing different outside of the fact that some chinese pieces of shit keep wasting peoples time with a base model uploaded in their own way.

I want to know that too

Honestly I think there is nothing different in terms of I2V. Just a complex workflow that makes 0 sense to any human brain compared to 99% of workflows. And the latest seems to give 0 fucks about any input image in any other workflow. Short answer seems to be this is just a waste of time to download. Idk maybe if you are some autistic nerd that only understands 0's and 1's maybe there is some difference but to anyone else. This just seems to be a waste of time. Hell they dont even say if vace is FP8 or FP16, like how is this being prompted lol. Chinese Trash with some data center farm pushing this garbage to the front.

@METAFILM_Ai thnx for the response.

@METAFILM_Ai 我生硬了

37 years with a GPU from 1970. 10 seconds with a GPU of 2050.
Your question is meaningless, especially since you don't tell what's your hardware...

@METAFILM_Ai thank you, much apreciated

@METAFILM_Aisounds like its going to render quite quickly then, im used to older methods that took ~5 min for an ~80 frame video

1 frame/s rate is insane with a 4090/4070 

@SD_AI_2025 mate, fk off with your bs attitude

your better option is to keep quit

hello, can you please post a working Workflow?? the the ones included here are confusing

@METAFILM_Ai Thank you for the clarifying information, I really appreciate it. Also, thank you for your hard work. Do you happen to know of a workflow for t2v, that can generate fast videos like causvid? Thank you in advance.

Forge has nothing to do with video AI.

FramePack is for Hunyuan.

Since using VACE ive notived the videos barely has any movement. I think it may be due to the fact im using the LORA to speed things up, which requires CFG1.

Its in the training data download, in the files drop down

I assume it's normal. My 3090 (24 GB VRAM) usually creates a video of around 81 frames (5 segs) with I2V-14B-720P in 12-18 minutes - depending of used LoRAs and a lil bit of luck - with use of SLG, Sage Attention and TeaCache. I carefully check that all the model and inference result stay in the VRAM, because if even a lil piece is allocated in the RAM (DDR4 @ 3600), gen time skyrockets for around 3 times.

I'm using the Q4_K_S gguf model and generate a 5 second video at 768x384 with 30 steps in about 10 minutes (20s/it) on my laptop 4080 12Gb VRAM. This is without Teacache and Torch. I do have sageattention enabled and I'm using the UnetLoaderGGUFDisTorchMultiGPU to offload to RAM.

@funscripter627 can you upload your workflow?

Its in the training data download, in the files drop down

rocky533 the only thing I see in the file drop down is Pruned Model fp8 (15.83 GB) which is top right corner, I am rather lost haha

if you download any of the example videos below the model title and drag the video into ComfyUI the workflow is contained in the videos already, and it will deploy into your ComfyUI session.. I did it and it worked for me

Its in the right by the creators name it says where you put the text encoders etc, when you go to hugging face those creators list where to put the files as well. usually models go in diffusion models or checkpoints, while gguf go into unet. But you can also use control+F for find and search the word "folder". Also check the civitai articles section they have guides how to do everything, usually made by the same model and lora creators

Do you have a 24gb card? If not then you might want to look at the 480p models possibly even gguf

Thanks for replying :D, yes, I read that on the right there is that section, but I couldn't understand what specific things to download and where to put them, specifically, since many files have almost the same name and I don't know which one to choose, if I had a workflow, I could specifically search for files by name, but there isn't one. I have an RTX 3090, I read in a previous reply that it can make 5-second videos in 15 minutes at 720p

@tomasvillarrealfederco536 First of all, I agree with you the guide is an absolute mess. It's extremely hard to follow for people with limited experience. But the workflow is provided, it's on the right side of the screen where it list the models and a training data package. The workflow is within the training data.

With 16GB vram, I'd use the Q4_K_M quant version. As a "quick" rule, the larger the file, the better the inference results, but the higher requirements of the GPU.

I use the wan q6 models with 12gb vram...though its nvidia. Always worth a try to find your ceiling with your set up.

I get 6 minute generation time with 16gb vram using the 720p Q3_K_S gguf model at 560x1024, 12steps, 16 fps, and 81 length. My optimizations are teacache, sageattention (gl installing sage on windows) and using causvid lora. 400x720 takes like 3 minutes

bhopping do you happen to have a rtx4060ti ?

1015 4070 ti su

same questions, hope somebody with better knowledge can help.

Or what's the best way to incorporate lora into the default workflow in the Training Data?

The fix is to use the wan video wrapper node to load clip vision for wan. The wan video wrapper is part of the kijai wan video wrapper node pack. Delete your current clip vision node, and when you drag the "clip vision" rope from your clipvision encode node, it will give you the choice of "load wan video clip text encoder" as a node choice all the way on the bottom of the list (usually).

I've done this with an image, but not with a video. How did you do it? Workflow metadata appears to not save to .m4a files (videos) downloaded from this page.

@megaman11 try another video perhaps. I tried it with a video (just the ones under the model´s title) and it worked for me.. e.g. the topless Chinese woman vid

I use topaz video AI that's a separate application but the nodes that ppl include in their workflows are still good, you can look up wan upscale workflows on this site and yoink their nodes and put it in your workflow

Seconded. I have a working Wan2.1 i2v, t2v, and VACE workflows without this... what does this add?

Look at the size, man lol. Ain't no loras that are 15 gigs lmao.

do a search for wan lora. you'll find plenty of them to choose from.

Geez man. At least attempt to learn the program before you make it sound like the Uploader's fault that you don't know the basics of the software. Click the red "X" on the right of the image feed to make it dissapear. Press the "Show image feed" button on the top right of your screen to make it appear.

@jazgalaxy581 And don't comment if you don't know, the button you're talking about is grayed out. If I hover over it with the mouse, the lock icon appears (a circle with a slanted red bar). There's no red X; it's put that bar in all my workflows. Anyway, it was in the options and I managed to remove it. The button you're talking about wasn't there because it was set to "hidden" by default. If we don't know the situation, we shouldn't jump to arbitrary conclusions.

model 480p Here is the video I just made 4s it's not terrible as a result 30 mn :(

It absolutely does. I have no idea what I'm reading.

They're just reuploads

Hi, sorry to bother, can you please post any workflow working with this model?? The ones included here are so confusing

ComfyUI

why the name onthefly? these are just all base models without lightning lora or any boost lora baked in. it's not fast at all.