🗡️💀 DaSiWa-WAN 2.2 I2V 14B Lightspeed | FP8 Safetensors💀🗡️

Sure model page or collection.

@darksidewalker I don't understand you. Where exactly is the usage example for the model? It doesn't seem like the example from the Wan2.2 official page is applicable here. As I understand it, here we need to connect the loraloadermodel only after loading the base model, or am I wrong?

@guljaca on my model page are the workflow as suggested resources or you pick it up on my collection. They are ready to use.

@darksidewalker Thanks, found it. Almost by accident in the depths of the local interface.

Wow! It works about 100 times faster than the original.

if i try to generate with it it gives me an error but the older ones works fine

@oussamajack0950599 Ask seaart...

@darksidewalker oh sorry, i thought you were in charge of it lol my bad

@oussamajack0950599 You know the most accounts from "publishers" there are just bots who crawl civitai and reup their stuff?

@darksidewalker holy, so that's not your account there? man do something about it, the account has your name and everything and making money on your behalf

@oussamajack0950599 I'm on that ...

@darksidewalker love your model man don't stop making more and here is a link to the account https://www.seaart.ai/user/01c9251620fb7d2a579199e24b59b6d3

Thx 😸

New model works great! Well worth the buzz!

@Gladas 😊

Edit: looks like it's the text encoder/CLIP (possibly) in my case. I just dropped one of the sample videos into my comfy and saw what they used. Edit 2: nope still crazy shit happening lol. Guess I'll stick to using Radiant.

Original message: This has been my experience as well with the new model. Radiant works great, but this new one is doing some crazy stuff.

The upload got corrupted.. No worries, my checkpoint is working. I'll re-up as fast as possible and post here.

The re-up is working now, did dl it and tested. No morphing. The last upload was corrupted.

@Gladas I posted a comment about getting the same error. For me, LureNoir High/Low would not run unless there was something present in the Negative prompt box (it was the only one that was empty; I already had a prompt in the Positive prompt box). You may want to try that.

I used the workflow out of the box, only changing the checkpoints, FYI.

@darksidewalker Yep that was the issue, thanks!

Personally I've literally never generated good results using Euler. I don't know if it's workflow related, some kind of configuration issue deeper than ComfyUI, or my hardware, but people making good results with Euler just baffles me.

Use what ever works for you. That does not really matter, just details. If it works it works, if it is doing artefacts its wrong. Euler is just mostly used by the trainers/creators of the basic tech and therefore most compatible.

I'll look into, maybe the upload got corrupted, no worries, the checkpoint is working as intended. If the upload is faulty I'll just re-up it. If it is the upload, just re-download it than, I'll post back here.

I can confirm after DL, its broken. I'll have to re-up this.

The re-up is working now, did dl it and tested. No morphing. The last upload was corrupted.

Awesome! Thanks for checking!

For me its working without negative and without increasing steps. Maybe you dl the corrupted version, I re-up them and tested. All should working as intended.

"ERROR: Clip input is invalid: none" is if you not use a own CLIP, it is not included any more in the "LureNior".

@darksidewalker Thanks! I only increased the steps to experiment and see what I'd get. It was working perfectly at the default 4. And I am fine with having something in the negative prompt box to get the workflow to run. The LureNoir Low version took almost an hour or more to download using Civicomfy, and I'd rather not have to go through that again unless I have to.

@PMAX365 That is also be a problem of the corruption. I even got on kb speeds... lol... You could dl in parallel, and just delete afterwards. Just to be sure :)

@darksidewalker I am very new to all of this 😆 Like I've only really dug into ComfyUI in the last few days after having dabbled with it on and off earlier this year. So, unfortunately, I don't know what all of that means, hahaha. But it's working as advertised thus far! 👌

@PMAX365 I mean the corrupted file cant be downloaded with original speeds from civitai. Should be going faster with a OK file. You could just dl them again, while using the others and than switch/overwrite them at the end if your dl speed is low.

@darksidewalker Got it. I may try that. I think there was just something going on with Civicomfy downloader at the time. The LureNoir High model download in seconds. The Low model was just taking forever. I have 1 gig fiber internet.

Not possible any time soon.

@darksidewalker, alright. There is no need to hurry that.

Hi! All is documented in "About this version" and some in my announcement 👍

My method is to use Qwen Image Edit 2509 (or Nano Banana) to create the storyboards/shots, and then use FLF2V to stitch the shots together and fix the consistency.

This feature is not possible with WAN 2.2 i2v, WAN AI did not implemented and trained this for the base.

The FLF approach of @PotatoChipO3O is a good idea to mitigate the problem, but with more chaining it will degrade because of motion artefacts.

WAN 2.2 does not understand anime very well, it will try to go into "normalised" eye forms if you have a spacial style. You may try to describe the eyes in more detail, that helps maintaining the eye style. Also describing the video as anime or cgi helps a lot.

Nice to know! Have fun, create some awesome art!

Oh really? Which workflow do you use? I keep getting OOM issues on my RTX 4080 (12GB VRAM) and 32GB RAM...

I tried a few workflows (fastfidelityComfyV1.2, fastfidelityComfyflf1.0 and fastfidelityComfyflf1.1), but no luck. I can generate a terrible video when loading only the "High" ckpt, but whenever I attempt to use both as intended, I get an OOM (even with 49 frames, 416:544, 4 steps). I'm on a fresh SwarmUI install.

@darksidewalker Sorry to bother you! Do you have an idea what my issue would be here? OOM always triggers right after second step ends. No other intensive active process, except browser running SwarmUI.

@NoctisLucis Maybe your resolution is too high. Or something is off with your install and the model gets not offloaded on the switch.
You could try to make your virtual memory bigger (zRAM, swap/pagefile).

You could also try to add --lowvram to comfyui.

Didn't work. OOM. All I did was use provided workflow with default settings.

@DeviantApeArt your virtual memory is too low.

I had an issue where it stopped working and was freezing my PC. What helped me was updating drivers. Switching from studio drivers to game ready drivers, clean install. Video gens are back all of sudden.

@darksidewalker I also have 8GB VRAM and 32 RAM though so why would it work for OP but not for me?

@DeviantApeArt As I said, you have not set enough virtual memory (e.g. pagefile, swap, zram). 32GB RAM is not enough to fit all in, but will work with virtual memory.

@darksidewalker I updated drivers and set --lowvram and now it's working. One weird thing, it keep crashing at the very end when I set --reserve-vram too low (anything less than (0.5). Also, I found it weird when using your workflow it only gave me the last frame of the video as preview, at first I thought it only generated 1 frame but I found the video in the output folder later. Anyways, thanks you for this, it's impressive WAN2.2-14b can run on a 4060 8GB VRAM laptop GPU. One generation at 512x512 only took ~3.5-4 minutes!! Also, please make a Discord community, it would be nice to discuss this model with other people.

Yeah. Kind of. Some dark-arts and such things included😆

@darksidewalker i waiting to get this :)

@darksidewalker ohoo i excited to use this

Does there is chances for gguf version for low vram user like me Ha ha hah

@speedofwar12993 Sorry not possible since there are no conversion path from fp8 to gguf

Did you install the comfyui addons from github? Maybe missing them?

@darksidewalker Hi, yeah installed the add on, but think I just worked it out. I downloaded the CLIP umt5-xxl-enc-fp8_e4m3fn instead of umt5_xxl_fp8_e4m3fn_scaled. I am not actually using an scaled CLIP. Thanks for replying, sorry it looks like I was just being an idiot. If it still doesn't work after I've downloaded umt5_xxl_fp8_e4m3fn_scaled I'll let you know. Thanks

@CanvasNSketch NP :)

@darksidewalker Update, I've added the right CLIP and it works like a charm. Thanks for the checkpoint and workflow.

But the clip is already baked in? use the clip that the model comes with

@DeHeinrich On LureNoir clip is not baked in.

I checked and it seems to Resulting in lightx lora, Have you ever encountered a similar problem? I saw that the video you showed is colorful

@1741673381619 You are not supposed to add an extra speed-up lora. There is no color shift problem with the checkpoint, this can only be your settings.

@darksidewalker I have found the cause of the problem. The video combine node loses color when saving GIFs, but when I changed it to another type, the color returned to normal. :( : (

@1741673381619 Gif is the worst of all format's. Nobody should use it😭

@darksidewalker Anyway, thank you for your workflow. It allows me to compare nodes one by one and identify the problems that have been bothering me for a long time. GIF is convenient for uploading to Pixiv, in a sense

I don't have this problem or encountered it. I do not think it is a problem with the checkpoint. Maybe your memory got full and did not clean-up, so you would use virtual memory -> slowdown

i do restart pc and generate again this help me always u should try tell me if its help

Yeah you likely ran out of VRAM.

If you're using ComfyUI, ensure you're not using the "High VRAM" launch parameter, unless you have an RTX 5090.

Also look up the "Unload Models" extension. It adds two nodes that allow you to unload models mid workflow. One node unloads a specified model and the other node unloads ALL loaded models (don't use this one).

I have an RTX 3090 but even with 24GB VRAM I'll run out of VRAM at the VAE decode stage if I'm generating at a high resolution. I use an unload node to unload the WAN 2.2 Low noise model after it has finished, which prevents the VRAM running out.

@MaGiC48x oh thats why, thx ill use the node extention

@Seii1 just add clean ram and clean vram after genration complete just attacched to vae decode

No, I do not have encountered any abnormal behavior and I did many samples and posts. Could you specify what your problem is?

@darksidewalker 
Example
https://civitai.com/images/107951063

@guljaca I always get 404 on the link. Can not open it.

@darksidewalker Strange, I can see it in the list of works for this model. And other people are leaving ratings. I would have sent a link to the image from a third-party resource, but I'm not sure the local rules approve of that. Try using this prompt: "A thick white sperm flows out and splashes from her crotch". In this case, it's as if the girl is being penetrated through by water, and it pours out of her mouth in a stream. Or take a closer look at other people's work. At the very least, the girls have saliva dripping from their mouths, even if you don't specify anything.

The problem is with RadiantCrushHigh
Image is in Lurenoir High

I can second this; there is a similar issue with other 2.2 loras such as the anime finisher and creampie ones where you have to set the lora strength lower to stop the cum vomit. I wonder if this was unknowingly cooked into the checkpoint. It is otherwise a fantastic model.

@guljaca I found it, manually searching.

No offence, but your prompt is very unnatural and lacks any details of description of the scene or persons and where things happen it can understand. WAN 2.2 does not inherently understand "anime" completely and therefore it can not estimate what it is seeing.

So you want to change/refine your prompting.

The used image is full of artifacts and distortions, it just pours out liquid everywhere it estimates a opening because it does not understand what you mean at what position with

1) "inside her" -> can not seen from the point of WAN
2) Where "crotch" is -> not visible
3) It "flows out" -> where and when
4) Combined with "splashes" and "a lot" it will just guess where could all this happen and accumulates "a lot".
5) "Flowing" as wording normally is from top down, like a waterfall.

So basically you got what you ask for 😁, lacking providing details and clearer composition.

@Pauper_Grave I did wrote this, to start with 0.3 and do more if needed and not overdo with LoRAs. The reward attention and distillation exponentially taking a lot of thing more serious 🙂 But I do not think he used any LoRA here. I think its the prompting.

@darksidewalker After playing around with it, I think it might be tied to the High CFG; I ran a few tests at 1.0 and it was a pretty frequent occurrence. The effect stopped completely once raised it to 2.0 or higher, so I think you might be right with issue being how literally the model is taking the prompt.

Thank again for such an amazing model

@Pauper_Grave It would be absolute amazing if I could manage to get it understand every concept. But I fear "cum inside" may not be a thing without a really good prompt and additional lora.

I did a short test with a fitting lora assertive-cowgirl (0.4) for an image and it turned out really well, despite the drooling, that I did not prompt for. I removed the example, because it was a fast sample and not worth the top gallery. I'll have a look at this. Maybe I can tune it or can find a prompt for this particular situation. I this was not in the scope of my testing lab before.

@darksidewalker In your example, she also has 'drool' coming out of her mouth. But it's not exactly drool. If you add 'a lot', she will 'vomit'. So there is indeed a problem. I thought that WAN didn't work well with long descriptions. It might start ignoring instructions as early as the second sentence. Can you recommend an article that describes how to create prompts for WAN?

@guljaca True this particular situation seems overtuned. I'll work on this.
Unfortunately I do not have a real prompting guide or article to recommend for that atm.

Sorry for the late reply. My comfyui was broken. I could not generate videos.
Here is a working example. Less prompting and a fitting lora seems the way to go for "creampie"
https://civitai.com/images/108078581

@darksidewalker Indeed, it's a 404 error. Perhaps the site provides different links to the user and the creator.

@darksidewalker I managed to find it manually with the same link. The site might be blocking direct links and only showing pages if you navigate to them logically from within the site itself. Thanks.

Anyone running this on AMD? I can not provide much help.

Try the distorch nodes. pollockjj/ComfyUI-MultiGPU: This custom_node for ComfyUI adds one-click "Virtual VRAM" for any UNet and CLIP loader as well MultiGPU integration in WanVideoWrapper, managing the offload/Block Swap of layers to DRAM or VRAM to maximize the latent space of your card. Also includes nodes for directly loading entire components (UNet, CLIP, VAE) onto the device you choose

Can you give me an example of what camera action you try?

@darksidewalker zooming out, but when i try it it just stays the same

"The camera pulls back until the subject is entirely in the frame."
I just tried generating with this prompt and it worked like a charm.
PS: Truly an incredible merge, thank you very much sir 🫶🏽.

@BobDF Cool hint! did not find the time to test the camera extensively.

I've been banging my head against the wall as to why I kept getting OOMs all of the sudden. Thank you!

Thank you!

@darksidewalker btw u said in the description no need any speed up lora, so i dont need lighting 4steps lora, cuz its already baked ?

@Seii1 Yes you don t need used it

@Seii1 Yes, don't use it

@darksidewalker i wonder is there a tool to make your prompt for wan more enhance or something like that

@Seii1 Can you describe a bit in detail what you mean? Do you look for a prompt enhancer?
All prompt-enhancer I know are not understanding NSFW content very well. All others are not worth the effort, too much hit-and-miss. They just guess way too much words.

@darksidewalker yes prompt enhancer

@Seii1 I have none atm.

Wow! I'm speechless! Thank you for this kind words and your elaboration!
I did not even think of going down with fps to 10 and if this is possible 😆 I'll try that too for some samples.

You are right "RadiantCrush" lacks some details for some "zones", but I really did my best to make this not happen on "LureNoir"! Plus, I incorporated anything hand-picked to get even more out of it in every tested scenario.

I hope, when the model gets free, you will have fun with it and can enjoy this too! I'm always happy if my work is used to create all these awesome art 🤩

Thank you! 🖤

@darksidewalker Yes, it's a bit extreme. Let's say I'm 100% safe with 12fps, but often even 10fps is enough. Anyway, yes, I can't wait to try the new model. Again, I'm really sorry I can't get it now, but if there are further versions in the future, I'll try to get them for a fee, I promise.

@BobDF Sounds interesting. I'll try experimenting with fps a bit. Just for science 🧪

And, don't consider yourself forced to do something. The major point is to have fun! 😉

If you did something you want to share, I would be happy if you post it to the gallery, that will support me too😄

The next model version will come at some point, if I can improve major things.🤞

What LoRA do you use for nipples?

@noog https://civitai.com/models/1419535?modelVersionId=2312689

It's a great model, but I find that I still need to use a LoRA for semen.

@magicballoon One says it is too much, one says its too less. I think give other prompting a try. And it is not a flaw if one needs 1 LoRA. A flaw would be stacking 2-6...😋

@darksidewalker Well, what I mean is, the fluid doesn't come from the right place unless I use something like the Anime Aesthetics LoRA.

I notice I have peculiarities with setup because I often have differences compared to what others say they use that I don't understand. Euler for example has never been a good sampler for video for me, despite me seeing many people say they use it. That's just one example, but it could apply to anything. It could of course be my settings or parameters but I wouldn't know what I'm doing differently. It may also depend on the source image I'm using.

So don't take my statement as criticism, it's just something I find interesting.

@magicballoon I'm not upset. Don't worry. Liquids coming from wrong places ... Yeah that happens. Sometimes it is just the right wording or tuning the LoRA strength used. It also depends on the image composition. Maybe it misinterprets some places here and there. I also found, that if liquids already present, like saliva or sweat it tends more to use them.

if you don't mind me asking what sampler/scheduler are you using?

@fierceys Euler+Simple. Works for me all the time.

@fierceys If I may, I also recommend euler+beta57, in many cases I noticed some better details, especially on faces, but I think it works best if used for 4 steps or more in Low

@fierceys Not sure if my input is helpful or not, but in case Euler somehow doesn't work for you like it typically doesn't for me, I've found DPM++ SDE and res_multistep to be good choices

Nevermind, I just tried out LightX2V's official workflow while going back to the basics and Euler works just fine. Guess it's time to contemplate my life choices.

@BobDF thank you. I would say faces generally are more important but balancing that when using these low step count workflows is difficult

Hi. How you ask? Reading the Version info on the model page.👍

LureNoir does not have baked clip, RadiantCrush has a spacial nsfw clip in. I decided to revert this. As there is not much benefit and it is a hard dependency. Lightx2v is included.
The rest is ... Here is the recipe:
- 1 drop of cat girl saliva

- 2 feathers of a gargoyle

- 23g pure masculine essence

- 1 hair of an elf woman

- 1 horn of an oni

... and a grain of salt. That's all! 😏

@darksidewalker So does that mean I'll need an NSFW LoRA as usual?

@rsflipson2 No try first without any. It really does a lot of things :)

@darksidewalker Sure does, tried RadiantCrush with default UmT5 and does NSFW well. awesome work!!

https://huggingface.co/Bedovyy/dasiwaWAN22I2V14B-GGUF

@qek 

nice. thank you. it is great

You just don't understand, those girls really love them!

More optimizations/enhancements and removal of the baked nsfw clip, you can tell by the size. And I don't think that LureNoir is for just Halloween content, that would be a complete waste. What made you think that?

@AI_man2025 it is like @rsflipson2 said. It is not made for Halloween content, It was just Halloween and ppl tend to make content for holidays!

@darksidewalker ah, just trying to work out what each version is more geared for, downloaded them all, so will test when i get a chance (dont have the hd space to keep them all)

@rsflipson2 I dont mean the version of the model, the variationsso not the latest version of Lurenoire, but what sets it apart from say Hotsprings, for versions, there is a "about this version" but nothing to say about the different variations. I am going to have to do some testing....still too many free models is a good problem to have lol

@AI_man2025 I would recommend the latest LureNoir, it's the most refined from my series.

@darksidewalker That is working really well, I shall stick to this one and leave the other variations a mystery lol

@AI_man2025 It's really just improvements, I don't think that the Creator has intended all these iterations to be variants that you have to choose based on your desired output..

Should work, if you have enough virtual ram (zram, swap, pagefile).

It works for me with geforce 4060 8gb, 32 gb RAM, 100 gb swap file

ok. what is swap file ?

@faryarjy100 the same as the pagefile, just on linux

@darksidewalker No, I have Windows 11

This computer - properties - additional system settings - advanced - performance - settings - advanced - virtual memory

@thismadworld8967 He asks what swap is, swap is the equivalent of pagefile. E.g. virtual memory.

It is a hard learning curve ... I would, say try to write like you would describe a video to a other person which can not see it.

@darksidewalker to be honest that doesn't make it easy since I am not a native English speaker, so I often describe things in a super broad term

@TOF_enjoyer I'm not a native english speaker too. I just use a translator or a tool to correct sentences.^.^

You can use grok ai to make your prompts its not filtered like chatgpt so NSFW content should work, i use below which works okay, im sure there is a much better way to make prompts tho

ask it to make something from this guideline:

1) Brief description of the setting/appearance

2) Actions, motions, story

3) Camera composition

so like below then you can ask it to change whatever is not good or to your liking

1) Dimly lit bedroom, soft warm light, wooden floor, large mirror behind.

2) She stands center frame in a cozy white sweater and black leggings; with a joyful spin she raises both arms high, hair tousles wildly, strikes a triumphant pose, beaming at the camera.

3) Static medium shot, eye-level, face and full body visible, sharp focus on her eyes at final frame.

@ghassananime1804 I was 1 hour younger when I decided to try grok, although I used it thro open router. boy is he nasty!

You can Vision-Language Models (VLMs) such as Gemini, Grok, GPT, or Qwen VL (Local), gemma3 for efficient prompting. This process requires defining a specific Role Prompt for the VLM and strictly adhering to the Wan2.2 Guide guidelines. if the input image contains NSFW content, you must NOT send the image to VLMs like Gemini or GPT.

@g1263495582 Neither the pretrained Qwen allows porn, use "NSFW" fine-tunes, there are some finetunes on https://hf.co/thesby

Not planned there are already some T2V models made, they are also very promising.

ComfyUI Manager V3.37

The error is a hint you ether use the wrong VAE or ur comfyui is outdated.

Getting the same issue. Updated nodes, comfyui and compfyui manager. Have correct VAE and CLIP. No other issues with other models except for this one.

Fuck, I'm stupid. I'm trying to run this as T2V. @fululin123 FYI.

@Kierkegaard420 you can say that but this error can only arise if you use the wrong VAE or CLIP. The last possibility is your backend is using SageAttention and it throws that error.
Or as I said, the backend has an error itself (is misconfigured/not up2date).

@Kierkegaard420 Okay. Yeah, that's not a T2V model ^^

@darksidewalker @Kierkegaard420 @Kierkegaard420 THANKS,I think my ComfyUI version is too low.

i don't have this error but an other one and it's due to the clip suggested in the description, try with the standard umt5_xxl_fp8_e4m3fn_scaled maybe it gonna help

Thank you again for your elaboration and kind words!
There will be a next version, maybe just not as fast as this one. But I have plans on improving... and I got some good ideas how to achieve that!

Try Rapid Mega then

RadiantCrush has a baked in NSFW clip, which was removed for LuneNoir, as it really wasn't necessary, resulting in a substantial reduction in model size.  

Indeed it is mentioned. "About this version", as mentioned on the detail page, there are the key points. Also I wrote a complete article about it.

@darksidewalker i don't see any loras mentioned here. Nor do I see your article linked in this model's description. I'm not saying you have to put it anywhere in particular, but when there are hundreds of nfsw loras it would be good to know which ones specifically to avoid. If that's your article, great.

About this version

✅ Optimisations

🌟 CFGZeroStar patch (better results and prompt adherence)

🍰 Baked in with latest VAE (2.1) BF16, Latest Distillation (r64-1022)

🍀Additional concept optimisation

🎬 Reward Attention (more realistic movements)

🌠 Enhanced motion compared to "RadientCrush"

✨Enhanced details even on low resolutions

🍌Enhanced basic understanding of 💦-related things (huge 💦 needs an additional LoRA)

😵‍💫Tuned for less hallucinations

🚫Excluded CLIP (again; more versatile)

🩻 Known issues

🤤 Sweetest Poison

@darksidewalker btw this checkpoint is great, don't get me wrong, i'm going through all of my more challenging prompts that base wan + loras struggled with now again, with very good results. I've always tried to create as much as possible with no loras since they generally always come with too many downsides, so I appreciate this idea a lot.

@IrakTheOracle Thanks, I'm glad it works for you and you have fun. You do not have to avoid any loras, except speed-up's (lower step). Just try without before, it can understand a massive amount of things. I also use loras, but the point of this is just that you do not have to stack a huge amount, only 1 or 2 impotent for the outcome. Or none at all.

on your screen, not in real life tho

@qek 😁You are sure about that?

@qek https://civitai.com/posts/24076317

https://hf.co/Bedovyy/dasiwaWAN22I2V14B-GGUF

I just want to say, that this quants are not made by me and I do not give any support on this or encourage the use. I tried all of them from gguf Q8 to Q2 and they are not the same quality.
Q6 till Q2 does not even follow the prompts like they should. Just saying.

@darksidewalker np, that is mostly because they are based on fp8 probably. if you share fp16 version of your models gguf would be easier to make and better quality.

@darksidewalker at least do a fp8_e5m2 for nvidia 3000 and amd users ... e4m3fn is slow for those

@xpnrt fp8_e5m2 will not provide speed boots to AMD or Nvidia 3000's GPU's. It just trades precision for dynamic range.

An FP16 version would not even run on my end or any normal consumer hardware.

I had to invest in external resources to do that and spend a huge amount of money.
And you know ... I do not earn anything from providing this model 😁

@darksidewalker it does , I can do torch.compile with it so do nvidia 3000 users, with e4m3fn we can't.

@xpnrt That's a compatibility thing with torch-compile. e5m2 itself does nothing for speed-up. Torch-compile, inductor or cudagraph are prone to errors and has almost no benefit on 4 steps, because the benefit will show of on many steps. So I see no benefit from providing these compared to the rising problems or edge cases. Just look up git how many problems rise with torch-compile. No thanks.😭
No offence, but I want the model as stable as it could be, without fiddling on dependencies. And e4m3fn is just the best for recent hardware.
I do not run AMD or Nvidia 3000, so I could not even test all that.

@darksidewalker if you arent using those then dont assume it doesnt help, it helps even with 3 steps ..

@xpnrt Okay. So if you need that, you could use Zluda for your AMD? I won't maintain and upload a different dtype any time soon.

You can detailed prompting for important parts. But just do not try to be extreme verbose. I got good results with more text and even with none prompt at all.

Thank you for your reply and model <3

wan 2.2 need to charge 2 model instead of one, so you need both

@Anima_Exhalare we start with High and refine with Low

If you are completely new to WAN 2.2, you could get the basics here: https://civitai.com/articles/20293/darksidewalkers-wan-22-14b-i2v-usage-guide-definitive-edition

Thank you, Sir!

Agreed. The model is crazy good. I just need to figure out a workflow to do start/end image for transformation stuff. This is fun.

@HYZ_AI My WF is really basic, but works with FLF, you can give it a try!

@darksidewalker Oh I will try. Kudos to your models. They are really fun to use!

@HYZ_AI Thank you 😁

Yes, 1022

I thought about that, but I decided against that:
- 5B will never achieve results like 14B
- I would have to maintain multiple different branches
- My time is limited, so the other model would suffer from quality and the 5B too
- 5B lacks extra resources and support from other creators

Overall it would be a waste of time and resources/money to do that for me. I myself would not even want use the 5B.

@darksidewalker Tried 1.3b and 5b countless times, not worth it

@qek My words 😁

If you want to prevent swapping you need more RAM. It loads one full model to your VRAM, therefore the 24GB allocation. But It needs the RAM for other purposes. With 64GB RAM there should be no swapping. But there are many dependencies like backend, configuration used attention and many more so I can not predict that to 100%.

@darksidewalker Ah okay, that makes sense. I'll try upgrading to 64GB RAM when possible. Thankfully, I can still use the page file method as long as I give my SSD breaks to cool down. Managed to generate over a minute of stitched together video using image to video and it looks great, thanks again for making this!

@darksidewalker I have another query. So I've made this nice long NSFW animation using image to video with one anime character on your workflow. Now I wanted to switch that to another anime character. I saw WAN has the WAN animate which allows character replacement from image to any video. But I think that model is censored? Do you have any workflows for this or any idea how to basically replace the character in my animated nsfw video without it being censored there?

@Jokerxx Sorry with WAN animate I can not help much atm. Didn't look into that in detail.

@darksidewalker No worries! For now I'm sticking to this i2v and it's incredible, I've been able to generate minutes-long clips stitched together with high quality, and it maintains character consistency too most of the time. Giving breaks in between, but 3-4 minutes for each generation of 5 secs, can only imagine the speed if I had more RAM haha

Welcome! Its warm here and we have cookies! That Catgirls are at your service, but beware they might bite 😁

Hi! First of all this is a checkpoint and would replace the standard wan 2.2 i2v checkpoint, not the loras.
Second I would suggest you this guide if you are totally new: https://civitai.com/articles/20293/darksidewalkers-wan-22-14b-i2v-usage-guide-definitive-edition
Third I have a fitting workflow for easy usage: https://civitai.com/models/1823089
Have fun and create something awesome!

@darksidewalker Thanks a lot! that's really helpful :D

@darksidewalker Hi, it's me again. Thank you for the guide you shared with me. I've read it and it was very helpful! Now I have a much clearer understanding of how the workflow in WAN works. I wanted to ask you something about the workflow you shared with me. When I open it, I get a message saying “Some Nodes Are Missing.” I was able to find some from the internal node manager in Comfyui (such as rgthree, for example). But I still need the following nodes:

VHS_VideoCombine
RIFE VFI
WanMoeKSamplerAdvanced
VHS_SelectImages

I haven't been able to find them in the Node Manager. Are they obtained from some other specific site?

As I mentioned, I'm still new to this, so there's a chance I just haven't searched for them correctly, haha.

Thanks again for your help, and sorry for bother.

@lalouwu This are the mentioned "About this version" on the workflow page. This are all official "custom_nodes". You can pull them from their github page or over the manager. I did pull them from the github. If you do not install them comfyui does not have the nodes for the workflow.

@darksidewalker Oh, I see it now! I feel dumb now. When I first read “About this version,” I thought it was some extra file inside the zip or inside the .json file. I didn't realize it was an additional section in the same page of your post. thanks!

@lalouwu Hrhrhr, no problem! Glad you found it 😉

For some specific actions/situations you need a LoRA. But looking at the samples and submissions there are plenty examples with no LoRA at all.
Prompting is also tied to the used input image. There is no prompt that fits all. Even the same prompt does not consistently give the same output on the same model. There is no way to compare WAN 2.2 models that easy.
You would have to make 50-100 examples with fixed seeds for a refined prompt used on every WAN 2.2 model and compare the median output of the best samples to get remotely an impression how the model performs.
Also you would have to find a situation all models can understand.

If you want just compare "penetration" as concept you should have to compare it without any LoRA and an initial image that remotely exactly shows a consistent understandable situation and clear visuals in at least 2d, 2.5d and 3d and a front, side and back-position. Making it to at least 27 simple initial images to test. You would have to use a fixed seed and the same sampler/scheduler, to get a comparable result.
A LoRA would always press its weights into any WAN 2.2 model. So this would not show the capabilities of a model.

That said, reducing a model to just one action and say this is better than others is ... well a bit too simple. This would never show the capabilities on overall performance.

PS: Basic WAN 2.2 will not understand that concept at all. Because it was not trained on NSFW content. You will see.

ah my bad i mean "im trying use this model and this model with same setting (no Lora) and prompt. then compare the result.. (im assume its same base model with WAN2.2).

the motion its good everything is good, just penetration its the problem, but when i see your sample (the BJ one) its really good, im have struggle with that too...
or probably Workflow affecting too?

hight quality? @saytolol12

@qek apologize its typo XD

@saytolol12 I do not know what to suggest to you. I really can not reproduce this problem. Maybe just try the prompts from other posts and look if that fits your image?

I think I have a similar problem. This model produced great results but it mutes deepthroat motion, both with and without deepthroat loras out there. Seems like it needs a different prompting strategy that I can't figure out.

That's not the checkpoint. You must have some misconfiguration on your generation settings.

I do not know what workflow you use. Could be various things. The thing is if you use the same image on 1st and last frame you get a double image at the end and dirty frames. My workflow prevent this with the "Exact loop" feature by cutting to the correct number of frames. There are other workflows that prevent this.
This is a WAN 2.2 problem in general, because WAN can not inject frames. There are some other reasons this could happen. There is no general parameter to prevent this.
Flashes or color shifts can also happen if used wrong loras or steps. Even the right frame count matters.

@darksidewalker I am using your workflow both with exact loop on and off.

In the end, after talking to @darksidewalker we were able to resolve the issue by updating my Python to 3.13.9 and my Pytorch to 2.8+cuda. I wonder what else these outdated dependencies were screwing up! Thanks bro!

@ambersai Why do I still have this problem after upgrading PyTorch to 2.8+ CUDA? I've also noticed that the higher the total frame rate, the more yellow the image becomes in later frames.

More versatile, less size, control to offload

@darksidewalker but it takes longer cause i have to use your radiant model to have a good clip

@PastellPastellPastellI No point to download the same text encoder two times, here's why they lack it

@PastellPastellPastellI You can just use the clip I mentioned on the model page itself. No need to include it. I run so much samples the benefit from baking it in is negligible.
This clip works for all scaled WAN 2.2 models.
You can just use my latest workflow or any other and set the correct clip, no need to mix it with my other model.

I imagine you are upset, looking at the page of DigitalPastel, that's why you are posting in rude language like this. That's why I thought I will not answer at all. You would like to know if I used your resources too. I'm quite sure I didn't. But you are doing awesome work and WAN 2.2 would not be the same without creators like you ... and so I decided to write you a DM, because I won't discuss private things and internal details on the page.
Have a nice day!

I'm glad we could have such a good chat and I hope I could wipe some concerns!
Good luck with your awesome loras!

They are refinements/iterations. More or less biased on something. I try to be as versatile as possible for me.
"LureNoir" is the recent update on my series. I just do not like using version numbers. I find it more entertaining to have names :)
I do not remember the time.

The workflowe is inside the video, what should I say? I prompted and after some samples (WAN does almost never get the desired result 1st try) I got this seed and outcome.

Honestly had the same thing happen too LOL, guy just wanted to keep going and going despite me prompting to stop it and hold it, or remove the pp, but I did manage it with an added lora for facial splash etc which OP also used in the workflow I think!

@Jokerxx I found that for better prompt adherence you can use a standard version of Wan2.2 I2V but only for HIGH noise reduction. For the LOW noise you can use the other fine-tuned model. This trick will produce the movement you want while retaining the great details of the fine-tune (like fluid simulation) without using any specific motion loras.

https://civitai.com/models/1981116?modelVersionId=2342740

You use the link @qek provided or you select the model you want from the names in the headline.

Thanks!

need a hand...

@nicholashoffmannewphone57 Hard to say. Which WAN is working, there are plenty.
Did you use the correct requirements, clip/vae? Did you try to update your CUDA/pytorch/python? Recent are CUDA13, pytorch 2.9.0+cu130, python3.13? Did you load High and Low and not mixing it up?

@nicholashoffmannewphone57 DaSIWA is Wan too. Also, you need 4-5 steps and CFG 1.0

@qek clean WAN (2.1 or 2.2, or with any lora) works fine. Problems only with DaSHIWA

@darksidewalker 

WAN 2.1 , or WAN 2.2 - zero problems

with Dashiwa - nan explosion

@nicholashoffmannewphone57 It is WAN 2.2 (fp8 scaled). So did you check and upgraded your dependencies? Did you read the explanations on the front-page?

@darksidewalker Can you tell me where to read all the necessary information? I can't find it

@nicholashoffmannewphone57 I can not tell you how to configure your RunPod and update the needed dependencies like pytorch, cuda, drivers, python.
The needed VAE and CLIP are on the front-page, with web-links.

@nicholashoffmannewphone57 What are your settings for the KSampler node? Specifically, which sampler and scheduler?

@g1263495582

"guidance_phases": 2,

"switch_threshold" : 900,

"guidance_scale" : 1,

"guidance2_scale" : 1,

"flow_shift" : 5,

"num_inference_steps": 4

@g1263495582 I'm using Wan2GP

@darksidewalker I'm using Wan2GP. Wan 2.1 works fine, Wan 2.2 works fine, LoRAs work fine. ONLY DaSiWa gives colorful noise (NaN explosion). If it was a hardware issue, wouldn't everything fail? What specific settings does DaSiWa need that differ from regular Wan 2.2?

@nicholashoffmannewphone57 I read sometimes Wan2GP is not compatible with scaled models. maybe it is just Wan2GP. It is not universal compatible or lacks features as far as I'm aware.

What specific settings does DaSiWa need that differ from regular Wan 2.2? - None. Just the settings describes on the front-page.

I may not help much with Wan2GP, because I do not use this, I just use SwarmUI and ComfyUI.

@nicholashoffmannewphone57 I was wondering what wan2gp version you are using. If it's not the latest, can you please try to update it? I can't help much, as I don't use wan2gp, but if that doesn't work, try submitting an issue on GitHub.

@g1263495582 WanGP v9.3 by DeepBeepMeep

It will do it all, there are plenty examples in the gallery :)

@darksidewalker Because they all are img2vid

@qek right this is an i2v model

There is a rogue one without permission, but since it is a scaled model it does just degrade. A real gguf is not possible atm for me. Sorry. I wouldn't recommend to use it.

Hehehe, no offense, but that only happens if you prompt is bad and WAN has to guess too much. Just write better prompts.

Read my guide.

The behavior may be slightly more because of reward attention, but that's a general WAN 2.2 thing.

Read the guide, it will explain many things in detail.👍

You guys should really read the info and my guide😁

Besides negative is deactivated on cfg1 WAN can not idle movement and guess things up if not prompted.

Use flf + loop if you want idle and a short frame count.

@darksidewalker I didn't even realised the cfg1 didn't take negative prompt, not used to the lightning lora, but super great for fast render.

I've actually been trying to upgrade my looping animation from Wan 2.1 720p to Wan 2.2 for the past month and holy snap there are so many tiny details of issues pilling up. Default model would incorporate color contrast over time, which breaks the loop effect. I think only your model removes this issue as apparently a merge of lightx2v lora + main resolves this compare to use them seperately? Not exactly sure what's going on behind the scene. I originally thought it was due to the Wan node pack. Wan 2.2 animations are just way too good + way more compatible loras for HD results.

I tried your last and new workflow, and loved to see the whiterabbit model in there as this workflow almost had everything I need, but 5B was too low res. Even in your workflow, there aren't much difference that my main loop workflow, so it is still moving too much at 3secs.

I guess I'll go back to Wan 2.1 for idle as I've tried so many tweeks the past week. If I use the regular model, I do get less movement, but color contrast issue appears.
Thanks for letting me know!

​I am using Normalized Attention Guidance (NAG) with the following negative details, added to the positive prompt: (silent:1.2), mouth closed, neutral expression, no speech, no lip movement, still face, expressionless mouth, and no facial animation. The results are generally good, although I still get occasional inconsistencies.

@g1263495582 I thought about incorporating NAG in my WF, but I need more testing.

@ItsThatTimeAgain I did a lot to make the model consistent. But you have to consider, that WAN 2.2 can not inject frames itself, so you will never ever, get videos without any artifacts when looping. WAN just does not understand this at all. WhiteRabbit nodes get really close and could be considered a perfect loop at this point, maybe it is tunable. But as you mentioned it is unbelievable difficult to manage.

My WG uses a different upscaler, which minimises colorshifts.

WAN 2.1 has just less motion at all, that's why it will be somehow better for this niche use case.

I think a perfect Idle would only be possible with a very good lora. The "Wallpaper" lora works somehow, but it also will not be perfect.

@g1263495582 Damn I didn't know about NAG, that looks awesome! Do you simply add a WanVideoNAG node after the negative clip and before the Sampler? I see there are some NAG samplers but not sure if it has a start-end value in them.

I'm not quite sure if NAG is that good, because it is not actively developed and the dev does not respond to any issues on git. Officially it is only for WAN 2.1.
So it is just not predictable if this even gimps anything from WAN 2.2.
I'm not so hyped yet.

@ItsThatTimeAgain ​I connect the flow from Model Load -> LoRA -> WanVideoNAG -> then into the KSampler. The Conditioning input on the WanVideoNAG node is connected to the negative prompt, and I perform this connection sequence for both high and low model. However, you must also specify the elements you do not want to change or move within the positive prompt.

@g1263495582 This actually helped out with the mouth a bit! Had a to generate a few tries, but it did in the end shut the character up and reduced the movements.

I've also added the live wallpaper lora to the low model and it kind of help. I might want to increase it's strength higher than 1 next try. Feel free to hit me up if you find better alternatives, but this is better than nothing for my needs.

@g1263495582 I try your method, but the character`s month is still talking and opening, do I set the wrong step?

@yazi6393wy136 Could you share your workflow and prompt ?

@g1263495582 Dasiwa`s workflow https://civitai.com/models/1823089/dasiwa-wan22-a14b-high-low-i2v-and-flf2v
and prompt is same as yours only type that

It is not a V2V model. The error could be a wrong clip or vae.
It is WAN 2.2 i2v.

@darksidewalker I tryed clip: umt5-xxl-enc-bf16.safetensors, umt5_xxl_fp8_e4m3fn_scaled.safetensors.
vae: wan_2.1_vae.safetensors, wan2.2_vae.safetensors.

Also found something like this in ComfyUI-WanVideoWrapper nodes_sampler.py:

if transformer.ref_conv.weight.dtype in [torch.float8_e4m3fn, torch.float8_e5m2]:

raise ValueError("Fun-Control reference image won't work with this specific fp8_scaled model, it's been fixed in latest version of the model")

There was also error 'NoneType' object has no attribute 'weight'.

ChatGPT said ref_conv is not provided in model. Could not find anything online, that why asked.

So, there is no way to somehow control video? Using controlnet or smth?

Checked vae Wan2_1_VAE_bf16.safetensors from this page, got same error
File "/home/.../comfyui-venv/lib/python3.12/site-packages/torch/nn/modules/conv.py", line 712, in convforward

return F.conv3d(

^^^^^^^^^

RuntimeError: Given groups=1, weight of size [5120, 36, 1, 2, 2], expected input[2, 48, 6, 64, 64] to have 36 channels, but got 48 channels instead

@trammediddusu9402511 You can not controlnet or v2v, as I said its just i2v. For that you would need other WAN 2.2 models like FUN-InP or VACE.

@darksidewalker seems like it. Thank you

@trammediddusu9402511 Kijai's Wan Wrapper supports ControlNet

@qek do you have a specific workflow to try? Because all I did before not worked (for this model at least)

@trammediddusu9402511 I don't use the wrapper

Don't use the custom node then

Using correct resolutions should fix that.

Check your clip model,it should be umt5_xxl_fp8_e4m3fn_scaled.I had the same error when I used no scaled version.

Update drivers and set --lowvram in config

@DeviantApeArt Sadly nothing changes, comfyui just shuts down when it tries to load the checkpoints now. My setup just refuses to run the workflow as it did before.

Must be something with your backend. The sampler part did not change and the checkpoint is also not changed.
My advice try a fresh install of comfyui and use python3.13, latest pytorch and drivers.

Some users decided to install an old version of the backend

Tried fresh comfyUI installs, nothing, tried reverting pytorch, no dice. The model was awesome when I used it, no complaints there! I made this comment just so I can give a fix for future users if I ever find a solution.

@rastaract --highvram --dont-upcast-attention

Yeah, its been happening to me. One day it works and another day it doesn't work, freezes my PC and stuck at 39%. It was working up to the point I tried out a different workflow and boom. I think the inconsistency makes me feel like I'm close, so I'll bite the bullet and trade my RTX 3070ftw3 for a rtx 5060ti 16gb. I'll wait until black friday and maybe I can get it a bit cheaper..

Update: Now CLIP doesn't even load, before even trying to load the checkpoints. Tried qek's flag suggestions, didn't work. Multiple reinstalls also didn't work, task manager isn't even able to show a spike in VRAM before Comfy gives out.

I think its something to do with the system ram not really liking using all 32gigs, once it goes from 29 to 31 it really starts having a stroke. Where as the vram sits at 40%-70% range.

@SugoiDess Yeah, pretty much exactly what I'm seeing on my setup. Really wonder how I was able to use it before, I didn't even try to close my other chrome tabs then and it worked fine.

There are bugs in the latest comfyui from before 2days. 0.3.68. Again they broke video generation for WAN2.2. You can downgrade atm or use this startoptions: --async-offload --reserve-vram 1

At least for me it fixed the long generation times and OOMs.

downgrading my version also isn't solving it. I'm less mad that it doesn't work on my hardware at this point and more perplexed at how my computer ever ran this at all. I didn't set up a manual pagefile, had all my chrome tabs open, and it ran fine then, yet now when I actually look into task manager there's no physical way my VRAM and RAM at that time could have fit the CLIP and both of those models in.

Actual witchcraft, I tell ya.

@rastaract you never load both model at the same time. Always only one.

But for me I had to completely reinstall comfy yesterday because one of their updates made it impossible to render anything with wan 2.2. even with version switches, something went wrong in the python dependencies I assume, had to recreate the complete install including venv.

@darksidewalker Really? Wait, then I'm even more confused...
The ComfyUI shows that on the Wan to Video step, it tries to load the high model, then the low model, right after each other. It crashes or throws out an error on the low model loading, but if what you say is true it shouldn't even be trying to load the low model at that point in the process anyhow. Really weird.

I wish I had remembered what the process looked like when I was trying it the first time. Can you let me know if the low model node glows green right after the high model loads for your workflow?

@darksidewalker what are you talking about? aren't you suppose to load High + Low model at the same time? It works just fine for me.

@DeviantApeArt no that's the whole thing with MoE, you have several checkpoints loaded for their purposes, each after another. You would need 2* memory if not MoE and would load a 28B model... You would never be able to use it on consumer hardware.

@darksidewalker Maybe you misunderstand what I'm saying. I'm using your workflow, I load both High + Low noise model and it works fine. Are you saying I should load only 1 model, either High or Low at a time?

@DeviantApeArt no I didn't say that, I say it will only load one at a time given. @rastaract mentioned it would load both and he gets oom.

Well, figure I should reply with the solution I found after everything... RAM. Had to keep most of my open chrome windows closed, and now it runs fine. Reset pagefile too, since I didn't edit it the first time.

@SugoiDess hopefully you won't also need to upgrade your RAM with the VRAM.

@rastaract My "lightweight" browser is eating 2.500 MB of RAM rn, probably need a real lightweight browser

@qek Firefox ist really memory efficient, all chromium based browsers will eat your RAM.

@qek What you should do is heavily debloat pc, idle it should use no more than 4GB RAM