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

update comfyui

@Darksidewalker Hello. Did you happen to receive my chat? I'm mentioning this because I haven't received a reply yet ^^;;

@Darksidewalker Haha, thank you. It's a small token of my appreciation, but I sent you another cup of coffee under the same nickname. Thank you always for being such a great model!

@Darksidewalker lovely mate

right, after I updated comfyui it worked, TY

Indeed, and not only by prompting motion-specific keywords. Using some classic motions loras at low strengths will allow almost anything using boundbite without impacting the quality. currently a 10/10. Will now see if SnatchKiss can offer something more :D

@Darksidewalker 

I just tried it with your default workflow and it actually worked! A 5-second, 720p video took about 11 minutes and 22 seconds to finish. It’s roughly the same render time as the Q4 GGUF versions I grabbed from your other posts before.

the best version for 3060 (12gb) is GGUF Q5 model, I've already tested it and works well.

@Darksidewalker i think its cause you added nsfw videos to the showcase on this one cause the others show up on there

Thanks for your reply. No by mistake i hit the combine switch thought it was my updates caused the trouble. Then i saw "load FFmpeg" was set to "output\video" causing an error, but offcause there was no videos to combine in the path. All my mistake, sometimes i hit things when dragging around. Once again thanks for your reply and for a very nice workflow.

yes, i've started using SmoothXXXAnimation lora, it helps a lot

@innyinny Are pairing that with this? I tired it and I'm getting blurry motion

Try to cut the last Frame out, because the last is also the first frame. The double can cause your slight pause. I use the ImageSelector Node for this. (Sorry I'm not a native English speaking guy :) )

@Urgathum Thanks for your reply! Which exact node should I use for that? Do you mean the “Select Images” node from VideoHelperSuite, the one with the indexes setting?

@rakkenshoo exactly. There set Selected_Indexes to 1:0 and done.

On that node of looping. Do you or any other here know how to create a Seamless Loop with SVI combined? I tryed WANFirstToLastFrameToVideo together with SVIPro. The length mismatches and won't work.

@Urgathum Unfortunately, that didn't work. Is there another method?

I'm sorry, I don't know.

@rakkenshoo I Can only try to discribe my Workflow for you (credit goes to a youtube user):
I use the Dasiwa BoundLite V10 safetensors Model High and Low (Credit for the model...works realy good thanks :) )
Load Clip and VAE, total standard. Connect the Model to the LORA Loader (I use the 4Steps lightning LORAs even if it is not recommended, just because I get better results that way...don't ask me why, just less SlowMotion).
Shift High and Low model by 5.
Connect the shifted outputs to the Advanced KSampler Nodes (CFG 1, Euler, Random Seed, 4 Steps, Return Leftover Noise enabled for high dissabled for los yada yada yada)
Connect the Loaded UMT5 Clip Model to the Positive and Negative ClipTextEncoder, Connect the 2 Conditionings to the WanFirstLastFrameToVideo Node. Connect the VAE and the Loaded Image to this node too. First and Last Image both imputs should be connected to the same Image.
Connect the positive and negative Conditionings to both KSamplers, Connect the Latent to the high KSampler. KSampler Latent high to the KSampler Low Standard.
Connect the Latent from KSampler Low to the VAEDecode Node. I double the resulting Images in a RIFE VFI Node for smoother Animations. Then I cut the First Frame with the ImageSelector Node (1:0). Last Connect the Finished Images to the VideoCombine Node (32 Frames/s).

Works, not always perfect, but it is most of the time Seamless. If you want you can add a ColorMatch with the StartImage as Reference, but in my cases it cause more trouble then worth. Try Prompting for a loop, like "return to starting image, seamless looping animation"

@Urgathum Thank you for your time.

I'm getting really slow motion in v10. Wonder if I should try v8

@badswagmuffins499 You can try, I think the generated results are also related to Lora to some extent

@voidyear My problem is I will run the same seed and prompt to test it and there is a major difference between the other checkpoints and V10.

@badswagmuffins499 Each checkpoint has its own characteristics, and it seems that the method of controlling variables is not applicable. I found that the reason why the video is slow may be related to my frame rate of 16. After I changed it to 24, it became very normal, but the 5-second video became 3 seconds, and the 9-second video became 6 seconds. I only added the following lora, https://civitai.red/models/1307155/wan-22-experimental-wan-general-nsfw-model

Does SVI deliver good results with lightspeed?

Hi, the more motion would be cool. Curently using Smoothmix 2.0 for High. Low Dasiwa is the King of the Kings for me.

hmm I never thought about trying to mix checkpoints like that, I usually try to stay with the High-Low pack they get released as but, its actually a really interesting idea thanks for the post

@FarAwayDreams The thing is that Wan 2.2 types of models have their place. HIGH offers the motion and the base of scene. LOW works as the details and finetune of the video.

After all, all of this is endless trial and error. When you think you found perfect combo (smooth high + dasiwa low) and then on another image it acts totally different, you play numerous attempts with different seeds, shifts, prompts only to find out that in this particular case dasiwa high gives perfect result in one shot.

@jnkraftmail703 thats how wan 2.2 works.if u change the aspect ratio,change resolution from Ref Image from 1:1 to 16:9.or if u have 2 person in the Ref Image and one person is looking at another direction is enough to change the whole outcome.a full "new Scene" is better but faces might shift more.i also use 4 total step with 1 step low and my results are the same. at 0.65 Balanced precision the outcome is usally the same if i go higher resolustion.

Although not recommended you can always use lightning LORA at the cost of quality and consistency.

I sometimes use it to make the scene faster in motion, but don't use motions which travel too far from the initial image or you'll see a lot of hallucinating.

I am using this with 8GB Vram.

Yeah, but short and not too high resolution

Use the GGUF version

@Darksidewalker  thank you so much. >-<

Increase the steps

Yep increase the steps or do it my way: I only use the Truevision checkpoints + An good speed lora with the right weights, if you are interested i can tell you my settings. Using your own speed loras and not the baked in ones gives you the best results and high details at only 4 steps!

@teeay25225 Hi, I interested in you recommendation for the speed up lora< There are so manz of them right know, when I tried the models without lightx loras baked in and used different lightx, the outputs differs much and most of the time losing the details. What is your best bet in speed loras?

@HugMeIntoFace Sure! HIGH LORA NAME: Wan_2_2_I2V_A14B_HIGH_lightx2v_4step_lora_260412_rank_256_fp16. LOW LORA NAME: Wan_2_2_I2V_A14B_LOW_lightx2v_4step_lora_260412_rank_256_fp16. Both are 2.32GB big. I think you cant find any better speed lora atm and the results are VERY good. Like i said use the Truevision checkpoint and the speed lora at: High: 1.5. Low: 0.8. And you will see magic happens because u get much more details at only 4 steps. EXTRA advice: If you use the standart Fp8 Checkpoint as high and Truevision as low u get even better results. Also: This is a 4step lora so you use them only at 4steps, higher steps worsens the result.

@teeay25225 u can share a workflow with your setting???

@teeay25225 OK, thank you I will test it. One thing what do you mean by standart FP8 checkpoint? (You mean original WAN 2.2?)

@enterkmfdm2016848 Sure! I use the "DaSiWa WAN 2.2 i2v FastFidelity C-AiO-12" since day 1. I tested alot of workflows and this one is definitely the best one i tried yet. You can use it with the default settings and with the speed loras i told. Dont play around with CFG or Sigmashift. What you can do is testing different samplers and schedulers when u get more advanced.

@HugMeIntoFace Yeah i mean that. You can search for "wanVideo22_i2vHighNoise14BFp8Sd". If you have any other questions ask me here or dm me. I like to help beginners because i was at your point too 5-6 weeks ago.

Where can I find that flow? I only see version 8. My apologies.

@enterkmfdm2016848 No problem im here to help. I could not find it either so i uploaded it for you, here you go: https://uploadnow.io/f/xYYBrBk

@teeay25225 Is upload got your settings you use or is this just the default?

@teeay25225 can you help me a bit? Basically I'm having too many issues with motions I want fast motions but I get medium to slow one's I most of generations. Also I don't know if its me, but most of time I don't get the walking of character right is it wan 2.2 issue or these checkpoints by Dasiwa as I also tried smoothmix checkpoint I don't get much issues with walking but in dasiwa character walks back after 4 or 5 seconds of video?

@yeahnohell617 Did you even try my recommended settings? If not do that. But the main fix is using the standart FP8 Checkpoint as high and dasiwa as low. This is empowering wan to faster movement. Also: Yes its normal that Wan is jumping back to the beginning pose after 5 seconds of video. There are workflows where u can make longer videos, just search for that like "wan 2.2 long video workflow".

@teeay25225 Thanks for replying! Is wan 2.2 i2v 14b fp8 scaled as high noise good? Also do I need to add speed loras if using fp8 scaled as high?

Sorry, it's 1.5. I didn't understand, I'm sorry.@teeay25225 

@yeahnohell617 I switch from the Dawisa Checkpoint as high to the standart one if wan should listens more to my prompt or i need more movement.. Which is maybe fixed with the new snatchkiss release, i didnt try it yet because i wait for the truevision version of it. And yes you should use the speed loras instead of the baked ones, they have more details at only 4 steps. Means: You can also use the Truevision Dasiwa checkpoints as me and the speed loras. And you switch between dasiwa high and the standart one in different situations. But if you are a bloody beginner you can use the lightspeed dasiwa checkpoints the first 3-5 weeks and must not touch ur own speed loras yet. I would not say that the standart checkpoint is better it's different and any checkpoint has its use :-)

@teeay25225 Reading your whole discussion. Which one counts as STANDART ONE?

@HugMeIntoFace "wanVideo22_i2vHighNoise14BFp8Sd" is the file name. It's the standard wan 2.2 14b checkpoint without any changes. You can find it here: https://civitai.com/models/1817671/wan-video-22?modelVersionId=2057465

@teeay25225 yoo, can u by chance recommend a workflow for making longer videos with the checkpoint mixture u recommended? if u dont have any thats fine, just wondering

@shindaiwa321 Wan2.2_I2V_PainterI2V_Workflow_Kenpechi_v2.4.. Is simple and good! Many workflows are too complex and making you overwhelmed but this one is fine.

Excuse me, how can I get more clarity in the videos, almost like Qwen Online Qwen 3.7 Plus?

I was having the same issue. Try Sampler: Euler; Scheduler: Simple.

After comparing around 200 clips between v10 and v11 to find the sweet spot, I used v11 for 9 out of my last 21 animations with Kushina (including the one I recently shared). It performed better than v10 on those 9.
v10 is still amazing, but v11 seems to handle some difficult motions better. I'll keep using both. Currently using Euler + Normal.

Would you describe improvements more detailed? I use SynthseductionV9 and cant find proper info about progress on V9=>V10 and V10=>V11.

The quality of the videos is sharper. When using the previous model I found them to be soft. Not blurry, but soft. The lack of a "warm up" is a big improvement for me, although some generations still have it. I haven't tested everything, but the movements usually look better.

I have the same issue with v11. To much looking around or yapping or smiling .very unnatural face expression.details are better and overall sharper.but that random face expression are so bad...

I will stick to v10.....

Currently im using remix v3(high) and dasiwa fp8 v10(Low)

@hartweizen I've made a bunch of generations this morning with better results. More detailed prompting can help, and so can adjusting the shift. For some reason the shift seems to make a bigger difference with this model than others I've tried, but I'm still new to doing video so take my words with a grain of salt.

Okay I just did a generation with a new image. The woman started talking and smiling in the first iteration. Instead of "she does not speak", I said "she clenches her jaw and turns her head slightly". Fixed. I've found that the negative statements don't work as well as positive ones (if you don't like what they do, instead of telling them not to do it, tell them to do something else). Maybe you already know all that, but hopefully it will help a beginner.

I tried 720p on my rtx3060 12GBVRAMM with 32B RAM, it works although it takes like 20 minutes for a 15 seconds SVI workflow, also try to use this version without lora first.

what do you use for facial consistency? and what have you found as a good sampler

If you switch schedulers such as from normal to linear_quadratic, make sure to check out the sigmas schedule using a node like SigmasPreview (R3S4LYF). AFAIK, for I2V the rule of thumb is to switch from High to Low noise model after 0.9-0.95 sigmas remaining. linear_quadratic schedule processes less sigmas in the first iterations. Therefore you likely need to shift your steps so there is more steps generated by High Noise model.

Facial consistency is great for me. I use 8 steps though, 3 for the high and 5 for the low. But, even 4 steps is great.

@Limeturt  Personally, I found Euler + Beta 57 to be the best in terms of face consistency and detail.

@altoiddealer So that was the reason. I'll have to give it a try. Thank you for the detailed explanation.

@DaddyWolfgang I tried it dozens of times following your advice, and you were right. Thank you.

@mangho9123389 Glad to help! It DOES take longer, but if time isn't an issue then the 3 high 5 low method tends to result in more accurate and better quality results. I've yet to find a workflow that works better than this, while they can be good, they often fall apart once you start asking for more complex movement and speech. When in doubt, increase steps to test. Higher CFG too, but that will increase gen times a lot. Darkside makes incredible models and has always been helpful when you ask the right question in the right way.

Not that you did, but if you notice when people come in and insult him or his model when things don't work right, he in turn gives that same energy back. Because he, and I think everyone else knows, these people most likely didn't try ANYTHING to resolve their issue first. In those cases, honey works best. They could simply ask, "It's obvious this model is great, given the examples I've seen from you and other users. However, for me it's a bit off. Why does X happen and what can I try to resolve the problem? Here's my workflow."

Instead they come in with, "FUCK THIS MODEL SUCKS I KEEP GETTING DISTORTION PIECE OF SHIT MODEL" And then they ask for help...

Darkside doesn't have to do any of this for us, they could keep the model to themselves, they could continue to charge for it as well and lock it away. But no, they release for free eventually. They also don't have to respond to a damn thing, but I think they truly do want to help but when you're an asshole then expect an asshole response.

And honestly, they're very calm considering some of the stuff I've seen said.

你好，

我遇到了同样的问题，我想知道你的问题是否得到了解决？能告诉我解决的方式吗？

https://civitai.red/models/1981116/dasiwa-wan-22-i2v-14b-or-lightspeed-or-safetensors?modelVersionId=2953485

Also, sometimes the mouth movement seems a bit excessive, as if the character is opening and moving their mouth too much.

Why does this happen?

@xuro To counter mouth movement: add "speaking, talking" to negatives. Don't know how to help your lora issue though. I'm using the WAN workflow by DaSiWa and all the loras work.

i had the same problem on my old install, no resolution control on SVI workflow and LTX refused to install correctly for the AIO; i made a new install using darksidewalker's own installer that i got from his page(that also automatically installed the workflows) and everything works now-- i still have one disconnected field on the AIO workflow(the very last one under upload watermark), but i don't know what it is or what it's supposed to do, ¯\_(ツ)_/¯

@aiaiaia4444 I had to downgrade the node version to an older one to fix most of the disconected fields except one, the field to select the seed behaviour. But I noticed that is not only on this workflow but on other workflows too. Sure it has something to do with comfyui cause I have the problems since it updated about 1 month ago.

same with nvfp4 idk why

No it's not supposed to take longer. GGFU is always slower. That's the whole point.

@ODBsGoonerExtravaganza I mean the results of fp8 lightspeed weaker than nvfp4 and Idk why

@Giey098 Al I know is that truespeed is indeed better. My current setup with Truespeed is 9 steps High Noise and 5 steps Low Noise with latest rank64 light2x lora. 15 minutes for 720x900 6 second clip, 2 of which spent on interpolating with GIMM. I run 12GB VRAM, 32GB Ram.

Me Too Same Problem i back to V10 for SVI 2.0

Can confirm the same issue.
Snatch Kiss distorts with a noise-like filter, sometimes starting as soon as the first second.
Bound Bite goes crazy with the camera movement.
Tasty Sin works marvelously all of the time.
Tested in multiple SVI workflows, including the one generously supplied by DaSiWa.

https://civitai.red/models/1981116/dasiwa-wan-22-i2v-14b-or-lightspeed-or-safetensors?modelVersionId=2953485

I had the same issue. The link in workflow downloads some 115KB file, which is too small for actual VAE. Use the link in dependencies instead.

Okay with further testing, it seems enabling NAG and putting "smoke", "steam", etc in the negative prompt works some of the time. It doubles generation time, though, so I'm hoping someone else might have a better solution.

@Darksidewalker I've rewritten and reviewed hundreds of times, and am still unable to figure it out.

I've tried a couple, yeah. It's a mystery.

I just posted two mp4s to the gallery. They both have the smoke thing but it's more subtle than it usually is. Even still, it's pretty easy to see in the second one I uploaded. https://civitai.red/images/134352963

I have a screenshot of the workflow made just after the generation, but I didn't think uploading it to the gallery would be appropriate. Let me know if you want me to send it to you somehow. The only modification I made to the workflow was to delete the save node for the mini meme. It was throwing errors and I didn't need it.

Anyway the last upload was with no lora. The first one did have a lora enabled, which is why I made the second one and uploaded it.

@Darksidewalker My prompts are usually much longer. I just wanted to show you what it was doing.

Try writing the word cold in the prompt. When describing the video maybe the words hot s$#@$ was used.

I finally figured out that this was happening with images I'd edited with QwenRapidAIO. Since I started using Krea2 I haven't had to edit nearly as much, and haven't seen this smoke again.

I have no idea why. Maybe something to do with that moire pattern that tends to show up in Qwen images?

@Darksidewalker Guess so. Weird, huh?

Happens to me sometimes too. But rarely, like 1 out of 6-10 ejaculations. My prompts are simple, it does tell ejaculate/cum. Like
"Animation of a woman with blonde hair performing oral sex. The male withdraws the cock and ejaculates across her face and mouth."
I use 1-3 loras. Most of the time I have high and low DR34ML4Y_I2V_14B at ~0.55

@naxo556x How did you generate your source image?

@hawkfreedom1177 I usually wait until I'm done before lighting up, but whatever floats your boat.

I've had this problem again a couple times this evening, despite feeding it very high quality images from Krea2. But it's not all the time like it was. I think the worst one happened when it was a 1MP reference image.

easiest if you are using their workflow https://civitai.com/models/1823089/dasiwa-wan22-workflows-or-i2v-or-svi-20-or-flf2v and putting the files in the correct folders

Even easier, ComfyUI has his own simple Wan 2.2 workflow that works perfectly well. Really dont know where you are trying to use it.

@fatessa ty, fixed

@SeikKV, ty fixed 

@Darksidewalker It was I2V, but I started out with a woman who was seated and her lower body was hidden. I make her stand up and I guess the algorithm decided she must be a mother of 6 with a steady income of child support. I eventually got it to display a less "MILFY" appearance, but it took some creative prompting.

@Darksidewalker ok then, thanks

Had the same issue occurring, it occurs a lot less when using a random LORA on 0.5 in my tests, since negative prompting and your other prompts like mouth closed didn't work for me too.

For me, it's not a problem most of the time. There have been several images I've tried to animate that the model seems to want the characters to talk, though. I've found the best way to counter it is to tell them to do something else with their faces. For example, prompt them to squint or grin or turn their head slightly to the left or right. Any kind of subtle instruction will usually stop it. For me, at least.

@potatometer350 Yes this works too, I added "neutral face expression" in my prompt but it does create a static emotion sometimes.

The best solution is just to render again, v11 outweighs the issues I have with much better results anyway.

I got it working! Had to load the SVI2 loras.

ANYHOW - any ideas how to get faster motion while copulating? :D Did mo aplitude to 1.4 and added lots of prompt and negative prompt stuff, but rather... well .. calm. Need really fast like jet engines, any ideas appreciated! :)

@Darksidewalker That's a good point of view =)

Alternatively, you can try running the workflow with either the bounce test or slop twerk lora's at negative strength.

@RL1775 Thanks

To create loops with any Wan checkpoint use PainterLongVideo from PainterLongVideo node pack (princepainter/ComfyUI-PainterLongVideo)

This node will also help you out, or rather is necessary-must-have if you want to make movies.

+1

+1

@Darksidewalker Oh my, thank you! Still not used to it with that new drop-down selection. I'm so sorry for troubling you.

Different models do different things. I appreciate the slower motions from this model, as I usually like to generate more intimate scenes. But if I want something faster, I can use Remix or whatever as the high noise model.

Im hoping for the same thing!

Author published the Int8-Convrot versions!

NOICE! Thank you sensei!

There is the list at the top of the page where you can switch between versions and high / low

@Darksidewalker Thanks!

Awesome!!!!

@Darksidewalker could you plz upload a int8convrot on huggingface？i really want to use int8conrvot，but what i download here is also the fp8

@Darksidewalker maybe not the fp8，i am not sure，they have the same file name

sry,it is the correct int8convrot，just the same name different file

If you hover the mouse over all the different versions (FP8, INT8, etc) every one of those has some special suffix at the end. Regardless of which you click, though, the final downloaded file will have the same name which has no suffix, for example DasiwaWAN22I2V14BLightspeed_snatchkissHighV11.

Anyway it's downloading the correct file based on what you click. I downloaded both the FP8 and the INT8 and they both have same filename - but are different file sizes with different hash, so they are two different files.

@altoiddealer Understood, thanks for that !