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

I did one (not public yet).

Here comes the BUT - LTX-2 is not anyhow good at the moment.
It sounds good in theory, but it lacks consistency, the audio is bad, the motions are bad, long videos have distortion, there is heavy warping and morphing of details and ... it is really low visual quality. You need extra loras for every camera movement... overall it is not very user friendly and unpolished.

If there is any technique or improvement to get it stable, I'll gladly release something.

For now I would not recommend it.

@darksidewalker Alright thanks a lot for the clarification :D

@darksidewalker In your opinion, which video styles is LTX-2 best suited for when it comes to i2v?

@g1263495582 Super short 5-10s tiktok like dance clips XD

@darksidewalker I don't have to delete all my WAN models and hack together a new workflow for a while longer :D

I have the same problem. The penis does not insert properly into the vagina. And the vagina orifice is too high.

"the penis completely disappears inside her vagina,"

you can also add helper actions like "his hips smash against her waist" etc.

i find the dasiwa checkpoints to need a bit more aggressive prompting than some other checkpoints for certain actions

其實提示詞應該可以控制，建議你找Grok修改，像是乳搖的敘述，就可以從單純晃動改到胸部互相拍打

Q8 offers better quality than FP8, while FP8 is roughly equivalent to Q6 or Q5.

@g1263495582 That's not entirely correct, unfortunately.
FP8+ (mixed, mine) > Q8 > FP8 (simple) > Q6 > Q5 ...
Also to mention that safetensor > gguf for efficiency.

Hi! Yes your comfy is not up2date

Yes, my friend.

Thanks to your suggestion, it finally works.

You truly are my AI teacher. ^_^

By the way, the latest Workflow (AiO v3.2) generates videos much faster, it's amazing!

You need some penis lora to maintain the shape. There are different sizes of lora for different penis behaviors, and it depends on your scenario and position to decide which lora to use to maintain it. If you don't understand, you can ask me specific questions.

KJ's workflow also encountered this issue, but the official workflow worked normally.

Guys ... just update your comfyui ... it's stated in the version notes, also it's mentioned countless times now in other posts. Stop running an ancient comfyui version.

@darksidewalker yeah i figured that out....works great after update haha.

@darksidewalker yeah...But the update screwed my Vace generations... no solution yet. But your Tasty Sin 8.1 Model is the best model out there ... So thanks! Worth the space of two seperate ComfyPortable ♥

My workflow has a color match future, you could try that.

Color match is on. Solved the issue (more or less) by going up or down a step if I hit the color flash. For whatever reason if it happens on 8/4, it wont happen on 10/5. Stepping up or down by one seemed to fix it.

You know, distilled models normally do not want such high step counts. 4/2 is standard, 6/3 is fine for some more detail ... everything over this will introduce problems.

prompting can also help, i've had success with "smoothly shift the color to match the last frame"

SVI or something like the color match node in my workflows should do, depending on the set strength.

@darksidewalker I see, Im also loading them using a regular model loader, could that be the issue? Or should I be using a checkpoint loader like in your wf?

@fierceys The loader does not really matter as long as the model is loaded.

Svi seems superior but doesn't work for every use case, so if you're manually extending, you can use prompts to guide the final frame.

an example: "her eyes are open and she's facing the camera in the final frame."

Other last frame prompts include 'clear face, fully in frame, high detail, visible penis, visible breasts', etc.. Use full sentences, but don't overload the instruction or it'll ignore most of it. Prompt specifically for what's being changed, and be concise!

you can also add NAG for negatives like jewelry, ponytails, etc..

if you're doing first to last frame, you can prompt "smoothly shift colors to match the last frame".

Thank you :) Glad you have so much fun with it!

Safetensors is always better, but if you are running into OOM, you may need a lower quant from gguf like Q6,Q5,Q4 ... But that also depends on your setup.

@darksidewalker I'm pretty sure that it will be not so necessary information for you, but still:
Intel(R) Core(TM) Ultra 9 275HX (2.70 GHz)
32GB Ram DDR5. On safetensors Im experiencing 5-6 mins generetion time for 5 sec video. That's why I'm searching for best model for my setup and asked advice for faster generation I2V type.

P.S.
Actually , I Tried Q6 & Q4 GGUF of yours, and it's works very very nice, I believe it takes about 4 mins on Q6 for same 5 sec video. <3

Hi!
this is not a model problem.
24 fps with just 5 sec are too short/fast to get a perfect transition. You need more seconds or reduce to 16 fps for perfect loops.

@darksidewalker Thank you for your response. Setting it to about 8 seconds at 24 fps works well as intended.

@lofean on v3.6 color match was broken, v3.7 has a fix, maybe that helps too

Maybe in the future, but not for the current version.

Make sure you did not mix high and low up.

@darksidewalker I'm pretty much sure, So I'm really confused right now.This is the first time updating ComfyUI, and it still hasn't solved the problem.

@darksidewalker I even checked the SHA256 , both the same

The reason why the generation itself doesn't work is not in safetensors, so there's no point in complaining about it not working.

@kenpechi Thanks for your nonsense. Maybe someone else have the same trouble and fixed it , he could tell me the answer

I tried the official workflow and it works, so it seems to be conflicting with KJ's.

@adminJP I do not use kijai loaders. Are you sure you using my latest WF and all dependencies are installed correctly?
The model itself does not do black-screens or pixelated videos.

I have the same problem and with gguf it's ok

@polish298 Update your comfy.

Since this isn't LORA, the prompts are free to let your imagination run wild and write freely.

You Can Refer This:https://civitai.com/articles/20293/darksidewalkers-wan-22-14b-i2v-usage-guide-definitive-edition

Make sure to install all dependencies.

I had this issue. A fresh install of white rabbit custom node solved my issue.

all dependencies are correctly installed... this was working fine a couple weeks ago. Even a fresh install did not do it for me

maybe is something due to being on amd and maybe some custom nodes updated... dunno

@counthareb And you are definitely not on an old version of the workflow?

And you did follow the instructions on the GitHub of the node and installed the 2 extra dependencies?

under the title there is a list of models, click on Low version and it will open a page with low version which you will download

@AlexMeowler thanks, i thought those were some tags related to this post 😅

The error comes from the unet loader because of mixed fp8 and is no problem, you can ignore that. It works just fine with svi 2, I did 8 batch videos with it. What ever you are facing is not a problem with the checkpoint.

@darksidewalker Awesome, thank you, will do my best to find a solution.
Can you please advise on where I should start looking into? The problem is that even at batch 1 resulting video is starting to become blurred at the end, at batch 2 end of the video has all details washed out and so on every batch until its just blurry mess of pixels.

@vladmight875 maybe you did not use high + low or mixed the up?

@darksidewalker No :( this part i did recheck many times :( high/low models + high/low sigmas exactly as in examples 4/2 steps also tried 6/3. schedulers wise euler_ancestral/bong_tangent does help a bit to reduce blur compared to euler/simple but still not much.

@darksidewalker I also notice that exposition with every batch is highly increased, in each batch after around 1st second of video its huge bump in picture brightness like al colors shifted towards white.

@vladmight875 So maybe you did not download and load the loras?

Thank you for the kind words!
If the motion is low for you this might be some setting you use, the motion is really fast if prompting and settings are correct!

Download from the page?

@darksidewalker sorry to ask where can i find these DasiwaWAN22I2V14B_synthseduction_q8High.guff

DasiwaWAN22I2V14B_synthseduction_q8low.guff

@darksidewalker im using your work flow wan2.2 i2v fast fidelity workflow. but slow motion bug from wan2.2 is still there.

but this work far better than wan2.2 is there any setting I can use to make the slow motion bug better?

@talestransmigration410 There is no slowmotion bug with this checkpoint, you may check your settings.
Or if you do not like like distilled models work, have a look at the TrueVision version.

@darksidewalker where is the TrueVision version.?

@talestransmigration410 just look on my page or my collection

@darksidewalker Thank you so much.

This ritual is more complex as known to the mere human. Stepping one close back will provide deeper insights. Beware the Ranfflix's.

Dude, you should think before writing such things. The path notation is correct if you are on Linux and changing the path to your files is always necessary.😏

im having the same problem how did you fix it. its not able to find my wanvae. i tried putting it in a folder named wan and also directly in the folder

@justicemanhero371 There is a folder inside "models" called "vae", put it there or do a folder inside and put it there, than select it inside the workflow. It is not that hard ... Maybe reload comfyui or press "R".

I just fixed. For anyone having this problem its because we are on windows and in linux the "/" are the other way around so you have to manually change it and find the vae dropdown. this workflow uses like a bookmark system so look at the top middle and click the bookmark symbol then click for the checkpoints section then manually change the vae there.

Poster boy for the PC Master Race

@darksidewalker ah that makes sense! ty!

Why don't you try and decide for yourself? Every model has other scopes. There is no "one fits all".

Why not try them all? Different tools for different scenes.

In my experience (very limited) remix tends to follow prompts a bit more liberally and morph characters into the ones it was trained on. Havent got to try v9 yet but DaSi's v8 is much more precise and follow almost to the letter. If you're trying to achieve convergence every time, DaSi works better. If you're more open to unexpected motion or off-prompt scenes Remix is fun.

@MiddlingMaker i'm able to get what i want so far, i just started AI video creation so very much a noob here. if you dont find answering my question. is wan not capable of camera movements? it never detects those prompts, maybe apart from zoom in and out

@NoScopeAi there are camera movement it can understand some basics, but the wording must be correct. There are guides on the Internet.

@NoScopeAi The syntax of the prompt controls the camera movements, especially on v9 I'm finding. If your prompts are structured into two or three "then this _____, description of this, more descriptors, more descriptors, then this _____, descriptors of this2, more descriptors of this2", make the final part of the prompt focus on what you want the camera to see and it will inform the movements.

Eg, if I a scene of someone eating and they start choking on their breakfast, and I want the close up to be their face my prompt might be something like "a man is sitting at a table, he has thick rimmed glasses and brown hair and a short beard [describes the initial image to the LLM], then he grabs the fork and begins to eat scrambled eggs [1st action, if the plate and eggs arent in shot the camera movement will very likely track out to reveal the eggs, or pan down to reveal the eggs], he chews thoughtfully, then [2nd action] he begin to choke on his food, [this will inform the camera movements as the LLM will try to make this happen by the end of the scene and the only logical progression would be through a camera movement] he panics and knocks the plate to the floor [this would focus the shot on the plate falling, likely panning away from the mans face BUT because I included that he's panicing the LLM will try and to include that which will want to pull the shot to the mans face]"

If I wanted just a close up on his face, my final part of the prompt would be "then he begins to choke while chewing (face centeric), his eyes bulge and start to water (face centeric), his cheeks redden (face centric), he shakes his head back and forth in panic (face centric). This has a much more likely chance of pulling the camera towards the face.

You can also give it explicit camera directions. In this prompt I'll also avoid he/she/they because the LLM can't really understand it. Referring to "the subject" seems to give better results especially with multiple subjects.

"a man is sitting at a table eating breakfast, the man has thick rimmed glasses and a short beard, then the scene focuses on the plate of eggs, three-quarter extreme close up, the eggs glisten and steam invitingly, shiny yellow yolks, clean white plate, glinting silver fork"

This prompt will very likely follow exact instructions, there's a clear camera movement, we've described the final scene as though the camera has already moved and we've only given it one thing to worry about, the camera movement to the breakfast. You CAN try to get subject motion, camera movement and details into a single shot but the more you give the LLM to work with, the less attention it can pay to any one thing - so details and identity start to get lost.

You also need to be careful with camera movements as they will invent the scene around the character as the camera spins. For a single shot, it works great but as soon as you move to the next shot the LLM has "forgotten" what the background is supposed to look like and will generate a similar but new one. If you had two shots where the camera rotated 180 in each, the first shot would present a very different background than the second shot, and it would be obvious when putting the shots together that the room morphs.

There aren't really and hard rules. Personally I try prompts and loras until you get close (render on really low resolution for quick previews), and adjust one thing at a time until you find the shot you want.

In my experience with dasiwa, it actually respects normal english like "... the camera zooms in slowly towards her face". I haven't quite been able to create more advanced camera movements like panning to the side or circle around a character. Maybe harder to pull off, or my prompting is not detailed enough.

@snobbias124 yes i keep experimenting as well, zoom in and out works decently, i just say "zoom out until character's feet are seen on the ground" and if i want to move location, i'll make the character walk towards some object in the image and tell it to "follow the character moving/walking" seems to work. yea but i cant do a complete 180 degree turn to camera angle, that would help me a lot

@NoScopeAi 180 is a looooooot for the model to make up in one scene, I do 90 degrees, then 90 degrees to grab an end frame and flf2v it

Thanks, but I'm not multiple guy's. Just me.

It starts to loop back as you do more frames over 81.

For me, I always just use the latest, but they are not all equally made, so some olders may provide some for spacial scenes.

They are all available here for redownload if needed.

@darksidewalker These are absolutely the best, V9 especially. The original Wan2.2 are terrible at following prompt. For example if I use "A giant paw appears and tears of clothes etc" in the vanilla model a paw appears that has no arm or the paw flies off the arm like a projectile. In Dasiwa model an actual paw with the arm that looks natural appears. Fantastic!! Not to mention the quality, no blurry crap, no issue. I replace most of the WF's default W2.2's with the DaSiWa ones. A+++

Excellent work!!!

@essence25 Thank you so much! I wish you all the fun crafting these art :)

Yes, baked in into lightspeed.

I would assume it is a prompting issue or the concept you want to achieve needs a lora.

This is because this is a fusion model, It may have been combined with“2d animation effect ”This lora will weaken character dynamics. So if you want to solve this problem.

1. You can use the official lightx2v high noise 1030.

2.You can add a wan2.1 high noise rank256 dynamic enhancement, intensity set to 3-4,

3. You can use dasiwa true vision, and rebuild lightx2v accelerated lora, and add high noise and low noise 2d animation effects intensity set to 0.5

@wjm171732988 Thank you very much for your detailed reply! Currently I'm using LoRa for strong movement tasks and the results have been quite satisfactory. From another perspective, this can also be regarded as a form of motion intensity control in a certain sense lol.

@meeseek42597 which LORAs are you using? I'm wondering about the right parameters, tried Dr34ml4y so far

@alexander13331593 smoothmix‘s adult lora

There is "TrueVision"

Update comfyui

"TrueVision" is without distillation. As mentioned in my announcement. If this is what you seek.

@darksidewalker - yes, that is exactly what I wanted. Amazing, thanks!

As a side note, I did SVI 2.0 pro with 8 sampler stages without degeneration of the video with my workflow.

This is just the headline, you did not install rgthree custom_nodes.

You know that this is not a problem of the workflow, rather than your implementation of runcomfy?
It can wherever be under VAE-folder, you just have to select it.

Even I update comfyui version, same result(2025-12-02 to 2026-01-29),anyone know how to solve this?

It only happens if your comfy is <0.4. ... you may not updated your comfy correctly. So check this.

@darksidewalker I check the version is v0.11.1,version ID is b0d9708,I change the model v7 can run my workflow,but v9 is still black result..my device is AMD R7 7700x + nvida RTX3090,maybe gpu is 30 series cause it?And i download your workflow, it show the wanvideoNag is error, but i update the the kjnodes to latest, still error. cloud you help me?

@andandeai0608 If you join my discord and post some screenshots, I could help maybe.
But if you get an error, you are missing dependencies or did not install it correctly.

I met the same problem. Did you finally solve it?

@xingxingxiao16254 I did create a comfyui-installer, it is a new release, but maybe helps to satisfy all dependencies. Maybe have a look? Any feedback would be appreciated!
https://github.com/darksidewalker/dasiwa-comfyui-installer

@xingxingxiao16254 my problem had been finished.Not the model proble, the node problem,my workflow use one node,after i update it all fine.

Maybe you try some kinks that is not working without Lora, because of tuned down bias. I bet that😏

@darksidewalker 
Hopefully my translator gets this right 🙂
------------------------------------------------
This is possible — I’m continuing to test the model.
Version 9 has excellent camera movement, but in every other aspect, out of 10 generations, all 10 were produced by 8.1, while version 9 completely failed all prompts.

From the NSFW v9 examples: the girl can’t simply spread her legs or take off her clothes without them tearing, whereas in version 8.1 this looks more than fine.
(Compared directly with 8.1 using the same prompt and the same seed; different workflows were tested.)

The issue is definitely on my side, since judging by the comments everything works fine for others. Now I just need to figure out what exactly I’m doing wrong.

I’ll continue testing!

Thanks for your work!

@nathanight85851 un-dessing / nudifying of real ppl is against ToS of civitai, so the checkpoint is not trained or optimized for that. If it works it is just coincidence. So there is no possibility in support for this.

The checkpoint is not for NSFW only, so I lowered the bias of explicit content to be more compatible and usable for a more variety of things.

You try to use the model for one specific case and judge just on this single case. V9 is superior to v8 on overall use case's.

It is not Our pleasure. Withdraw thy suit, for We shall not speak on’t again.

I don't know if it's a bug or not, but I've noticed that the camera tends to naturally move to the left and the subject tends to move to the right edge.

What do you mean by long generations?

WAN 2.2 is made for ~5s generations. Anything above tend to loop back or produce unwanted motions, normally. For long videos you do multiple 5s-7s videos and combine them or you may use SVI 2.0 pro.

@darksidewalker Yup, I'm using SVI with your workflow to make multiple sections into a long generation. It works great, but even in single 5s generations the camera tends to move to the left. The more sections I add, the camera keeps moving further to the left.

@Firo901 No clue, my camera never moves in SVI, except I prompt it. Even with 8 sample sections. Did you make sure to use the recommended sampler/scheduler?

@darksidewalker Yeah, I'm using the recommended settings that you provide on your workflow. I will keep testing, thank you.

@Firo901 Maybe the resolution? Do you use the SVI standard resolution?

@Firo901 Try removing 'the' from in front of 'camera.' I've played around with Wan 2.2 a lot, and whenever I use a prefix before 'camera,' it doesn't seem to work as well. Maybe it's just me, but that's what I've noticed.

@darksidewalker Yeah, I'm using option #3 in your workflow which is 480p.

@g1263495582 Thank you for the tip, I'll do more testing. 

All these merges usually do is add in the porn stuff, so you can only do whatever else base WAN knows, nothing more

@Xeno443 Well, you may think so, but my merges are a bit different. Just read the announcement and descriptions. There is more at it than just lazy merges for nsfw.
One thing is right, arbitrary spacial moves that are not trained are hard to achieve ... but that's a totally different thing.

@darksidewalker 
my apologies, you are right and I was not clear, I meant this mainly for the question about movements, scenes, concepts that you can do with any given wan merge. while there are some non-porn loras that add those to wan, the majority seem to be about nsfw concepts, so if base wan does not understand "kicking a ball" its unlikely any of the current merges does.

@Xeno443 No problem, mostly you are right there are many lazy merges which just throws some xxx loras in and call it super-multi-uncensored-whatever checkpoint. I just wanted to say that I really care about what I'm doing :)
Yeah such conceptions like "kicking a ball" is unlikely trained or merged. It is just too niche.

Thank you for the kind words 😊

Well, I did notice that with no prompting, but with prompts it stays for me.
Why v9 decides to do this is not clear to me, but unfortunately I can not patch SVI, that or finetune the model for SVI. The model was made before SVI was a thing.
Maybe for the next release I'll look into it.

Yeah, with SVI the subject’s face changes drastically, but I found a trick: using the official model with the Lightning LoRA just for low noise. That fixed about 95% of the face/subject consistency. Sometimes it still recreates the teeth, but overall it works, hopefully next releases fix it.

@rangarajanvinnan537 this is a complete other thing and happens mostly if your res is ultra low or you do not prompt for the characters appearance

@darksidewalker it only seems to happen with other workflows. Using the main one and it seems to work fine, might just be a quirk with how the other two are set up.

With the official workflow it’s amazing once I got everything dialed in

I'm not having this issue. It's important to know that if you're chaining multiple videos together, the next set takes an image from the previous gen and continues from that. If the face is obscured by a lot of stuff it'll have to guess what the person looks like. This can absolutely trash appearance consistency. That's why it happens randomly a lot of the time.

@LetTheBassDrop I know this was a response to the other poster, but I only noticed inconsistent faces when I either messed around with the sampler outside of Euler/Res_multistep or was doing lower quality previews. Once I jumped up to the higher resolutions or upped it to 6 steps instead of 4 I rarely have had a problem with consistency.

Sure

No it does not, your settings or wf is broken

Sounds like you don't have the correct VAE set. If you're using someone else's workflow you most likely need to manually set the VAE as it's probably pointing to the wrong directory.

This is not a workflow it is a checkpoint.

This is not a workflow it is a checkpoint. You may post the problem to the workflow you face the issue, that would be better.

I encountered the same problem, which occurred after enabling upscale, using the workflow in the model example.@darksidewalker 

You have no clue what are you talking about

Try using a lora for the specific type of situation/motion on low strength.

@darksidewalker thank you and thank you for quick reply

Yes.
https://civitai.com/models/2272580/dasiwa-wan-22-i2v-14b-or-truevision-or-safetensors

@darksidewalker Thank you very much!

Yes, it's very good at that. The character remains the same. If the only thing that matters is that the characters remain consistent, I can recommend this model.

Hi!

I investigate this, but unlikely for "SynthSeduction". More likely for the next release.

The reason is, I would have to build the source anew and nvfp4 is without fp16 base, like de-quanted from this base, worse than gguf q4.
This would result is extreme low quality, poor adherence, trash temporal consistency and bad visual fidelity.
Also, only Nvidia 50xx users would even have a benefit in speed, but non the less the poor quality.

@darksidewalker  okay i wasnt aware of the significant difference in quality. Probably best to just stick with Fp8 then. I tried Flux2 in nvfp4 und only noticed the speed boost.

@AmC13 For image creation it is like 85-90% of fp8, but for video it is more like q4-5 gguf. When done from a fp8 base even worse.

Thank you very much!😸

Yes, v9 completely ignores the camera prompt. This means this models is a failure.

If you don't prompt any facial details/expression it will guess probably some, if you mean that by "yapping".

@darksidewalker Thx for responding so quickly, nah my characters are having like whole on conversations with each other

No clue, must be a prompting thing, my characters are not chatty if I prompt their motions properly.

@darksidewalker Lmao alr thx

@Toufh its hard to fix, you have be super detailed and run a few gens, I ran like 10-15 with different prompts before I could get it right, even putting talk, speak, chat, in the negative prompt didnt work. even then sometimes you just get lucky and they dont talk

@johhnnymann1 negative don't work with cfg 1, also I have no clue why this is considered a thing, I write like 2 till 5 sentences and they never talk.

@darksidewalker ahh makes sense then, I've only been messing around with comfy for a week so I have large knowledge gaps. Maybe it was a lora that was causing the talking, I dont remember what I was running with the model where I was getting the talking

@johhnnymann1 I'm only using the base model without any loras, I think it's just you need to explicitly provide the expressions

You are confusing some things here. I suggest you reading my guide, if you are totally new.
https://civitai.com/articles/20293/darksidewalkers-wan-22-14b-i2v-usage-guide-definitive-edition

The other things are clearly written inside the workflow notes, where to put the models, also with a chart. Just follow the description.

If you really put the files correctly, you can select them in the gui of the workflow, after restarting comfy or pressing refresh "r" -button.

@darksidewalker so.."This is a WAN 2.2 Model: You will need one pair of High + Low." if I use tastysin, should I put this high and low model in diffusion_models folder or checkpoints folder? I tried to use the wan workflow from the template and it's downloading high and low noise as diffusion model, what's high and low noise? are these models also high and low noise?

nevermind, it seems it only detects files that I downloaded directly from comfyUI, the vae and text encoders already shows up, if I put it manually into the folder like these high and low checkpoints models, it won't be detected. I'm using desktop version of comfyUI. it seems to detect if I use the one from github

@lizardman72119 well, if that happens you did not place the files correctly, it has nothing to do from what sources you load them.

FP8 may be too much for 8GB VRAM, you need a large swap/pagefile to offload or use lower quants like q5/6.
Beware, large offloading to SSD will degrade your SSD over time.

@darksidewalker  wow that was fast man, thanks for your response! I have got the 12GB variant :) I wondered if I scrapped somehow the workflow, I noticed that if I create Images (other Workflow) on low resolution its worse than using High resolutions. Thanks for the help, im tinkering for days but at least im learning fast that way. 957MiB / 12288MiB just checked through terminal just to be sure

edit: python main.py --listen 0.0.0.0 --port 8188 --lowvram --disable-xformers --reserve-vram 8.0 --disable-cuda-malloc --use-sage-attention --disable-smart-memory --cache-none thats my start prompt, maybe im overusing

@civitaiscam182 --reserve-vram 8 will reserve 8GB from 12GB for your system ... so comfy only will use like 4GB for generation. Lower to --reserve-vram 1

@darksidewalker forgot to mention I already tried reserving like 4gb but it was still OOM or even Desktop freezing. I tried your idea and it actually worked to the point where it wanted to decode, it killed my Terminal weirdly, like a crash.

@civitaiscam182 This happen because your system RAM getting filled completely, freezing your system

@darksidewalker alright , I stepped down a bit with the workflow, I wanted to use the AiO version and went down to the DaSiWa WAN 2.2 i2v FastFidelity C-I2V-17. That seems to work instantly without crashing. I guess the rest is User being noob at prompting. Thanks for taking your time, I guess once im not completly noob anymore I will try again :)

Hey, side note just for who is interested, thats the BEST Creator among the Platform, hands down, learned so much since the last post and you are great, full SFW usage and NSFW, even can use it for epic bossbattles

Thank you!

I see your account is really new and you did not contribution on the page. So... I really don't know what the intent of this question could be to me. My workflows I use are all available.

Glad you like it, have fun with it!

@darksidewalker I'm sorry, but I have a question. How would I attach a camera node like WanCameraEmbedding to the workflow? I'd like to force a static camera in my videos.

@goonie888 I don't know this node and never used it, so I may not provide any detailed help here. But it seams to replace the normal "WanImageToVideo" node.

NAG for distilled models with cfg 1, or a non distilled model with cfg 3.5.

@darksidewalker Thx for reply.

It is not, except some rogue uploaders who sometimes change my quants making them worse.

Update your comfyui, black output means old comfyui install.
If in need, try my installer.

Can only be wrong settings with wrong usage of checkpoint or missing loras, what I can think of

@darksidewalker I've changed to your workflow and the problem is gone, thank you. If i have your attention here ... can you kindly tell me what am I doing wrong that with SVI (or WAN in general) model tends to return back to the start frame. I prompt to remove clothing, she removes and then puts it back. Same happens with other actions. As if it's drawn to this starting frame with a huge magnet.

@mjanek20 Looks like you used more than 7 seconds, after 5s it tends to loop back.

If you have ghosting, double check that you did not mixed up high and low or forgot to use the loras

@darksidewalker holy shit, you´re right. I must have been tired and confused it. Works really really great. I tried so many workflows for long videos and this one is the most promising so far.

@SensUnique Have fun! Glad this helped

Sorry, I have no clue what you mean, this is the model page. Maybe you are posting on the wrong page and want to know about a workflow? Please refer to your workflow page than.

Hi! 👋 Fp8 safetensors is too much for your specs. If you want detailed information about WAN 2.2 have a look at my guide. Otherwise maybe try a gguf q4-q6 and try if it will run.

It's running freely on 8gb vram and 16gb ram but only used high ver on custom simple wan workflow.

@deadpolo9090976 no way, you ether did very low settings or extreme swap/offload to SSD.

Thank you for these lovely words :)

Glad you enjoy it! And thank you for the friendly words 😊

@darksidewalker Amazing work, DaSiWa.

rife is implemented, but you must manually add the file for rife, since all rife custom node dl's are broken.

I'm not sure if you've resolved this yet, but there is a "Color Match" node that can be used to align colors—you might want to give it a try.

@futa2088 yes, using wan vace + hm-mvgd-hm method in colour match is the best way I know to make a seamless continuous vid.

Ignore this.

I get that too then the sampler says the tensor values are incorrect. (48) vs (16). Feels like something is glitched here.

@aaronisl33t406 that's a totally different thing. Wrong clip or VAE used.

https://civitai.com/models/2272580/dasiwa-wan-22-i2v-14b-or-truevision-or-safetensors

@darksidewalker So the name in the latest AiO WF that I mentioned above is actually intentional and referrs to the one you linked? Just to clarify, since even the one linked is still named differently. (assuming you renamed it on your end)

@essence25 I don't know what you mean, there are multiple versions on my collection. Distilled and not distilled and gguf (distilled).

@darksidewalker OK let me clarify. In the C-AiO workflow under the BACKEND main node I2V_High/Low_safetensor the default model is called "DasiwaWAN22I2V14B_synthseductionLowV9.safetensors" notice the word "Lightspeed" is missing from the name. This default name refers to which checkpoint? because that does not exist anywhere. I'm just curious why the default checkpoint does not have "Lightspeed" or "Truevision" in its name. And what is the recommended checkpoint for the C-AiO workflow?
The Lightspeed or the Truevision?

Here main question to clarify the default checkpoint for AiO workflow:

DasiwaWAN22I2V14B_synthseductionLow/HighV9.safetensors = DasiwaWAN22I2V14BLightspeed_synthseductionLow/HighV9.safetensors

or is it:

DasiwaWAN22I2V14B_synthseductionLow/HighV9.safetensors = DasiwaWAN22I2V14BTruevision_synthseductionLow/HighV9.safetensors

@essence25 it's just my naming on my SSD. It is the lightspeed version.

Your comfyui is absolutely outdated if you get a blackscreen. It is mentioned in the description tho.

I am keeping my comfyui outdated on purpose which is why it's been flawless for the last 6 months, without a single minute wasted on fixing broken dependencies etc etc. I'd rather take a bullet in the leg than trying / building a new workflow from scratch. So back to my question, is there any specific element/dependency that makes your checkpoint go blackscreen within my workflow (and not the other checkpoints) ? I just need a clue. But theres no way in hell I'm going to randomly update everything on somebody's hunch.

The frontend ui does not matter, prior to base comfyui 0.4.0 there is no fp8 mixed precision processing, resulting in black screen.

@darksidewalker I mentioned the UI as its an easy thing to point out as being different between the versions that will work and the versions that won't work. This makes it much simpler for people to understand what the issue is when they run into this problem. Thank you for confirming.

@darksidewalker tbh, there was certain version of comfyui where my gguf models went from generating videos in 5 mins to generating in 45 min, certain comfyui version just kept crashing, I had to reinstall the very first version of comfyui I ever used

@ItheriaSK I think there is no reason to be on old versions because there's a version with a bug that maybe fixed in the next. It is the same with drivers, if your driver is not supported you can't run the game. Comfy and models is the same, if you run a version that can not understand a model type, you can not run it. Updates are mandatory at some point.

🤷 I think this is a prompting issue. Sometimes heard this from others, but did not face this much myself if prompts are good

I've tried a lot of promts, but for some reason, some videos actually turn out to be like I wrote earlier, with a right-leaning effect, which is very strange XD

It's a bit disappointing, because I really like the model in terms of creativity, but this defect really bothers me.

@daminik124124519 maybe it's a setting like the sampler+ scheduler?

@darksidewalker I checked everything, but the result is the same. I used three of your models for testing, and here's what I can say:

1) https://civitai.com/models/2190659/dasiwa-wan-22-i2v-14b-tastysin-v8-or-lightspeed-or-gguf This model turned out to be the best in terms of stability; the camera doesn't move to the right at all, so I use it.

2) https://civitai.com/models/2269796/dasiwa-wan-22-i2v-14b-synthseduction-v9-or-lightspeed-or-gguf Here, the camera moves, as I wrote earlier about the shift.

3) https://civitai.com/models/1981116 Exactly the same problem as in the second case.

No. Nothing.

@darksidewalker Same question. With 'nothing' you mean you prefer the default wan 2.2 T2V model or you prefer to use any T2I flow and continue with your model in I2V?

What would you recommend for T2V -> Nothing, since I never use T2V.
I always do/prefer T2I -> I2V

@darksidewalker Interestingly, during a workflow test, I accidentally selected "synthseduction." It successfully generated a video—and while it was slightly rougher than what a dedicated T2V model would produce, it nonetheless followed the prompts to generate the correct output.

@darksidewalker As a side note, I originally intended to purchase the previews for TrueVision V10 and BoundBite today, but I don't currently own any cryptocurrency. Have you considered releasing your models on other fan-based platforms?

@futa2088 Yeah, should be somehow, since it inherits some features to enable the t2v understanding. But it is not a full featured t2v model.

@futa2088 No, I do not release on other platforms atm.

Gazing till they gaze back! 😶

dasiwa is right but if that just refuses to work....
aggressive prompting usually works. "her expression does not change, she is {pouting/grimacing/etc}, her expression remains static, her facial expression stays the same" as well as using negative prompting with >1 CFG or video nag node.

@Nk0GGry Or your input image has no smile, normally the model copies the expression.

The easiest way is to use the First/Last Frame feature (FLF2V). Simply feed your starting image into the last frame slot. This forces the animation to return to the beginning, making it loop perfectly.

To be clear, flf2v will not loop seamless/perfectly. There are tricks like in my workflow that work better as just flf2v, but they also will not work all the time. It depends on the input and frame timing.

I'm also intrested in perfect loop.I'm using your workflow but the loop is not really perfect.What should I set my fps for 5 seconds of video ? I'm doing 1920*1088 whats the trick ? Do you guys know how to loop perfect or near perfect ?

Camera is controlled through prompting like on every i2v model.

I'm having good results using "...while a static camera is recording the scene" after the first sentence. Ex "A girl is standing while a static camera is recording the scene".

There is no hidden technique, most of my videos has metadata with all settings.

Install all dependencies, delete not needed custom nodes, update comfyui, or use my installer.
If the switches don't work, you installation is not correctly setup.

@darksidewalker The problem was the 2.0 nodes; I hadn't read your message telling me to disable them. Thanks!

@darksidewalkerHey, man, your workflow is amazing, and I'm using it a lot. There's just one thing missing to make it perfect: a camera control for WAN. Is there any chance you could add that in a future update? Thanks anyway for this great work.

@tob11 There is no camera control for WAN 2.2, camera is controlled through prompting on i2v models or some spacial loras.

@darksidewalker Oh, I've seen several really promising workflows for controlling the camera. The truth is, I read the WAN 2.2 documentation on the camera, but it never works right for me.

@tob11 I never seen one working as intended, technically there is no extra camera control possible without training.

@darksidewalker I have a question, friend. I'm an animator and I use your workflow to create my own animations, but I can't seem to achieve the same level of sharpness as the reference image, even when I set it to 1.15MP (I have an RTX 5080). I see in your examples and those of others that theirs look great. Is there a setting I need to adjust for better quality? Maybe more steps? cfg? I'm a bit of a newbie.

@tob11 Not that I'm aware of. I just use 4 steps and one of the upscaling features, sometimes interpolation. Most of the time even 0.6 MP. Cfg is 1.
Nothing spacial.

@darksidewalker And with the 4-step process, I guess increasing the step size or setting a higher CFG won't work, right?

@tob11 6 total steps are good for some details, till 8 is possible, everything else is placebo. cfg 1 is a requirement for distillated models to work as intended.

You will have to fix your comfyui installation or custom nodes. Or make a clean install with my installer.
https://civitai.com/models/2364056/tool-dasiwa-comfyui-installer

@darksidewalker can it be run on rtx 4070 8gb vram and 16gb ram

I have the same problem, and no matter how many times I do a clean install, the same problem keeps happening :(