Not the biggest expert but I did some research on this myself and generally the halfway point should be the best I believe, so 4 steps with cutoff at 2 or 6 with cutoff at 3.

I found the same issue, but I’m not sure what’s causing it.

This problem happens frequently. The breasts move eerily.

I get this a lot, if there's a breast, even slightly exposed it will bounce like it has a mind of its own, really killed a lot of otherwise good prompts. 😕
I think it's some NSFW LORAs merge, so not sure how to cancel it.

That's most likely because you are running the --use-sage-attention flag at launch. Happens to me when using core nodes with certain models, particularly some safetensors, while the wrapper doesn't do the same. So run the Patch Sage Attention KJ within comfy instead.

Yes

Could you please provide a link to the workflow you're using? I don't know where it is.

https://civitai.com/models/1847730?modelVersionId=2264611

@DBrepair Thanks for recommending and linking to the workflow. As you said, it works perfectly out of the gate. Just had one custom node to install (mxtoolkit) and that was it.

Try GGUF https://civit.ai/models/2083588
https://huggingface.co/Bedovyy/smoothMixWan22-I2V-GGUF

Already available

@qek 
Thanks! but really?
I've tried the official I2V workflow of this model with reference image by using this mode(v2.0), but result was T2V's one...

Any tips?

@dtablesiffusion You need to select the right model. There are 4 different models. Look at the top left corner of the webpage. t2v high and low v2 and i2v high and low.

@Dimyo 
Thank you for the kind advice!
But, I think existing I2V model is v1.0, right?
So, the I2V 2.0 is not deployed yet.

Of course I tried I2V 1.0v and it was good!🥹

Is it possible that you run the light loras on top of these checkpoints? They are not required here.

@Amatiramisu No, I don't have any lightning loras loaded. In fact, this issue occurs even with no loras loaded at all

wan 2.2 is divided into two because it would be too big for GPU's. High and Low. High is the first steps and does the rough motion and concepts... Low is improving over that. You can also just use Low, but you won't have all those great motions that come with 2.2

@agustincaniglia575 thanks for the reply, i was also wondering. so to use high, its mandatory to also use low? but using only low is okay? (i have rtx 3090, 24 gb vram)

@dpons222 If you decode the latent output from the High model, you'll notice it's very grainy and blurry, but you can still see the shapes and outlines of the characters performing the motion you want. The Low model sharpens and refines the image, so the output from Low is always clear and crisp, but it has little to no effect on motion and using it alone results in fairly static videos - characters walking in place without moving, etc.

@dpons222 basically what Google ai told me. High gives smooth animation, but low is for details.

Try turning off LORAs one by one to identify the culprit and reduce their strength dramatically afterwards.

With only the model I get really good results. With more LORAs or when turning up the strength of the LORAs the whole character starts to bounce / jiggle

I ran into the same issue,

I found that running this checkpoint with the smoothmix animation lora at 0.5 high/low minimizes breast jiggle to a tolerable level.

I have tried probably 40 seeds without any lora whatsoever, that is what I described above.
I have also tried including the lora between 0.1-1.0, but the jiggle stays the same or gets worse the stronger the lora is.
The best results I have gotten so far, but imo still WAY too much jiggle is with the "old" nsfw lora at around 0.25.
So the problem lies within the model it seems.

I was wondering if someone had found a negative prompt or smthn that could mitigate the jiggle...

were you able to fix it? i kinda have the same problem, it was working until yesterday. now it just stops when loading low noise model.

@razor3208 the cache-non flag is supposed to work with nodes that clear memory. Ideally you would arrange the nodes to ensure the 2nd model is not loaded until the first model completes and then is fully unloaded. cache-non can't help much unless you set up the ordering.

@charge266 i dont use cache -non, or anything, i just use the default, it was working fine. then it stoppped

@razor3208 That sounds like a different problem than mine- but yes, I was able to "fix" it- One, I realized I did indeed change a variable: I had tried using a WAN 2.1 lora and that was what was causing the issue, I thought they could be used on WAN 2.2 but apparently it can cause the issue I had and even "poison" the model so even after its off there were still issues.

Two, and what really made this workflow and these checkpoints shine- I finally switched from Windows to Linux. Should have done that years ago apparently. Everything works way, way better now.

Sorry for the slow reply, I hate Civit's notification system lol

Use i2v models

I don't know your sampler and scheduler

@qek Euler a/Normal as recommended

@edsgadhgsrerthsdffgdhb In my workflow: 5 steps, euler a, beta, shift=5, cfg=1.

Do not use step+cfg distill loras! Smooth Mix includes it
Other things: Tiled VAE is slower and lossy. There are 2 models and it takes some time to load and unload if you use both. You don't have to use the default length value.

@qek trying a run with your settings soon. I havnt found out if umt5-xxl-encoder-Q6_K.gguf is tiled or not... it might be time to order a 5090

@qek getting great results with your settings now! is it the shift that fixed it?

@edsgadhgsrerthsdffgdhb I said "Tiled VAE is slower and lossy", not umt5

I see

For my GGUF workflow using the quantized V2 T2V checkpoints I had to change the lownoise to the standard quantized Q8 model for Wan 2.2 A14B T2V while leaving the hghnoise smoothmix checkpoint. I don't have a checkpoint loader with weights so I went this route, still has better range of more natural motion in scenes without the ridiculous adherence to model specific proportions, I use the older loras to add back the smoothmix look at lower weights using the lownoise loras in the WF, waiting for the V2 loras to leave early access.

flan_t5_xxl_TE-only_FP8 =+~5 GB

@qek why does it need a text model integrated? Is it for something special that the standard model doesn't do?

can you plz upload model without text encoder like others. thanks

@charge266 and it exist on both low and high model, kind of insane

It's better to ask about RAM too, it depends on multiple factors

Your best experience would be with atleast 16GB VRAM and 64GB RAM
You can run on lower specs, but your result will get worse because you then need to cut in quality

I tried with 32GB of system RAM and it took me over 15 minutes to render 6 seconds, after my RAM upgrade only 5 minutes... so it does matter a lot.

@Setian91 What's was specific GPU when it took 15mins? Seems this model is poorly optimized. Is 24GB VRAM not enough to run this model fully on GPU?

@DeviantApeArt Don't listen to him, he's wrong

@DeviantApeArt I used a 4080 super which has 16GB of VRAM and I used the scaled clip version

@Setian91 I have less RAM, my GPU is worse, but I've been running WAN 2.2 without getting OOM

@qek we might have different settings or a different workflow, I used Smooth Mix's workflow without sage or triton since my ComfyUI refuses to install it with al the compatibility issues.

I didn't claim I got out of memory though, it just used the hard drive.

I use DaSiWa's workflow now with some small edits like adding clip vision for better results, for me that workflow works a little better.

I also use Windows 10, your experience might vary, but this was my experience and the slow generation was instantly fixed after adding more memory

@Setian91 fr? I also have 16gbvram but its a 4070tisu but only with 32gb of ram. i think it took like 10-20 minutes for the hunyuan base model so it sounds like im in the same boat as you? I'm always resorting to GGUF's atm and then hear about ppl being able to use the raw models with 16gbvram somehow, maybe i do need the 64gb of ram

Edit: RAM prices are atrocious rn nvm lol

@bhopping yeah, there are ways to run it fast on lower RAM... and I tried but the results were so bad compared to what I create now... and yeah RAM prices has gone up a LOT sadly

@Setian91 thats crazy, how much ram made the difference? i have 2x16 ddr5 ram (32gb) and have 2 slots left. i could try to upgrade to 64, is that how much u have? sounds like its worth the upgrade

Replace comfy essentials with this version: https://github.com/cubiq/ComfyUI_essentials/tree/9d9f4bedfc9f0321c19faf71855e228c93bd0dc9 that worked for me. hopefully it does for you

@fairypants Thanks. I'll give that a try. So I remove the old folder in custom_nodes and clone that one?

@jonk999 Yep. I just overwrote the files in the original comfyui_essentials folder. Just make a backup in case anything goes wrong - but it should work fine.

@fairypants Thanks heaps. Will give it a go once I have some time.

same issue. i tried to use use both positive and negative prompt to suppress the body movement, but did not work

A bit, depends on what it was trained on, might have been overtrained on dance or nsfw. I'm trying to figure this out myself, what negative prompts have you used?

Also remember if GFC is set to 1.0 it ignores the negative prompt so I usually try 1.5 - 2.0, above 2 tends to lead to overexposure

Always even if you specify something like "static" "natural" "at ease" "calm" or "rest" haha

It leaves you with goofy results sometimes

Sounds like the same issue I've been seeing with the Remix checkpoint, too. Must be a Lora causing it.

I've found giving them something to do, walking, standing up, sitting down, etc.. Sorta suppresses the urge but the moment they stop they go into their idle animation of the boogie hips.

I've had this issue as well. Good callout.

You probably trying to generate videos longer then 81 frames?

I echo Lora_Addict. In my experience this never happens when generating videos up to 81 frames (5 seconds at 16 FPS) in length. Sometimes I can push it to 96 frames (6 seconds), but beyond that the video almost always returns to the start position by the end of the video

@wildkrauss It seems it's the reason why some Hunyuan Video users say that Wan 2 can't generate long videos. There are some LoRAs making it possible to generate better longer videos

@Lora_Addict yep i was

@evilenerjohn190 yeah, Wan 2.2 can't do more then 5 seconds / 81 frames. Or lets say, you CAN but it will do what you witnessed :)

--lowvram ?

I use the Q3_K GGUF models from BigDannyPt on a 3060ti 8GB VRAM + 32GB RAM

Then just follow the recommended settings for Mid End. Use the Smooth Workflows V2.

It takes around 5 minutes for a 5 seconds video at 1 CFG.

Y para una 7600xt (16gb) ?, sigo viendo HIP out of memory...

Bedovyy's gguf of i2v, i use q6 on my 4060 laptop gpu 8gb vram + 16 ram and its work pretty nice
without sage and any loras its like 500 secs for 81 frames 480x848

And remove the text encoder not to reload the same one, @DigitalPastel

I'm also going to add that this model has an over infatuation with making that stupid Pokimane face where she sticks out her tongue.

没有，内置了，别再加了。

Can be both, thank umt5 and the Wan team for training on Chinese captions

i maybe wrong but use a different text encoder.

Just get the links and download without the VPN

@qek the links are still on the civit domain so I can't.

You can use i2v

@qek you can also use the damn base model and learn to combine loras yourself... -.-" this merged bs models looks like wan2.1

@K3NK or complain about women with dicks and cum leaking from everywhere because they have been merging random loras and say that it works as intended

smoothMixWan22I2VT2V_t2vHighV20.safetensors
SPEED_HIGH.safetensors (outdated Wan 2.1 lightning?) strength=3.0
SmoothMixAnimation_High.safetensors, strength=0.5
SmoothMixStyle_High.safetensors, strength=0.2
SmoothMixStyle_MoreOptions.safetensors, strength=0.2
-----
smoothMixWan22I2VT2V_t2vLowV20.safetensors
Similar, but loras for the Low model

Thank you qek. :) Do you think the same loras are used for the v1 t2v version. Imo it's better than v2. I found the first two but can't seem to find 'SmoothMixStyle_MoreOptions.safetensors' anywhere. :/ I guess it should be found here?: https://civitai.com/user/DigitalPastel/models

I like how faces look with this model.. Less ai look than the vanilla model. I'm trying to tone down the massive loads of artificial looking cum and strange and loose looking vagina and anus. Subtracting the first two loras doesn't seem to do a lot but I'll keep on testing.

is there a way to remove a lora from the model? i would like to remove the lightning lora

https://civit.ai/models/1419535

@qek I've tried using this LoRA, but in I2V, there are issues where the face becomes blurry or turns into a completely different person.

Yeah, unfortunately it seems like the model is incapable of making small breasts. Even using loras, like the one mentioned, doesn't help.

its already here. google smoothwanMix gguf in the google

Ok, I changed the start from Unet Loader to Load Diffusion Model and increased the size so it doesn't need to be scaled at the end. Looking much better.

However, there is only a single SmoothMixWan and it's high. I used that for both but where is the Low for this checkpoint?

TIA

@david469 there's a high and low download tho?

@civitaisucks777 I'm an idiot - there is. I didn't think to look at the download choices (still learning CivitAI).

Nah, just a weak merge
https://civit.ai/?dialog=commentThread&commentId=1014429

Faces are pretty good on version 1 of this model imo but for best realism try t2i (960x1440 or higher) with the base model and this lora and the workflow embedded in the image: https://civitai.com/images/98215009

Res_2s/beta57 (or bong tangent) are pretty good for t2i. You'll find these samplers/schedulers in the RES4LYF repo.

Then do i2v with that. I tried t2i with the smoothmix merge but the images just came out distorted or strange looking.

Something like the new nano banana pro is properly best at ai realism but I suspect they don't allow nsfw content.

For generating i2v I recommend the enhanced merge. It doesn't go crazy with loras and improves motion I think. https://civitai.com/models/2053259/wan-22-enhanced-nsfw-or-camera-prompt-adherence-lightning-edition-i2v-and-t2v-fp8-gguf?modelVersionId=2409571

If you're using Lightx2v speed up loras , turn those off. Those are already baked into the model

@Melty1989 Yes i know, the point is if characters movement are dynamic and fast, then automatic interpolation does not correctly record strongly shaking parts, e.g. breasts

try using CFG 1.5/2.5 in HIGH noise and use negative prompts to avoid those extreme moves

@simonishere for me, negative prompts changes absolutly nothing

@yukiookami_8593027523 negative prompts only works with high CFG, that's why i said, ''try CFG 1.5/2.5'' (even 3.5). CFG 1 ignores negative prompt

@simonishere raising cfg will increase the time it takes, @yukiookami_8593027523 try a NAG node for negatives before raising cfg from 1

(or try a diff checkpoint, another one recently came off early access and it's smooth as butter 90% of the time)

@civitaisucks777 can you tell what it is name

@civitaisucks777 of course it will, if you want better results sometimes you SHOULD wait longer times, that's a direct correlation

I type prompt to take out dic from girl, and thats it, but generated video showing a sex animation, I can type sex, doggystyle, sex from behind... in negative prompt and increasing CFG all day long but effect is the same

The second problem is that the animation is often jumpy, for example in cowgirl position, woman rises and stops halfway for a split second, then resumes, and so on

@yukiookami_8593027523 @yukiookami_8593027523 what sampler/scheduler are you using? shift? are you using loras?

@simonishere just default and close to default settings in Smooth Workflow Wan 2.2 (img2vid/txt2vid/first2last frame) workflow

@simonishere are you intentionally giving bad advice? more time for better quality makes sense, but that's not the case here. you can only raise cfg to about ~2 before you REDUCE quality. and at ~2 cfg, the negatives aren't very strong anyway.
so you're adding 1.5 to 2x more time for barely any difference. when you could just enable NAG with zero added time. NAG also gives stronger negative guidance than ~2cfg and you can change the NAG scale to make it even stronger if you need to.

you shouldn't ever need more than 1cfg if you're using NAG and prompting properly. it's better to use that time saved for more steps or higher res

@civitaisks777 you sound like an ad bot

You could just download the base models and those SmoothMix loras instead

Any chance you have a workflow that works for the initial image. I got e.g. the 'WAN 2.2 Smooth Workflow v2.0 (img2vid)' workflow to work but in the end the image doesn't appear in the video at all.

Any chance you have a workflow that works for the initial image. I got e.g. the 'WAN 2.2 Smooth Workflow v2.0 (img2vid)' workflow to work but in the end the image doesn't appear in the video at all.

"Do I need the high and low models at the same time" We start with High and refine with Low. Maybe you should try only using one, some have been only using Low, also, Wan 2.1 and Wan 2.2 5b don't need a refiner model
"if yes" If High doesn't get unloaded, then yes, both will remain loaded. If it doesn't unload for some reason, then you have to separate
"40gb of vram?" You can use GGUF to save more memory, lower quants of the transformers and umt5

@qek ty for reply
just to be clear can force the unload of one model before the second one gets loaded?
since when the second model starts loading the comfy just stops working

@yar4ik141 I think you can play with some arguments related to memory management and lower quants of the model(s). It appears that the backend doesn't want to unload already loaded models unless they no longer present in a workflow (deleted loader nodes, not just bypassed), or you press Unload Models and Execution Cache to unload everything completely. You can try memory cleaner nodes, but it didn't really help me in the case. You can save a latent made by High, then unload High and load the latent for Low to finish, you would have to reload them anyway

@qek did you found a solution for it or just decided to you a lower model?

@yar4ik141 It's already stated

我也是3090显卡，我在同时使用高低频型号，实际上它们是运行完一个之后再运行另一个，所以24GB显存可以完全正常运行。40GB显存是指你的3090显卡有24GB显存和16GB共享内存。你不需要去管它，因为你的显卡不会去使用那16GB共享内存，共享内存的速度比显存慢太多了，用它反而会大幅拖慢速度。

More frames/higher resolution, more VRAM, that's it. How many frames and resolution did you render?

on a working workflow. 1 model load at a time so about 21gb (for the models and the text encoder) of vram need total, and when it run the vae it should be ok. using both high and low is recomended but dont worry as only 1 work at a time if set or workflow intend it to. resolution wise dont go higher then 480x720 or vise versa and if runing 16 fps keep at 5 sec of video length increasing these will require more vram. use their official workflow for that model to test and tweek . once you learn more and want to branch out then you can start experimenting.

@haxo A total of 113 frames, with a resolution of 1536 × 1024. Higher resolution may cause graphics memory to burst or run at low speeds using shared memory.

Only use the high model...

Just use i2v workflow.

Without lightx2v 4 step

are you using any loras? and prompt it for smaller bust? to help we need more information. so we can try to see what the culprit is.

bigger the better :3

@charlenebelmont Just the lightning step Lora. Here's the prompt that produced quite the knockers:
SmoothMixAnime. A young woman wearing a skin-tight gymnast outfit spreading her legs seductively. She is seducing the viewer. She has small breasts and a flat chest. She has wide hip bones and a toned stomach. she is impossibly beautiful and cute and innocent. she is petite. Blonde hair and blue eyes. Refined and mature expression.

@Seikkailijapoika from a lot of my testing with this model it get confuse with conflicting words sometime. my suggestion is try using description that wont overlap or contradict each other or to much detail in one area. for Example: try using "her breast are small and flat" . when you try to give it to much detail this model for me at least get confuse and break sometime. what i find also work is . describe the women feature first before any clothing detail. this way you can fine tune fore the look you want and when you start adding loth it wont mess with the prompt for her feature and description

so my promt would be like:
(1)her look and feature (physical appearance)
(2)clothing

(3) then the prompt for action and or camera and setting or what you want the video to be about.

this way you can say change her look without messing up the other stuff or change her action or cloth and it wont interfere with what you want to do. and it alot more organize :)

Nope

Add the Load Lora node yourself

@qek if I add the load lora node, where should the node connect to and where should the trigger word be typed? Should it be typed in the Positive prompt node in the workflow?

@leepeter1231 Their workflow is a mess. Should at least connect to "model". The trigger word(s) may help

@leepeter1231 yes, add triggers to positive prompt

They can just be added to your workflow. No lightning loras are needed.

It isn't VACE

https://huggingface.co/Bedovyy/smoothMixWan22-I2V-GGUF/tree/main

@qek thats what i meant, i have already found them and im trying to use Q6 and Q8 no errors in the generation but the videos come out like patchy, exaggerated movements and bad overall. i tried changing the lighx loras, the wan 2.2 that comes with the wokflows, the rank 32 mentiones on the description, the rank 16, 64, disable all loras, including the ones to speed up generation. changed the clip for a cuantized one, changed to the wan2.2 one, changed again foor a xxlbf16, xxlfp8, the one in the confyui wan2.2 github page, im trying so many options and every generation gets worse than the last one. i thought that if the models where posted on civitai some people would generate and i could see how they used the quantized models and see what i am doing wrong

@Fgonax It's a simple merge with porn loras, the users should have downloaded the base models and DigitalPastel's loras instead. I do not recommend this SmoothMix merge, it isn't the first comment about output videos with strange and/or unwanted movements

It's a simple merge with porn loras, do not use it

This happens to me too, and it drives me nuts. even when I am trying to get it to do other NSFW things where the woman's hips happen to be in view, she wont stop gyrating or bouncing

Of course

@qek i thought that too but its randomly now taking up 100percent of my ram for a single render, using the smooth workflow and not changing a thing. Any tips that could help?

@Aisamplewalrus I don't know the workflow. Make sure you run the models in fp8 and use a lightweight VAE from lightx2v

@qek Oh I’m using the smooth workflow that this model recommends. Sorry noob question but how do I run the models in fp8? I’ve just been using the smoothmix high and low from here and their smooth workflow.

it runs on a 3060 12gb just fine

I2V. Mostly no. At high resolution 1024 and a 5-second clip, I use up 60+ gigabytes of memory during upscaling.

@DBrepair Wow, I need less

Yes

Yes and if you add a clip vision to your workflow you don't even have to explain the image being realistic.

@Setian91 Thanks a lot, where can I find a simple workflow for i2v with Wan 2.2 smooth ... ?

I kind of like DaSiWa's workflow since it doesn't require Triton installs:
https://civitai.com/models/1823089/dasiwa-wan22-workflows

You need to attach clip vision yourself though, I don't think the newest version has that.

I haven't tried the newest yet though, but its fairly simple to use, Smooth's workflow is also good, but I keep getting errors after the 4th video < maybe because I disabled the triton nodes... but w/e.

Futa transformation

Okay I've finally managed some yank sollution using the pull out cumshot lora (0.5 weight) along with prompting "Woman vagina is behind man penis"

It probably has to do with all the futa training which I'm not against but if its the only result even for females having straight sex this can be frustrating...

I have the same problem, especially when the person is at a distance. I understand that a face detailer is needed, but I don't quite understand how to integrate it into the author's workflow

No, it's just the pretrain fron Wan AI + their 3 porn loras, nothing fantastic at all

可以的

I run the GGUF versions on a 3060 12GB. Think I had some issues with these ones...

I'm so glad i recently bought 5070ti version with ddr7 and 16gb :D runs so fine

The icon: no b1tches?

hm https://civitai.com/models/1995784?modelVersionId=2324440

Do not use. It's just base models with loras from https://civitai.com/models/2040641

https://huggingface.co/Bedovyy/smoothMixWan22-I2V-GGUF/tree/main

@qek thanks mate

one says i2v the other says t2v right after the "_"

They will continue posting such comments

What's your sampler and scheduler?

@qek Euler and Simple !

@mnlapn My combo: Euler A + Beta

It completely changes the face on every setting, Useless for i2v

Same here. For me t2v give "Pixar / Disney style"

My workflow run on a 3070 8gb with the GGUF version - Q4_K_M
https://civitai.com/models/2083588?modelVersionId=2357653

I'm not sure but in my understanding, lora makes the generation faster by reducing the steps it needs.

As this model is already able to achieve good quality in 4 steps, it seems to me not necessary to add extra lightning lora to it.

https://huggingface.co/Bedovyy/smoothMixWan22-I2V-GGUF/tree/main

I have the same issue, its ignoring the image and just making a picture video using the prompt

@ThatSoKitten 
I figured it out, it is really the models. I have downloaded the "I2V High" and "I2V Low" version and now it is working.

If your using comfyui, try setting to --cache-none, and --reserve-vram 0.9

@ThatSoKitten i tried that, still didn't work. it was working fine weeks ago but for some reason comfy crashes when loading low noise model. i have 16gigs of Vram. all other models load just fine. something is seriously wrong with this low noise model. still unable to run it. i hope they fix this issue on the V2 version and release it soon!

@razor3208 use --normalvram or --lowvram

@qek yeah running low vram offlloading onto my system ram

@images101 does it work when set to low vram mode? cause mine is set to normal

@razor3208 for me smooth mix does not work thats why i am here asking for help

I'm also having this problem.

It's a simple merge with 3 porn loras, I do not recommend

The ultimate helpful thing would be for some kind nerd out there to provide images from the checkpoint as compared to it's counterparts, and on each variant, and so on. Out of curiosity, what is the benefit of merging the loras like this, does it work better than summoning lora like normal? If your baking in loras, and new versions of loras release, or better loras, seems like bad idea.

Well, it's my fault, I used the wrong model and used I2V instead of T2V