Is a 53.9 gb pagefile enough to run both?

42 gb is sufficient for caching both.

only one is loaded at one time. You dont need memory for both simultaneously. Just use the Smooth Workflow.

@DigitalPastel Well it worked. Amazing job.

I got it to work. but added a "purge vram" after the first ksampler, before it ran through. Not entirely sure, whether I had luck with this, or not. Will give it a few more tries.

For me it works with 16GB of VRAM with the WAN 2.2 Workflow v.2.0. But I have also 64 GB of RAM. I don´t know how much the RAM helps out here.

@Pergy25 Was probably the RAM then. I only have 32 GB ram and yeah it's full to the brim.

I'm using a 4070 with 12 GB VRAM and 32 GB RAM. I don't use the GGUF and it works fine in 480x720.

@fatberg_slim In that case the problem was probably just randomly stopping when Chrome ate to much RAM and my purgeVRAM "patch" is nothing more than superstition 😂.

works fine on my 12GB 4070Ti

Like all wan checkpoints it actually natively does 24fps

this just isn't true wtf. why do you think the motion goes slow when you gen at 24fps? jesus.

@crombobular It... doesn't go slow when you do 24fps? If anything it goes too fast (because 24fps means faster playback).

The actual generation doesn't care what framerate you output the video in, it only cares how many frames you ask for in total. When I say it generates in 24fps natively, I mean the motion and physics are in 24 FPS.

HOWEVER, most speedup loras are trained on lower FPS, which biases them towards fast movement at lower FPS. Most of the time they still generate small motions that match 24 FPS though, but often with large motions being in fast-forward. You can match it back up by specifying "slow" for the big motions.

Some of the wan models natively do 16, some do 24. This particular one we're commenting on right now seems to be doing a hybrid, because big motions are fast but small motions are slow. At least, that's what it seems like to me. You can bridge the gap by specifying SLOW large movements, which matches them both up. You can't specify slow small movements, so this is really the only way to do it when they mismatch.

I've been running a lot of gens with this and it's just what's working for me.

I used 32GB (5090) to make mine, 928x640 resolution, 73 frames, 6 steps, took about ~3 mins to generate and was using about 25GB VRAM at any given moment. So using GGUFs on a 4070 will probably be more like 6-10 mins I'd guess. Probably much longer using the non-GGUFs.

@nsfwVariant Thanks for answering. Your estimation is close what I would think a gguf would need to generate. But I'm pretty sure safetensors above 15GB in Size will take forever... SO. Nice to know that you really did have a beefy card :-)

@Adaptalab0r I don't, I spun up a VM online to be able to do it. My actual GPU only has 11GB lol

@Adaptalab0r yep it took 13 minutes to generate a 6 second video on my 4070 ti super. the OP said they will work on the gguf

@coolstrad thanks for the heads up. I dont know which resolution you used but 13 mins sounds not too bad :-). Since you posted your comment, I'll started downloading. Can't wait for the gguf version to be dropped

@Adaptalab0r I really didn’t check the resolution, I just downloaded the workflow of OP of this model, and just used the base settings, I didn’t change anything, I used one of the picture that the OP used in showcase. I mainly did it to check how long it takes. But I am guessing it was probably on 480p cause I do y think so my setup would be able to handle all that in 720p. I can check later and let you know.

@coolstrad thanks

im using 12gb 4070Ti with this model and it works fine.

@coolstrad and @nsfwVariant nice! thanks. In the meantime my poor internet connection managed to finish the download ^^'. AND You are totally right! It works and woooohooooo it is beautiful and fast! 720 x 720 x 81 frames = 6 mins and 27 seconds (including loading the models) I posted a result. Thanks for giving me motivation trying the safetensors !!!

more info. I didnt change anything no lora or stuff, it was on the base config. I used the workflow of the creator (digitalpastel)

Agreed with all your comments, it affects the cumshot aesthetics LoRA quite heavily but it can be made to work well if you lower the strength to very low levels.

Also, have you considered using GIMM-VFI? It's a newer way to interpolate than RIFE and I think it produces better results. All you have to do is install the node for it in ComfyUI. It is a little bit slower, but not by much.

@magicballoon thanks ill check the gimm out, havent really tried other kind of interpolation

Can you give an example what you mean by very different reaction to prompting? I'm having good success with it and have had success with niche concepts the q8 gguf just couldn't even begin to understand (realistic samus x realistic floating metroid blob). I am using my experimental workflow tho and it had good prompt execution on the q8 gguf before. I've found that you have to be more exact in what you prompt such as the woman takes of her tshirt(in the examples he posted) she turns her button up shirt into a normal tshirt but when prompted correctly unbuttons it.

The most I've had problems with is how the nsfw lora baked in handles bjs the woman turns the guy into ken dolls (prompting "the male's intact penis slides in|out of the woman's mouth|vagina" solves it mostly ), or licks the topside on retreat(can't figure this one out).

Also I'm having problems with game screens stuff speech blocks/bubbles, ending screen fades, etc. But the quality of physics this has on both realistic and animation is greater than fp8 scaled and is way smoother

speed loras I believe was baked in don't know which ones though which has killed a lora I used before that can't activate with those.

speed wise q8 guff is 50%(5-6 thrusts) speed, this is(11 thrusts) ,and the f8 scaled is 170% speed (17 thrusts) on my tests. Can use the sd3 model shift to adjust some but can cause weirdness like looping or visible glitches (I settled at 15 I think on mine with )

@AnonBlah ah yeah, well mostly in my limited experience it was that you didn't really have to be so specific, you didn't have to write a book to get what you wanted, and often times it seemed to me like the semantic neighbors of words was a much bigger cluster, like you could use a much greater variety of words to describe something and it would still know what you were talking about even if you mixed terms, like there wasn't as much differentiation between words which in a lot of cases is nice because you don't have to worry so much about exact phrasing. and in one of the videos I posted I did a two video set with one from the standard gguf and one of this model and the motion was completely different, much faster and everything is more 'dialed up' so its important to know that. The only downside I have seen in my experiments so far is that it seems this 'semantic collapse' of terms and concepts also tends to oversimplify. So just about any liquid tries to turn into cum, a dick anywhere near a mouth tries to turn into a blowjob, it just seems to want to pull hard towards certain concepts even if it isn't in the prompt, so its a balancing game and will take experimentation to get to know more about how it all works.

@clzpetn804 LoRAs help, it's just that some of them work better than others. For example, Huge Titfuck works really well because it was seemingly designed to be used with Lightning anyway, and POV Insertion really helps with penis shape if the entire penis was not in the initial frame

I put it in the Diffusion Models folder

There's plenty of other methods, you just need to look. I've been making easy videos with my 4080 super and 32GB ram, using models like this (fp8) ever since wan2.2 surface with no probs.

Is there something special that i miss because for me the first two frame start to animate my picture then it just change everything and i have a complete different style, like i input an anime picture and i end up with a realistic style scene

Solved it by switching to unpc, the best model ever!!!

What workflow and Lora did you use? And how were you able to generate it in 4-5 minutes? My rtx 4070 ti super did it in 13min 12 seconds. I did not add any Lora or stuff. If you are okay would you mind sharing your workflow?

@coolstrad The workflow I'm using was a standard 3 ksampler 6step that I have modded to 4 kasmplers(I'm back down to 3 and 3 minute 720 vids) and modified the model pipeline if you want to try it I'll upload a copy after I've cleaned it up a bit.

If sounds like your offloading to page file with your diffuse times if you are there is no workflow that will help. You can check by opening task manager and watch if your harddrive has a bunch of activity while you diffuse. You can try closing background task to free vram space

@AnonBlah Can you share the workflow? Yes it is using page file memory.

@coolstrad  If i don't share the workflow today I won't be able to get to it for a few days super busy.

I am going to restate that if you are already offloading to the pagefile my workflow will not help you tbh. You need to lessen background tasks

Also got an RTX 4080 but cannot run with this setting without running out of memory

@supertype3 At first, I had Comfyui

via Stability Matrix—that didn't work properly either.

Then I reinstalled Comfyui locally, and everything worked.

Give it a try.

Hi where can i change for 960x1440px resolution ?

Yea, especially since the lightning loras could go through future iterations and perhaps different variants. Merging whichever makes this checkpoint a lot less flexible. Hoping OP would reconsider.