>How
Download the package and extract it on some fast drive that have good amount free space (100gb at least). Then
----Run 1stAIdBag_install.bat
--------Install ComfyUI and Manager. This is kinda bare minimum you need to install. You can install workflow's missing node via manager. Some user inputs needed during this.
--------Check if your ComfyUI runs after this. If it does....proceed
--------Install Custom nodes. This will install all custom nodes needed in my workflows, also other workflow's commonly used nodes (this might take some time)
--------Install Attentions. Highly recommend this as they give speed boost for generations.
--------Download essential WAN files and one of the WAN GGUF models (Q8, Q5 or Q2) Better your rig, bigger Q to get
------------NOTE: Downloading these models take time as those are several gigs
--------OPTIONAL: Download QWEN, GIT, or SDXL stuff if wanna play with image workflows

Run comfyui. On the first run it will take some time to load as some nodes like reactor will download some models needed. Once loaded, open the latest version of the workflow, re-select the all the models, vae, clip and loras. This is something you need to do every time you open up a new workflow, so the files point into locations in your computer. See the vid: https://screenrec.com/share/mkgZDHG5uI

After that, check other settings, type in prompt and hit run.

Buy me a beer: https://ko-fi.com/1staidbag

>Already got old install?
----Backup (cut) models and user folders somewhere
----del old ComfyUI_windows_portable folder. Then do as guided above
----paste model and user folders back into your fresh ComfyUI folder after install

>What do i need?
-Windows, RTX card, pref at least 8GB of VRAM, 32GB of RAM, drivers up2date
-You can run stuff with lower rig specs but then you would need lower models, lower resolution, lower clip duration, etc
-There are some info/help to run stuff with ATI cards in prev threads
-NOTE1: Major change on the pack, nothing is pre-installed, all nodes are getting installed via 1stAIdbag.bat now. Pack is now much smaller, 7megs

>If/when you need help....
A) take a screenshot of your whole workflow (and log console window if possible)
B) state what workflow are you using (i'm not going to try fix other's workflows for you)
C) what specs on your rig

>Where to get loras?
https://civitai.com/search/models?baseModel=Wan%20Video&baseModel=Wan%20Video%202.2%20I2V-A14B&baseModel=Wan%20Video%202.2%20T2V-A14B&baseModel=Wan%20Video%2014B%20t2v&modelType=LORA&sortBy=models_v9%3AcreatedAt%3Adesc
(you need to make account and enable show adult content)
https://civarchive.com/search?type=LORA&base_model=Wan+Video+14B+i2v+480p&is_nsfw=true&sort=newest&page=2
(old 2.1 loras that got removed from civitai due to their nsfw bullshit policy)

Old thread(s)
vol1 https://archived.moe/r/thread/19910536/
vol2 https://archived.moe/r/thread/19938024/
vol3 https://archived.moe/r/thread/19987607/
vol4 https://archived.moe/r/thread/20011173/
vol5 https://archived.moe/r/thread/20035627/
vol6 https://archived.moe/r/thread/20053516/
vol7 https://archived.moe/r/thread/20096113/
vol8 https://archived.moe/r/thread/20137696/
vol9 https://archived.moe/r/thread/20195961/
vol10 https://archived.moe/r/thread/20238038/
vol11 https://archived.moe/r/thread/20264942/

SVI workflows
https://files.catbox.moe/hns335.json (1.0)
https://files.catbox.moe/kct5jp.json (1.1)

Needs svi 2.0 pro r128 high and low loras:
https://huggingface.co/Kijai/WanVideo_comfy/tree/main/LoRAs/Stable-Video-Infinity/v2.0

and update kjnodes

...still kinda wip thing but you gets you started

File: 123731_157573958331029965(...).mp4 (2.78 MB, 480x640)

2.78 MB MP4

>>20332217
Can you recommend an alternative to "Assume the position" I look for something that does a similar job, but WITHOUT her putting her hair behind her back.
She also doesn't need to put her hands on her back.

Basically, just getting on her knees and opening the mouth / sticking out the tongue.


Reason:
I find pulling her hair back often destroys the hair style and general look of the woman from the initial screen.
This is an example of what I mean.

(saying in the prompt that she doesn't pull back her hair doesn't work, it's too strong in the lora and always comes through.)

>>20333257
She also doesn't need to put her hands on her HEAD*

File: 004308_157573958331029965(...).mp4 (509 KB, 720x400)

509 KB MP4

>>20333257
well how about just? no loras needed

"Woman starts kneeling down on the ground. she looks at the viewer, fondles her breasts, opens her mouth and puts her tongue out."

File: 233504_157573958331029965(...).mp4 (2.38 MB, 640x432)

2.38 MB MP4

Second question:

I have issues with the dick coming out of the man's butt instead of crotch about 1/3 of the time.

And tips on how to lower the chance of this happening?
(posting the workflow in a reply to this)

File: dickbutt.png (651 KB, 3540x1423)

651 KB PNG

>>20333460

Thanks for the tip! I will try it with no loras!

Here is the workflow for the dick-butt situation

File: 1771368783068260.jpg (57 KB, 640x432)

57 KB JPG

>>20333465
lets say extend1 starts with this frame....I would guess model has hard time which side of that man is the crotch, which is ass

BUT let's say it would do what you prompted; you would mostly see only that man's ass, right? You cannot see "woman looking at the viewer" when there is a nude man between her and the camera shoving dick in her mouth.

Try to define where the man is coming from. Example "Man with black shirt and blue jeans enters the frame from the right side. Woman turns facing the man"

Also do not use words like NEVER, NOT, or other negative def words on positive prompting as wan doesn't really understand them....So "woman never stands up" get most likely interpreted as "woman stands up".

prompting is tricky when using lightx loras because cfg must be at 1.0...this means that model rely more on what it knows to do, rather than what you want it to do. When using loras, lets say oral_insertion, that gets loaded into the model and when you type the trigger words "suck penis" in your prompt model knows what to do, even with low cfg value. In other word there no point writing complex prompts when using low cfg values as wan won't follow it

Also, like i've said in many threads before. there are no golden prompts. One prompt work for some pic, not for other. There are so many variables what makes your gen....image, resolution, random seed, cfg values, steps, duration.....prompt is one of those variable

you test and try

>>20333597
i still do not know the difference/how to know what is a lora and what is a lightx lora, i am missing something simple lol

>>20333653
consider lora as mini-model that has been designed to do one specific thing, that main model can't do, or does it poorly. So it's like an addon pack to the main game.

LightX is lora too....but rather than adding something new to the main game (model), it makes it run faster but taking some toll on quality (in game terms, it tweaks the options so you get good quality in best possible speed)

>>20333884
but how do i know which is which when on civtai etc? does lightx only go in the lora loader part in comfyui, or does it go into the high/low noise lora part?

File: 2026-02-18 050239.png (271 KB, 1772x439)

271 KB PNG

>>20334049
lightx loras...like most of wan 2.2 loras have high and low parts. They go into high and low lora loaders like other loras

In my workflows i've made separate loaders for lightx loras as they are things you always want to keep on (aka separate them from rest of the loras so you don't accidentally disable them)

on civitai lora page, different parts (high and low) are usually listed like this, click high, click download, save as in lora folder......click low part, do the same

File: Screenshot 2026-02-17 195240.png (73 KB, 662x468)

73 KB PNG

>>20334075
ok, i am pretty sure i understand that. you would put the froggy lora as an example into the parts i circled in orange right?

my confusion is what else, other than what you have in as default (the 4step.safetensors) would you put in the red circles?

you wouldn't put froggy, or oral insertion etc, into the red circles right? so what WOULD you hypothetically put there?

>>20334207
those red circles are just and only for LightX (4step) loras yes....i could've titled those better. Select them once and then forget the nodes in red circles

everything else goes into orange parts, oral_insertion, forry, pov_missionary etc etc

>>20334289
you're good, that is what i thought

so where does something like these go?

https://civitai.com/models/1995784/smooth-mix-wan-22-14b-i2vt2v?modelVersionId=2513182

https://civitai.com/models/1981116/dasiwa-wan-22-i2v-14b-or-lightspeed-or-safetensors?modelVersionId=2555640

>>20334346
those modded version of base wan model. those go into diffusion_models folder OR if you download GGUF version of them they go to unet folder (thou i think latest version of gguf nodes can find them in diffusion_models folder too)

Note that those have lightx loras "baked in" them....this means DON'T load your own lightx loras into it

I can see your enthusiasm.....but don't try to learn everything in a day (or week or even month). Learn the basics, what different lingo means, etc

>>20334383
roger that. yeah i was not planning on using those yet, just wanted to know what they even were lol

>>20332217

What would I have to do if I don't want the workflow embedded as metadata in the video file?

Thanks in advance

>>20334889
Nevermind! Found the button!

ComfyUI ran the first time before nodes attentions and WAN, but now I get an error when starting:

ImportError: cannot import name 'is_offline_mode' from 'huggingface_hub' (D:\Comfy\ComfyUI_apprentice_portable\python_embeded\Lib\site-packages\huggingface_hub\__init__.py)
[W218 22:07:41.000000000 AllocatorConfig.cpp:28] Warning: PYTORCH_CUDA_ALLOC_CONF is deprecated, use PYTORCH_ALLOC_CONF instead (function operator ())

Any tips?

Google said to update huggingface_hub via pip with:
pip install --upgrade huggingface-hub

>>20335919

The issue only happens when I install the attentions. If I go through the whole installation without the attentions, it launches fine

back to top with ye

File: asd.png (23 KB, 979x150)

23 KB PNG

I'm running an RTX2070 (non-super) with 32gigs of RAM.
Should I be having issues running Q5?

I can run Q2 and get some very hallucination-y results, but Q5 just fails with no error once it goes the low noise step.

Not sure if a problem with me locally but if anyone is using RIFE VFI I'm getting 404's all day when it tries to download .pth files.

You can manually patch this is you are handy with python. https://huggingface.co/jasonot/mycomfyui/blob/main/rife47.pth is a link to 47, and you'll need to patch vfi_utils.py -> load_file_from_url. You can hardcode the url there if you only need 47.

File: 157981921180745268_00001_.png (176 KB, 272x480)

176 KB PNG

I'm getting some incredibly hallucinated and blurry outputs. Doesn't matter which lora I try

Is it because I'm using Q2? Q5 just closes without an error for me.

I'll post my workflow after this

File: 157981921180745268_00001.mp4 (724 KB, 272x480)

724 KB MP4

>>20337247

Oops, got the last frame instead of the video

File: asd3.png (480 KB, 1828x997)

480 KB PNG

>>20337248

This is the workflow for this

>>20337253
I don't use bags workflow but looks like you don't have lightx2 loras added

Anybody know of any vore loras for i2v/t2v or qwen/sdxl? I don't mean the gory realistic kind, but the swallowed whole fetish thingy.

>>20337253
Like anon said... You don't have lightx loras loaded. You need them order to make gens with low step count.

Load lightx loras where your blink ones are now, and load blink (and other loras if you need) in loaders below

Is it possible to retain the face/likeness of the woman through several extends?

I once saw another workflow that promised to achive this with the "ping pong" methode.

I tried activating the "ping pong" button in your workflow (rifle version), but all it did was, that at the end of the video, the same video would play again, vut in reverse.

File: asdasd.mp4 (522 KB, 272x480)

522 KB MP4

>>20337585

Thanks for that, I knew I had to be missing something obvious. I thought I read on the previous thread that conflicting loras were a thing.

I'm still getting something really blurry. I don't know if that's just Q2 or not.

But whenever I try to run Q5, whenever I reach stage 3 for the low model, the browser windows blinks grey for a second, and then there's just a disconnected message. And on the cmd, i just get a message saying press any button to continue and no errors every time.

Any way to troubleshoot this?

File: asd3.png (442 KB, 1796x916)

442 KB PNG

>>20338011

This is the workflow for this video

>>20338011
That could be just that Q2 is shit.... I think it's only good for anime or stuff like that.... What specs on your rig? Vram ram situation?

I also notices you got line missing that connects into tripleksamplers base_steps...it says -1.... Try set it to 0

>>20338046

I don't think I changed anything on the workflow, I just opened it and selected the loras again.

RTX2070 8GB VRAM
32GB DDR5 RAM

It's not great, but I was hoping 8GB would be alright for passable quality, just slower.
Any way to make bette results even if it takes longer?

>>20337955
Yes, that what ping pong does.... Play and reverse back

I think what they meant was to use first frame and last frame. Use same image on both.... Make the length bit longer, few sec. Prompt it like usual and run

What this does; it guides the Wan to do your prompt BUT whatever it does, scene should end with end frame you have chosen. This also prevents the face shift in someway because Wan knows how the face should look at the end

Reason for those extra sec is so you can cut the video at the point where it seems wanting to go back.....on some cases this is not needed, for example if you animate a blowjob image

>>20338075
Main models you use are main factor on quality.... Sure increasing resolution and stepcount help with quality with cost of time... But then again you would most likely run out of mem

Have you tried using a Gen without any loras? If you remove lightx loras, set steps to 20 or higher

>>20338091

I've been troubleshooting a bit more and I can use Q5 or even Q8 as the high noise and it's fine. It also improves the results a bit.
But if I try to use Q5 or Q8 as the low noise, the cmd just stops responding after it says "requested to load wan21" and it fails without giving me an error, and I get a red banner saying reconnecting.
There's no message on the cmd or the browser window and I have to manually close both.

>>20338516
sounds like OOM issue to me.....
you could try;

a) add --lowvram flag into comfy launch bat (edit the bat file with notepad)
b) increase system pagefile size to 100+gb
c) load clip into cpu instead your graphics card (cuda)

start with a, move forward if no help. Also is your comfy which version? (you can see it from the manager)

>>20338538

I'll give those suggestions a try. Thanks a lot for your help, you've been brilliant.

I've used Q2 for high noise, and Q5 for low noise and just got some really great results compared to before, night and day!
I'll give those three things a try tomorrow, I've definitely got something to work with now with this Q2 Q5 setup

I'll have some more time in the weekend to start learning the ropes a bit more and try to understand more the details.

My comfyUI version is 0.14.1, and I'm running the v1.5d workflow.

>>20338607
i guess you could mix match those.....as high models (and loras) generally contains just the movement and low ones has the details. Leave your disc if you have one and i can help ya on weekend, so we don't spam this thread,

>>20338091
How do you do it without lightx, just bypass the loraloadermodelonly?

Also, I've been using 12 steps with the lightx Lora you have in there by default. Does that accomplish anything or should I just have 4 steps?

>>20338707
that should work....or set lightx lora weights to "0". 4steps is the min lightx needs....this is the fastest. In my runs and test, 8-12steps tends to give better quality and movement. Set the seed to fixed and run same image with same setting with 4. 8. 12 steps, you'll be the judge if those extra steps are worth the extra time

When running gen without lightx....steps shoud be around 20-30 and cfg around 2-5

>>20338747
Is that base cfg or lightning cfg? I've had base at 3.6 and lightning at 1 this whole time

>>20338762
lightning cfg.......but to be honest i don't know how it works in tripleksampler, never tested it...it claims to set it automatically to "1" on some scenarios

File: Screenshot_19-2-2026_1950(...).jpg (868 KB, 1981x1066)

868 KB JPG

Workflow. using wan 2.2. using the q5 smoothmix.
Igoon blink handjob + light loras. I interchange the igoon lora with whatever lora i want to use, keep the light lora and settings the same.
result in next post.

File: ComfyUI_00229_.png (485 KB, 448x800)

485 KB PNG

>>20338848
results are solid. keeping the face identical comes down to the quality of the photo it seems. it seems to like real photos a bit better than AI ones, and good lighting etc. U can also up the "output longest side" to match your photo and get better quality, but it will take longer to gen and have minimal improvement sometimes.

File: 157128488479040731_00002.mp4 (1.32 MB, 448x800)

1.32 MB MP4

>>20338852
whoops. heres vid

File: 157679474435941022_00001.mp4 (476 KB, 368x640)

476 KB MP4

is there an igoon "blink to doggystyle" or something of that nature where the girls face is still visible?
trying it with different loras and prompts with poor results.

>>20338888
Yes, but when I've tried it, the face looks nothing like the original, but I might be doing something wrong

https://civarchive.com/models/2232764?modelVersionId=2513563

>>20338888
What did you use to make that video?

File: 157106151461466796_00001.mp4 (533 KB, 368x640)

533 KB MP4

>>20338907
just the light loras + blink handjob but turned the handjob strength to 0.35 for the high and low.
im getting ok results but cannot get a man to show up without it being really deformed

>>20338911
Maybe this or another from this guy

https://civitai.com/models/1855263/pov-insertion-wan-2x

File: 157217397915682923_00001.mp4 (608 KB, 368x640)

608 KB MP4

>>20338907
>>20338911
oh i also changed the prompt to "The video begins with a close-up of a woman. after 1 second, The video then jumpcuts to the same woman in the same location on her knees and bent over fully with her hands and forearms on the ground. she moves very slowly and stays stationary and still. her ass is the main focus she is looking back at the viewer, her face is fully visible in view. she stays stationary with her ass up higher than her head as she looks back at the viewer. her ass moves slowly as she models her body, the movement is slow and natural.. she is wearing light blue thigh high stockings and lingerie", just switching out the last line of in the prompt to be whatever i want the girl to be wearing.

>>20338852
yes face coherency will always go down, there is not much you can do about it

consider that WAN only has one reference frame of your girl's face (the input pic) and loras are trained using images and videos NOT using images of your girl...so it's pretty much give that face will change some...some loras change them more than others

always use image with good face, good details. lowering lora weights sometimes help to keep face "intact"

>>20338923
this being said...lowering the loras too much --> lora won't work correctly, missing movement, odd textures

>>20338923
It always seems fine with bj Loras but anything else and they're unrecognizable for me. Even for this missionary which you'd think would work since the face angle doesn't change much. So I think I'm doing something wrong..don't have a workflow screenshot bc I kinda gave up on it a while ago lol

https://civitai.com/models/1983608/wan-pov-missionary-insertion

File: 157508543366344570_00001.mp4 (715 KB, 352x640)

715 KB MP4

>>20338896
wow not terrible results for first try.
same workflow as >>20338848
just switched out the lora and kept the light lora

>>20338929
its not workflow issue.....its about lora and how its trained, what the dataset are etc. There are example pussy loras out there that you would think that they wouldn't not alter the faces....makes sense, they should only alter the vaginas, right? But they do.

>>20332217

Penis thrusting movement?

I feel thick as a brick even having to ask this, but how do I prompt actual thrusting for sex and blow job scenes? every gen I've tried so far with Q8 literally just leaves the guys dick in her mouth like she's having her temperature taken with a cock thermometer. Same with sex, it just stays there motionless or at best twitching slightly. I've been using Wan2.2_Remix_NSFW diffusion models which basically just get it when it comes to sex scenes, it's like the girls mouths and pussys are god damn cock magnets using that thing, but Q8 feels like it's asking me "so what now?".

I've prompted every variation of thrusting, sliding, moving in and out, rapid penis movement, moving up and down you can think of and it's grandmother, but they just stay there like someone put 2 naked coma patients in a sex pose. My guts telling me I'm missing something stupidly obvious like Q8 may be heavily lora dependant unlike Remix which does nearly all the work itself, I'm only really using the ultimate pussy & anus lora atm but the prompts ain't working.

Since Q8 very obviously has better detail than Remix I'd love to migrate to it, I just can't figure out the how to make it work part?

>>20338951
image of the workflow and settings?.....if lightx in use, that will limit the movement a lot, there are some workarounds for it now

File: download_V1.mp4 (3.86 MB, 490x848)

3.86 MB MP4

What program do u all use to string vids together? im using an online program and not happy about keeping them in the cloud.

>>20338977
losslesscut....comes with the installer. Thou haven't used that since svi came out

File: workflow q8.png (1.17 MB, 4481x2450)

1.17 MB PNG

>>20338963

Ok, so all she did in this was turn and put her mouth on the tip but not suck, then water not cum dribbled out her mouth. i had a prepared image made in Qwen of the guy already standing there. the two 4step loras seem to be needed as every gen I 've tried without them come out as a blurred mess?

I'm pretty baffled at this point?

I realize I’ll come off stupid I’m sure. But I download the package and do all steps. I get to having to check if comfy runs and I’m stuck. There is not run file or anything. Where is the actual run application for comfy?

>>20339141
well....first of all you don't have any lora loaded. wan doesn't know what sucking dick looks like, you need some lora for that. try oral insertion, load both high and low in their dedicated loaders and try again

>>20332217
Dumb question, but does your download include comfyui or do I need to download and install that separately and then follow your steps/downlaods?

>>20339789
It installs all you need to get things going

File: Screenshot 2026-02-20 113158.png (91 KB, 1112x622)

91 KB PNG

>>20339941
I went through the steps and when I run run_nvidia_gpu or sageattention i get the following error. Any idea how to resolve?

>>20340236
what hardware? delete the whole folder and run the install again....BUT don't install attentions

>>20340248
nvidia 5070

>>20340258
would be more than good then....

on installer....just install comfyui and after that see if its starts

(installer always install the latest comfyui version.....on hindsight, this was stupid decision from me as, as new versions always break something)

>>20340281
i installed comfyui and the manager addons (option 1 in intaller), and comfyui opens fine. This happened last time too, but you're saying DON'T install the attentions this time?

>>20340297
yeah...skip the attensions....try installing the custom nodes now and after test it if it lauch

>>20340301
alright it opens now. should i continue to install the WAN files (i did Q8 last time)?

>>20340333
yeah....getting models are essential....they won't break the comfy

File: Q8 oral cum lora test.mp4 (1.58 MB, 592x592)

1.58 MB MP4

>>20339465

Right, so...deep throat H&L 0.5 oral inserion H&L 0.5 low cum 0.5 thiccum H&L 0.5

That did the trick. As I suspected it's far more lora dependant than Remix is, that has most of it built in, I only really needed loras for highly specific stuff and better genitalia. So any insertion needs the lora to go with it on Q8 essentially?

btw is there a high and low cum lora? i can only find a low noise one, though I also have the thiccum lora, or doesn't cum need high noise?

File: Screenshot 2026-02-20 134223.png (911 KB, 2486x937)

911 KB PNG

>>20340428
man this prompt took 13+ minutes to complete, which seems way too long. Is that because the attentions weren't installed?

>>20340517
thiccum is old wan 2.1 lora....in 2.1 there was no high and lows. Wan 2.1 loras work in 2.2...load the wan 2.1 loras in both, high and low loader (thou i think only low is needed)

and on your screenshot....you are not using remix version. The thing on these "baked" you can't really control the lora weights.....and loras get updated. its much easier to get new versions of loras than getting 15gb version of new remix

>>20340533
sage attention gives you around 20-30% speed boost. How much RAM you got? Also set enable third sampler to 0 and interpolation to 2, steps to 4-8

get a screenshot of your console window after you hit run and it start generating

I've got various things working with q8 GGUF and 4-step lora with real images.

Can someone let me know how I can use I2V on anime/cartoon/drawing?

File: Screenshot 2026-02-20 141621.png (1007 KB, 2449x1183)

1007 KB PNG

>>20340553
Okay that took about 6 minutes. I have 32gb of RAM. I have steps at 6; does it make it better if steps go to 8? Or should I move it to 4 to make it faster if it wont impact the result

>>20340586
just by prompting it is anime?........

i'm not into anime but there are loads of anime loras for wan 2.2. I guess you could use other loras (ment for real-life images) as well, but then i would lower the low-lora weights to avoid i them pushing real-life textures into the gen

>>20340605
Honestly I'm not really into anime either but if I get a pic of a person like a cartoon character say and I want them to flip off the camera, doesn't seem to work but I'll keep working at it. Works fine with normal people. I'll get it worked out thanks man.

Is there any way to run this in linux?

>>20340601
when using lightx.....4steps is only needed, but i've found 8steps works better

>32gb and 12gb of vram
that card should sample it in pretty good speed. most of the time goes loading parts of the model into ram->vram and off

>man loses weight and becomes muscular
ain't going to happen

>>20340616
like i said....if it is some lora designed just to show the flip off motion.....keep high lora at 1 and low lora at 0.0 - 0.5

then promp

"anime style video where girl flips off the viewer and laughs"

or something like that (dunno what the trigger word for that flip off lora is)

>>20340618
lol on the prompt response (I was just trying to test something). So basically this is as optimized I will be able to get it with high quality then? Bummer about the attentions it would be nice to shave 60-90 seconds off the gen times

Anyone got any blowjob/CIM workflows that are pretty much press and play (with the exception of installing the loras of course...)?

>>20340631
no it's not optimized in any way....this is just my workflow and how i run stuff.....there are loads and loads of different workflow in civitai you to check out

it seems rtx50xx cards (blackwells) have some issues with sage attentions....installation wheels from my installer might not work....leave your disc and i will contact you and troubleshoot this shit out

>>20340658
I don’t have discord but let me see if I can get something set up

>>20340658
May I get input as well on a 50 series? Notabot4090

File: 02_12_2026-11_00001.mp4 (1.33 MB, 592x880)

1.33 MB MP4

>>20338896
>the face looks nothing like the original
This seems to be a common issue with jump cutting, especially with other loras are involved. How much of the body shown determines what type of results you will get. For example, I'm trying to do jump cutting of a woman taking a facial while lying on the floor. This wont work unless the woman's ENTIRE body is in the image or the results will be some goofy looking shit. Not 80% of her body, ALL of it from head to toes. Even then, there is still a chance it wont work. So if you just have the upper half of the woman's body, Wan is too stupid to compute the rest of the human body or how it should position regardless of how detailed your prompt is. Sometimes you are better off just having a prompt telling what position you want the woman in and its the best way to keep her face intact and not sort of look like what she did in the original image.

File: 20260220_153459.jpg (416 KB, 641x1056)

416 KB JPG

Can someone do a test with this one for me? Thanks. Im still experimenting.

>>20341063
got the same error when attentions installed?

>>20341077
in blinks.....did you prompt it "same woman laying on her back at the same location"?

Does anyone know any good loras to get two women to make out, fondle each other, etc?

lmao these smoothmix models turn everything into anime. i wouldn't recommend using these for any realistic stuff.

File: 234406_157573958331029965(...).mp4 (905 KB, 384x480)

905 KB MP4

>>20332217
Can you help me make POV Blow Job videos?

My problem: The camera always zooms in super far.

What I would love:
The kneeling/sitting woman remains in frame, so I can have a cumshot that drips down on her breasts and legs in later extends.
A higher "looking down at her" perspective, not camera at her eye level.

(Workflow is following)

File: pov problems.png (928 KB, 2364x1956)

928 KB PNG

>>20343544
Workflow (rifle version)

>>20343544

Only use a BJ Lora:
"A woman is seen standing. (((The woman then quickly sits up straight on her knees on the floor.))) The camera performs a zooming arc shot, moving upward while staying close to the woman's upper body. The arc continues until the camera reaches a close top-down view of the woman's head. (The woman's upper body is in frame) at the end of the video. (Her hands rest on her thighs.) Her breasts do no shrink in size. She is looking up at the viewer with her eyes wide open the entire time.

Then a man's extremely massive and extremely long penis comes from below frame. The woman sucks the head of the man's cock."

This will help in not reducing or changing body parts because Wan just loves to do that shit. It also keeps most of her body in frame so you can continue off Last frames if you are at that level yet.

Use BJ Loras that can generate dicks that aren't tiny or look like shit. I recommend BBC Extreme.

>>20343567
The reason why I used oral insertion is, because I have no penis in the initial frame.

I will try what you wrote, and see if I find a lBJ lora that doesn't create a mutant nightmare dick.

>>20343544
well you don't necessary need the girl's face in the endframe when using svi....svi is clever enough to sample her face from previous latents and from your input image

also don't use phrases like "NO zoom in" in positive, it will most likely end up wan doing the zoom in. Use "fixed camera, static camera" or something in those lines. for more wider perspective.....i dunno. "man lays in the ground, girl puts her hand on his chest while sucking the penis" <-- this kinda force more of the man's body to be incl. in the gen

also try other loras as well, there are at least few pov bj loras out there

>>20343596
Does the SVI setup use up a lot of memory? Not a fan of the workflows I have seen that uses it so far. Not a fan of AIO workflows.

>>20343615
it's not aio workflow, it only has one purpose, svi.....memory-wise doesn't take much mora than any "vanilla" wan workflow

it just makes 5sec clips with different prompts in them and at the end it merges them together. So 20sec clip is actually 4x5sec. Doing 20sec in single run with wan would most def OOM you out

>>20342500
Nope, doesn't work. I am probably better off prompting the bitch to lay on the floor with a tracking camera then have a man come from offscreen which I am going to try.

File: 021316_157573958331029965(...).mp4 (2.47 MB, 384x480)

2.47 MB MP4

>>20343567
>>20343596
Thanks for the tips! Getting pretty good results.

File: 024611_157965822253586679(...).mp4 (2.26 MB, 384x480)

2.26 MB MP4

>>20343836
i guess the best lora to use to get what you are after is blink blowjob. this will give you man standing, woman kneeling pov scenario.

on the field of view...like said prompting something off screen elements to be shown can do the trick....like "man's stomach"

File: 035130_157573958331029965(...).webm (3.38 MB, 480x640)

3.38 MB WEBM

>>20343845
Thanks for the tip, I will try that.
I tried me previous workflow (from the vid with the black woman) with a few more pictures, and it's still getting a bit too close sometimes, causing even her eyes to sometimes go out of frame.

>>20343845
What prompt did you use for her to take of her shirt in such a realistic manner?

I tried a lot, but it always goes through her neck if I don't have her tear it apart. (and I don't want them to tear it apart, because of the unrealistic "OBVIOUSLY AI" nature of it.)

>>20343845
Actually, I would love to take a look at the entire workflow you used here, if you still have it :)
You did a lot of things here that I like.

>>20343974
https://files.catbox.moe/0h7o4p.mp4
save and drag into comfy

>>20344023
Having success with Big Love. How do I use lustifynsfw?

>>20344023
Thank you VERY much!

I can't find the lora "shirtlift_wan22_14b_i2v_high_noise_v01.safetensors" - or at least I can't find one where the file name is the same.

Is it this one, and your file just has a different name? :
https://civitai.com/models/1586786?modelVersionId=2707115

The result that I found with the exact same file name is on this website, but I can't find a download button:

https://www.runninghub.ai/model/public/2001610908197634049

>>20344863
Go to civitai archive

>>20344878

omg, I've always used civarchive.com thinking that it's the same website as civitaiarchive.com
*facepalm*

I want to give this a try with the wf from the original post
https://civitai.com/models/2053259?modelVersionId=2376074

Which version should I get, with or without lightning?
If I get the ones with lightning, do I just not use the loras that come with the workflow that are always on?

>>20345072
that is correct...if the model have lightx baked in...don't load your own

File: 185620_157920566775850394(...).mp4 (2.3 MB, 384x480)

2.3 MB MP4

>>20344023
Okay, I tried to re-create your video.
I think I got all the same loras.
I didn't change any of the prompts or settings.

The video is generally good, but the guy is jerking his own dick instead of her doing it.

Can you spot the mistake?
(Workflow screenshot follows)

File: upper half.png (1.04 MB, 3334x2025)

1.04 MB PNG

>>20345109

File: lower half.png (533 KB, 3609x1192)

533 KB PNG

>>20345110
lower half

File: asdasd.png (78 KB, 1093x384)

78 KB PNG

>>20345100

So do I just bypass these nodes like this?

>>20345110
update:

I just realized that I'm using
wan2.2-I2V-A14B-HighNoise-Q5_0
while you're using
wan2.2-I2V-A14B-HighNoise-Q8_0

I'll go get the Q8 version and try again

I

>>20345110
try again with different seed.....wan never follows prompts 100% accurate....lower the cfg is, less accurate the prompt-->action will be

>>20345124
yes...or use workflows designed for that model

File: 202511_157704426502615374(...).mp4 (2.27 MB, 384x480)

2.27 MB MP4

>>20345136
I tried again (with the Q8 model) and this time she jerked it :)
She took her shirt off kinda weird, but not a disaster.

Now I'll try to do it again but with a white guy.

mql_stabilizer_rank16_high_noise.safetensors
mql_stabilizer_rank16_low_noise.safetensors

i am looking for these two loras, can anyone share them plz?

also holy shit!
>>20333257
this is such a great gen!

>>20344023

Can the lora
"Wan22_BiggestCock_high_noise_V1"
only create black dicks?
I'm trying to make it a white dick.
I can make the dude white, but not the dick.

If it can't make white dicks: Any alternatives for this purpose? They don't have to be quite THAT long.
I found one result in the archive that's called "big dicks", but it has no preview graphic, no light version, and the description says that it's faulty and only works 1/5 times.

File: 987987987987_00033.mp4 (1.1 MB, 576x720)

1.1 MB MP4

>>20345910
i think biggestdicks lora dataset mainly contains bbc stuff....that why when you have it enable, no matter how you prompt it, it will be big black cock

disable that lora and use some other....here i used blinks blowjob loras

"(ebony woman pulls her dress straps aside, dress drops down, reveals her big nude saggy-tits:1.8). same woman kneels down in front of nude Caucasian white man. She starts giving a blowjob to man. Man skin is very pale."

>>20342489
Yea I got around it though with clean install but could I still get added?

File: dajo059.jpg (1.88 MB, 3648x2495)

1.88 MB JPG

Excellent guide 1st.AId.bag

Fresh install, want to try I2V with wan 2.2, which workflow do you recommend ? Just pick the last one ? 1.5d ?
I'm fairly knowledgable in T2I I2I in comfy but never done videos
5060Ti 16VRAM

pic unrelated

>>20347005
well not much difference...instead t2i model you load i2v model....everything else is same, load additional loras, prompt it. Of course you can't make output highres like making images, you will run out of mem (and time). And yes 1.5d is the latest for singe segment video gen.....latest SVI (longvideos) in embedded in the video (thou there are elements that don't work correctly, yet. like detailer nodes, keep them disabled)
:
https://files.catbox.moe/0h7o4p.mp4

File: 157862557580054313_00001.mp4 (912 KB, 432x640)

912 KB MP4

>>20347023
thx
first test, promising, but I guess I should use loras for better movement

>>20347043
seems like your interpolation is set bit too high.....set it to "2".....then it runs 32fp. But yeah, getting some masturbation or fingering lora should help. There is also motion_amplitude setting, this for battling the slowmo lightx is causing

>>20347123
thx fren, Q8 is definitely a bit too much for my VRAM
will try with your setting.
Is Sage off by default ? I guess I should try to turn the patch sage node to see what it does

>>>/gif/30309174 is Q5 with a lora.

>>20332217
This looks like a great resource, exactly what I was looking for. However the download seems to be down, I'm getting site not found for files.catbox.moe . That filename gives a 404 on just catbox.moe

Am I missing something obvious?

>>20347130
Q8 should work fine on your card....how much ram you got? i've got 4070 and use Q8 without issues, got 64gb ram, but used to run with 32gb before

> prompt and lora were supposed to make her roll on the bed and present her anus

it doesn't matter what you prompt if the model and/or lora don't know how to do it. That lora is trained with videos where woman is laying on her back and rubbing/fingering. You might want to do the first scene with some pussy lora only. might have better luck.

Sage if off by default, if you installed it, go ahead and enable them

>>20347152
works for me....here is mirror
https://mega.nz/file/L1JWUBZJ#LhLjidTwMbaKr-ZrJNLlhjLPHerazaTjBEXkiVsRGa0

>>20347165
Awesome, thanks!

File: 193423_157219655766619591(...).mp4 (1.47 MB, 784x528)

1.47 MB MP4

>>20347005
>>20347160

>>20347320
magic !

File: 02-23_13-33-08-SD_00001_.png (1.15 MB, 832x1216)

1.15 MB PNG

how can I get two girls with their own Loras to gen generated into a single image. I tried taking this image, masking the other girl in a separate workflow but it acts retarded

>>20347383
what are you trying to make, image or video?

>>20347398
image

I have an issue with character consistency, I am using full fp8 model of WAN and it seems like character consistency is rough meanwhile I see clips here with celebs where they do jumpcuts and it keeps their character consistent throughout. Any tips?

>>20347443
have your tried using qwen edit? if you wanna merge two girls in same image (or remove one girl form existing image)....i'm not the best guy to give you tips on image gen/masking, i rarely do them

>>20347489
depends what loras you are using

>>20347502
using f4c3spl4sh mainly for cumshots/facials, having pretty good character consistency now with wan 2.2 remix gguf, it has a ton of nsfw loras baked in and was built for higher character consistency, but would like some more control with fp8 since it is slightly crisper with results.

I have a 5080 and 32gb of ddr4.

>>20347524
do your tests one lora at the time....lower the lora weights (especially lower lora ones) until the lora does what it should and the face consistency is on acceptable levels.....if that doesn't help, just dont use the lora

>>20347383
controlnet with face reactor set for two faces and donor images.

Qwen edit to swap heads from multiple donor images.

Regional prompting image generation.

Inpainting can work, but as you can see, can be finnicky. Your mask may be too large, your prompt innappropriate, or the loras are working agqinst you.

File: results.jpg (1.04 MB, 3636x1200)

1.04 MB JPG

>>20347383
>>20347580

>controlnet with face reactor set for two faces and donor images

>>20347165
could you share just the 1.5d workflow pls, don't want to reinstall everything!

File: tessst.jpg (605 KB, 2717x1024)

605 KB JPG

>>20347580
>>20347383
>Regional prompting image generation *with loras
>I don't have celeb loras because I of my half ass workaround workflows, so this is regional prompting with facereactor

>I say face reactor, but for this and >>20347678, I use a combo of IPAdapter and face reactor for improved results

File: tst.jpg (486 KB, 2097x1052)

486 KB JPG

>>20347383
>>20347580


>inpainting can be hard to convince the AI to match the sizes of replaces heads, sometimes it is easier to replace just faces, or as you tried, whole upper bodies, but the larger the mask, the more ambitions the AI can be in trying to create an image, which could be one of the causes of your inpainting issues, or as stated, the AI having difficulty finding a reasonable size match

>Depending on how the lora is trained, it might try to create more than just a head or a face, but I am not certain about that

I would like to start making some i2v of undressing, but i have a 7700 sapphire. Should i give up?

>>20347781
my pack is not for you but look and read
https://github.com/patientx/ComfyUI-Zluda

File: tts.jpg (1.66 MB, 2339x1760)

1.66 MB JPG

>>20347383
>>20347580

>though qwen and it alikes have in the past had skin issues and genital issues, they can be quite effective for simple object and face swapping

File: wan22_00548-merged-177006(...).mp4 (1.61 MB, 560x720)

1.61 MB MP4

>>20347708
nvm found it in the OP my bad. looked before but didnt see it. random gen unrelated.

>>20347870
all workflows are in the installer rar...extract that then the addon.rar....wf are in user folder

btw gotta love that man in the bg getting an inhaler shot =D