think i got it fixed, tried updating comfy to nightly like u mentioned and now that node is no longer highlighted red

k that didnt fix it i redownloaded and set up your latest. could it be where i changed the lora in the bottom loader from jellyhips to none?

ya not having any luck. : "NoneType' object has no attribute 'lower'" -and as soon as i hit run 2 nodes turn red highlight one is the preview on the far left of workflow and the other is the one i mentioned earlier

@K3NK  I assume you mean the "batch png combine for video extend" option? (its the nodes I see on the upper left) its 4am here for me so I can't confirm my usage problems being fixed anytime soon once you confirm or deny.

If that other method was supposed to overload me, it never did. God bless my 5090.

@K3NK well I "fixed" the issue. Apparently you need all three extend option enabled for combine to work. Thank you for your help, I appreciate it. much love.

Hi, @badhandproductions . I enable the upper left button, and run a task. Why can I not find the folder storing the PNG files? Where did you store them? Or how can I specify the directory?

@K3NK Do you have like a discord or some way you prefer to chat on the topic of this workflow? There are some paths, like the Extend sections I just cannot get to work, that I would love to pick your brain on regarding settings to get the best outcome.
For example, anytime I try ANY of the extend sections, I get an OOM error, even though my system has RTX 5090 and 48 GB DDR 5. I also have a VERY generous pagefile allocation. I have also tried shrinking the resolution LOW..we're talking 256x256 or 240x360, etc and still having no luck. It gets to the part where it outputs the last frame as the input for Extend_01, but then never gets past it. I love this workflow, the complexity is growing on me and making more and more sense every day, but I just can't get the extended sequence stuff working unfortunately.

@K3NK figured it out, thank you very much)

now everything works as it should, this is a great workflow, thank you very much

@mr1sokoo230 can you please share what have you changed in the workflow for best results (and no oom) on a 3090 ?

@cucurucu i believe the error is in the Wan I2V extend 1 section. Both the high and low samplers there have 32 blocks of swap, which might be draining the system Ram. Drop that to 20 or less and it should work

check the models and lora you have are the same in the nodes. since in the template there is a lora that i could not find just make sure you remove thigns you dont have or load your own. and yes its tons of new stuff but download what its on the workflow and it shuold work is not that hard.

@bahaba7208305 Ok fixed It wanted to update Wan Rapper and obense or something in the manage issue resolved on the the next.

Please forgive me as I'm still learning a lot about comfy and the various nodes and workflows but I'm trying to research without a ton of questions to the community. I'm stumped also on being able to run this workflow with many runs qued. Typically I'm able to just adjust the run count at the bottom, run, then load a new input select the run count run and stack up a bunch of runs. On this model, it seems to drop out many of the runs or do a single run per image. Any advice on what to change within the workflow would be greatly appreciated. Thank you!

@K3NK I don't think my issue with the que's is from freezing or becoming unresponsive. It seems as if when the load image node isn't activated or reloaded between runs, that the run jumps straight through to the preview/save image step and skips everything else. If i toggle in a new image, click run, load new image, run I can get a que going. Thoughts on what might be happening?

I had the same issue from point 1. You can replace it with the FaceRestoreCFWithModel node from the "Facerestore CF (Code Former)" pack.

I dont get it. I get the same error when using context windows. What do I need to do to fix that?

Sizes of tensors must match except in dimension 0. Expected size 80 but got size 128 for tensor number 1 in the list.

Think I figured it out. I needed to connect the "Resize Image V2" node to the "Wan Video Encode Latent" in the context option and now it works.

@K3NK i suppose i could like it if i could understand it. Maybe there are som interesting creative ideas. You know, not all people are here for a quicky. You could be a better engineer if you had ability to present your work in a correct way. But maybe you dont care. if so OK .

@wylmquest Seems more likely that you're just a fucking noob

Yes your 3090 runs on the Ampere architecture which doesn’t support that

@Melty1989  So how are people running it on Ampere? Do I have no option other than switching to a GGUF workflow or turning off triton and sageattention?

update: gguf works just fine

@jonnoel145658 You just change to fp8e5 in the quantisation dropdown in the model node: I had a 3080 and did just that. GGUF isn’t needed for you since you have good VRAM

@Melty1989 Thanks! Fair enough, a bit of a shame Ampere does not have acceleration for fp8, but it is what it is.

@K3NK Sounds great! Do you think you could also make v2.0 of this advanced workflow run on 24 GB RAM by default so it’s easier for regular users to use?

@K3NK I also solved the problem of quality loss from segment to segment - how can I get this edited workflow to you?

@m0no1 upload a media with the WF attached

@juliusmartin 
> 24 GB

> Regular User

wtf did I just read

@m0no1 Now I am interested. How did you solve it?

@haenlesn937 Reference frame reinjection in every segment based off the I2V image and upscaled/interpolated -2 frame crop from GIMM as the first frame starting point, though RIFE 49 is better for this task to be frank, with x2 upscale before interpolation you get cleaner details. I’ll post and link a video later tonight.

@K3NK haha, was just looking at MMAudio for the last step audio dub + muxing.

@haenlesn937 To be frank, I went to 32GB vram and feel limited. I can use fp16 models without quantisation (MUCH better quality) but even then I need to use sage attention and I still OOM sometimes past 720*1280 if I bump fps past 20. 80GB in ram with some workflows. K3nk’s workflow keeps my overhead much lower and I can power through 4x81 frames at 512*860 in 13 minutes, I’d still consider 96gb+ vram card for LLMs

https://civitai.com/posts/22429354 - edited workflow is embedded. Thanks @K3NK

@m0no1 404 for me. Ty for sharing

@juliusmartin https://civitai.com/images/101096736

@K3NK Then Civitai is buggered. Go to my profile and my videos and the latest one of the redhead strapped to a chair

@m0no1 pastebin or continue having fun with us

@K3NK https://civitai.com/images/101227924 - last try before I use pastebin

@K3NK yeah, that's not the reason I removed the interpolation in one go - I added interpolation to scale up the last frame x2 and then inject the reference frame (you'll see x2 load image nodes) into the latent at the start so that wan is reminded what the subject looks like between segments. as you can see from my video the subject remains pretty constant throughout.

@m0no1 @K3NK 

hey guys, i waited a few days so you could double check your workflows 🙂 now i got a few small questions if that’s ok:

the seeds, sometimes fixed, sometimes random. how do you pick? why not just one seed for both samplers or even the whole 20s clip?

did you try the workflow from m0no1? did you get better face consistency, movement or color?

is there a switch i’m missing if i only want to work on extend video1 (like 5-10s, the second part)?

the lighting loras, like the 256 lora from wan2.1 on the noise model only, do you choose by eye/gut or some script?

if you had a local 4090 with 24gb vram what would you change on the wf?

i’m following along and like many others really interested in your results, so feel free to share 🙂 most of these q’s go to whoever reads this, anyone can jump in with their experience

@BrAInB0t hi, have you updated your frame injecting wf or only your username? what image would you load as second/ref image? a close up from the face, the same image again, a new portrait. ty

Is it possible to show me your workflow? I didnt really understand how to use this Kj's node.