Innovate Futures @ Benji patreon

Native Wan VACE - Long Video Gen (Unlimited Length) - Version 20250708

🕑 Added 2025-07-07 19:31:09 +0000 UTC

Comments

Franco Aiello

Hey! Great workflows. Love your work. v2v worked perfect for me. Now when using i2v. It seems like it's not working with the reference image and giving me a prompt base output only. I'm using the Wan2.1_T2V_14B_FusionX_Vace-Q8.gguf just like you. Haven't modified the workflow at all. Any ideas what might be happening? Thanks!!

Benjamin Law

You need something like iClone then. Yes , open source Controlnet is like a toy grade. When you go for more detail like cg movie, that is another level

Pedoneit

I feel like the controlnet nodes that we use in our workflows are pretty basic (I also use these nodes thanks to your idea). I would like to find a way to be able, not only to replicate movements but also replicate much more detailed expressions. Sometimes I get the eyes look at somewhere in a weird way, or get non-human gestures, etc. I am looking for a more advanced controlnet technique or whatever it's called

hf21

The long video Gen works perfectly (i did one with 11 parts) but this one with loop does not work, after the first video part there is a big glitch and the video result after that is ugly

Benjamin Law

Its normal, because the overlap frames are coming from the previous batch. If you use it in the sitiching, your final video will get weird. Btw, 20 is too much. As I mentioned 6-10 is good enough

Кьюбит Шоу

every time I create a video and set the "Overlap frames" to 20, for example, I get a video in which these 20 frames are cut after each iteration.

AARON SOF

If I join-up is there any kind of faceswap vid2vid you can show me??

Benjamin Law

Length of the video missing in some part in your setting/ process

Jim

hello. i am new to this. i saw your vid and downloaded the workflow. but i get this error when i try to run. the only change i made was the upscaler. i used the simple one since for some reason i couldn't install it properly. (Prompt execution failed TypeError: Cannot read properties of null (reading 'length')

Vicente

I'm getting better results each time depending the lora combination, very impressive! There's something i don't underdstand though, on the far right of the workflow there're 2 nodes, the first "video combine" node that produces a video, and if i understand correctly, the final one at the end on the right that connects the upscale node and should produce the final, upscaled video. However i'm only getting a still image in that final "video combine" node, is that normal?

Vicente

I think this is your first workflow that works for me, almost all i had tried before gave me some form of OOM or other error. This one gave it too, but lowering resize_to HD instead of defeault 2k fixed it. Amazing tech here, 12 sec vid in 11min with great motion and okish image quality, which is probably related to the lora selection i have used which could be better, so i'll tweak some of this stuff now that this works. Overall amazing job, thanks a lot for this amazing piece of AI tech.

Benjamin Law

Self Forcing have 14B? Are you talking about LightX2V lora? If that so, that is just a change in the model and lora loader drop down. Doesn't need additional workflow change, if you really know what are you asking.

2X

Can you add self-forcing 14B when you have time, and adapt the workflow to only sample once? That would be a great performance improvement.

Benjamin Law

you can DM me in here or Discord. Since, I have my own company running. I am pretty much busy most of the time, and doing AI content here for hobby. :) I have to check how many in the work load and what is the budget first.

DBJ19

Probably have to set up an offload as you go

DBJ19

Hi Benji, is there a way to contact you directly? we are working on a project and might want to hire you to consult. Thank you!

Nocare

This has been really cool. It seems RAM ends up being the limiting factor. With a long video I eventually run out of ram and comfy crashes.

Benjamin Law

you mean yours?

Dominique B

all video outputs are blurred

Yih-Jia Tan

this is so good! thanks for all the math work behind!!! would like to ask, using the i2v workflow, does lora work for native Wan Vace? I experience lora key not loaded mainly like motion driven Lora. enhancement Lora (example: detailenhancement) sometimes loaded ok, and if there is the lora key not loaded error, i close and restart my comfyUI fresh then the error seems to go away. Most motion driven lora i use is for i2v, some t2v, using on kjwanvideowrapper is working fine. Model wise i use T2V or I2V model. i tried to test on VACE but due to Vace module and T2V model load separately in WanVideoWrapper, i get OOM once I add in Lora. Not sure what is causing it, maybe you can shine some light for me :D i am using this setup Wan2.1_T2V_14B_FusionX_VACE-Q6_K.gguf umt5_xxl_fp16.safetensors

Benjamin Law

v2v have controlnet, and i2v just use the image, how can you have the same output?

Al Go

Hi. Could you please elaborate the difference between the v2v and i2v of yor latest iterations? The only difference I spotted id that v2v of them has prompt (which is completely ignored) and throws loopEnd exception after first batch of 81 frames (for the same input video/image as i2v)... The exception is: easy forLoopEnd '626.0.0.3.0.0.655'

Benjamin Law

Might need to check your reference video size . Looks like after resize the mask are not the same res.

PeterFuchs

I get this error when I process a video thats 201 frames long and the frame length is 81: IndexError: The shape of the mask [81] at index 0 does not match the shape of the indexed tensor [49, 480, 832, 3] at index 0

Clayfacer

Wow much better. Thank you

Benjamin Law

300+ fps?😱😱😱that is very very fast like 100M running speed, are you sure?

Rahul Kumar

I am getting this issue The shape of the mask [81] at index 0 does not match the shape of the indexed tensor [37, 832, 480, 3] at index 0 my video is 300+ fps

Benjamin Law

Its based on the video length now. You don't need to set. But its better the video is over 10 seconds let the beast run.

Rahul Kumar

You got an amazing piece of brain Benji. Thank you so much!

Benjamin Law

how long is your video?

A Syed

how do you set the # of loops, or the length of the video?

Al Go

Thank you Benji. Constantly hitting "start index is out of range" from "Get Image or Mask Range From Batch" node (one, connected to the "DWPose Estimator")... it calculates start index as 152, while num_frames is 81... any suggestion?

Lionel Vinuales

Benji, you are truly amazing ma man. Thanks a lot.

Benjamin Law

Haha thanks, you can try this one rather than the last version. but just be careful, summer time, have a extra cooler blow that gpu. Lol cause when running this workflow, its hot. Eyes, it need close up shot and use the face detection (i forgot the name, so many AI..) And then combine with DWPose.

Temp Mail

Hey benji.. You are such a gem always comes with something exciting ❤️❤️. I'm also keep working on your last long length workflow it was working fucking good and I ma damn sure this one gonna boom the whole journey. I have a question can you do something with eyes moment this is something which isn't getting controlled. I have generated more than 20 videos on different settings and value but everything it gives robotic eyes moment. Can you look into? Because this is something I'm not able to control and it's obviously differentiate between real video and AI generated videos.

Benjamin Law

I think this to be more dynamic way now. So we don't have to keep having so much nodes... Make me dizzy lol

Intellectz Productions

This looks incredible. I want to see how this works later on when I can try it. This is how you move things forward. Great work. I have been working on a V2V custom temporal fix for Wan 2.1. it's a very meticulous set up but it produces really good results with Low Quality videos. I'm still working on it but I am very happy with how it's turning out.

Benjamin Law

Use this gen movement and bring to the v2v mutlitalk that I posted previously, and its done.

Benjamin Law

This update I solved it, and run few video, using color match it fix. If I use Self Forcing 1.3b.... yes it works without change all workflow structures, just change the model. It is consistent, but the quality is very low. So I haven't use self forcing 1.3b after all.

AiInfluence

Can i use this workflow to give body movements to a lip synced video?

Hayden Martin

I got very close to solving color shift by sending in my reference image alongside all my FFLF controlnet frames and masks. I’ll probably make a video about it - you just need to manually trim the first 4 frames of the output - Trim Latents node makes for ugly start frames