>What is it?
Modified version of comfy portable
Aimed for new apprentices who wanna jump in AI world. In the package
----Kijai's WAN-Wrapper nodes pre-installed
----Pre-installed nodes needed for my wan2.2/wan2.1 workflows (wf included)
----Pre-installed nodes needed for pisswiz's wan2.2 workflow (wf included)
----ReActor preinstalled, NSFW patched (note that updating the node might fuck it up)
----Sage & Triton installation
----Simple batch file by me that downloads all needed files (vae/clip/lightx-lora/ggufs) to get you started
----Incl. program called LosslessCut to join your clips together
----Couple upscale models included

>How
----Download the package
----Extract it on some fast drive that have good amount free space (100gb at least)
----Run the run_nvidia_gpu and see if you get your comfy running
----....if so, good, close it up. It won't do anything as you don't have models yet
----Install VC_redist.x64.exe and reboot (most likely you have this installed already, needed for triton)
----Download (1staidbag.bat dload fix into your comfyui root, replacing the old one, fixes lighx lora dowloads)
----Run 1stAIdBag.bat
----....Download options 1&2 and pick one of the GGUF models (Q8, Q5 or Q2)
----....Downloading these models take time as those are several gigs
----....Install sage and triton (also updates flash attention)

>Already got old install?
----Backup (cut) models and user folders somewhere (pref same drive's root where your comfy is)
----del ComfyUI_windows_portable folder
----read how section above
----paste folders back into your fresh comfyui portable

Once all downloaded. Launch Comfyui using run_nvidia_gpu_sageattention.bat, RESELECT ALL MODELS, to make sure
they are all there in your computer. Load image, type something in the prompt, run.

If expecting path/string(NoneType) errors ---> you didn't reselect the models
If sage related error = you didn't install it correctly/ or at all --> install sage, or disable the nodes
If out of memory / allocation to device --> lower output resolution, shorten vid length, lower steps, enable blockswap

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
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)

Is there a specific lora or prompt you use to undress quickly?

Any non-mega links? I'm already at their download limit and don't want to use a proxy

>>20053712
i've also struggled with this. prompting to remove clothes might take up 80% of the video.
'clothes disappear' or 'she removes all clothes' or 'with one swift movement she pulls all of her clothes off'
sometimes it just refuses to work no matter what

any nsfw lora seems to help

>>20053712
>>20053979
try "scene suddenly change where girl is naked"

>>20053759
yuo can resume it after few hours

>>20053999
yeah if i prompt that she is in a completely different environment from the picture, there is sometimes a sudden change - sometimes it's a strange picture wipe effect, other times there is a fast fade transition. yet to work out how it decides which it goes for.

File: Video_04230.webm (1.38 MB, 1184x1440)

1.38 MB WEBM

>>20053979
if you only care about getting them naked, i like to throw this in front of my prompt to just get the clothes off,

the woman has a clothing malfunction causing the woman's clothes fall apart at the seams. clothes falls off her body and onto the ground quickly, exposing her naked breasts and naked vagina and completely naked body the woman is completely naked leaving her wearing only white sneakers. then <YOUR PROMPT>

it basically makes the clothes fly off the body, so it doesn't waste time trying to take it off then to do what you want

yes it does have a lot of the same prompting on there, it's for confirming prompts if that makes sense

>>20054098
naked with white sneakers, a man of fine taste

What is the difference between Lightx2v and Lightning?

>>20054165
i'd say nothing if you are talking about the loras...same lora, different names

>>20053516
Have you been using wan animate? Wouldn't it be better than the RNG we generate normally?

>>20053522
>enable blockswap
finally it works

>>20053516
Do you do a T2V pack ? your workflow is one of the only ones I've managed to get working.

>>20053522
just a question, what's the best sage option for potato pcs?

File: 1111566074470230_00001.mp4 (1.54 MB, 480x640)

1.54 MB MP4

getting these hoes to strip is so cumbersome,any decent add on for it?sometimes>>20053979
this might help but weirdly it also zooms a ton on the bellies(?

for boob reveals in the wan2.2 1.14 workflow, i'd use I2v-tittydrop-e26 for both high noise and low noise? is there one specific for wan2.2?

Just a quick question.
I'm currently generating pictures using

wan2.1-t2v-14b-Q8_0

Is there an improved NSFW specific unet GGUF to create porn pics?

File: 1761860172842546.mp4 (1.64 MB, 800x960)

1.64 MB MP4

>>20053522
Which lora is this and the workflow is quite successful can you help me

>>20055405
If you mean something similar to some sdxl checkpoints.....no, or at least haven't seen one. So just the base model and loras is the way to go

There is specific nsfw Wan text encoder that should do better with xxx prompts

up

>>20056370
Thanks!

Where do we find larger than Q8 GGUF models?

>>20057347
Q8 are the top gguf models, if you want better quality then use the diffusion ones fp8 (quality almost the same as q8s) or better fp16. Note that you need better rig to run these, they also eat more time to gen

>>20057361
https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/tree/main/split_files/diffusion_models

>>20057361
>>20057374
Gotcha!

Is there any way to increase video quality by sacrificing the time it takes to make it? I don't care if it take up to an hour as long as the quality is good. Or will it just run out of memory and fail?

>>20057516

what's the point

>>20057731

Because i dont have a very good card

>>20053516
How much are the pc requeriments to run this?

>>20057516
use better models, increase the reso, increase the steps, don't use acceleration loras....but if you don't have good card/ram amount you will face OOM pretty quick

try increasing the output resolution (need to have good resolution input image of course) and steps little by little at first

>>20057959
i'd say min = rtx 30xx and 32gb....lower can run it but very slow and less quality (as you would need to use lower models)

File: which one.jpg (19 KB, 277x221)

19 KB JPG

>>20058000
question not same persom, how does sage help and what are the best ones to use for lower end shit. For context have a 20x, generates at the low end settings and generates just fine, just find it slow. heard sage can increase speed

>>20053519
How much VRAM do I need in to use Q5?

up

>>20058245
i've always left it to auto...there is not that many options if you wanna test them out. sage 2.2 should boost the speed about 20-30%

File: 157253377212423942_00001.mp4 (1.83 MB, 960x1280)

1.83 MB MP4

>>20053516
just wanted to say thanks, this helped accelerate my understanding of how this all works. still a long way to go, but a great springboard.
got any loras you like?

>>20053516
Anyway to get your StartFrame/Endframe workflow to output perfect loop vids? the loops it give are a bit choppy.

Up

>>20059158
never planned on using that for loops.....try adjusting/disabling the wan skip last frame node...... i dunno really, would make sense to skip the last frame if it is the same as first frame thou (to avoid stutter at the end

I tried this and keep getting this error:

Power Lora Loader (rgthree)
Error while deserializing header: header too small

# ComfyUI Error Report
## Error Details
- **Node ID:** 103
- **Node Type:** Power Lora Loader (rgthree)
- **Exception Type:** safetensors_rust.SafetensorError
- **Exception Message:** Error while deserializing header: header too small

I didn't change the workspace at all, everything's the same as yours when you uploaded it.

I cant figure it out, pls help

>>20060146
download the lightx loras manually into models/lora folder (my auto-dowload is broken in the pack)

high: https://civitai.com/api/download/models/2361379?type=Model&format=SafeTensor

low: https://civitai.com/api/download/models/2337903?type=Model&format=SafeTensor

>>20060158
oh and once dowloaded...restart comfy and re-select them on the dropdown list (they are newer versions)

>>20060160

hey thanks for the quick response, it seems to be working now from my test run!

so, to get add new loras, I click "Add Lora" in the "High noise LORAs" node? I still need the lightx loras, right? I just put the new loras under that one?

Also, this works for pictures of all sizes (resolutions) right?

Also, any way to change the length of the video?

Sorry I'm brand new to this

>>20060172
no need to add new lora.....click the existing one and select the one you just downloaded (on both sides, high and low)

>Also, this works for pictures of all sizes (resolutions) right?

yes...input image sizes won't matter, they got resized before the gen.....some say 1:1 pictures work better, like 720x720...but i dunno about that on wan 2.2

>Also, any way to change the length of the video?

there is a slider for that in the workflow...keep it in 4-6 secs thou, wan can't handle more (atm)

>>20058640
thanks for response. Did so did get faster but sort of negligble once your pc has been on for a while
>no need to add new lora.....click the existing one and select the one you just downloaded (on both sides, high and low)

also what

>>20060486
Just make sure you have one lightx high lora on high-lora loader and one lightx low lora on low-lora. You can remove a lora from the list by right clicking and remove

>>20053516
Got a question about fine tuning, I got the Q5 model with the apprentice pack and 5s videos come out relatively fast on my 3060, but most of the time they're messed up (I'm trying mostly high noise models, idk if that has something to do with it). Is there any.
Is there some parameter I can mess with so that the ai spends a bit more time on it but makes videos that look more consistent? Or is it something wrong with my loras? I started using more, but right now I'm only using the lightining ones that come with the download and a sex mashcut one I found on civitai.

>>20061769
post screencap of your whole workflow so i can see the settings, and example of your output would be good too

>I'm trying mostly high noise models

wan 2.2 wont work like that....if you got a high model/lora you need a low one to company it

File: snapshot.png (158 KB, 332x358)

158 KB PNG

>>20061773
It's all still on default, I just added some loras, a starting picture and a prompt. But if I need both low and high, that's probably why it's so fucked up.
Gonna download the low noise counterparts and give it a try, thanks man!

File: llama-stare.gif (617 KB, 220x196)

617 KB GIF

>>20061778
......yeah, that seems pretty fucked up

File: 157698395660621476_00001.mp4 (1.18 MB, 336x640)

1.18 MB MP4

>>20061778
>>20061784
Still odd, but at least now I think it's just cause my prompts aren't specific enough. Thanks again
While I'm at it, does anyone have recommendations for loras? Or some uncensored ai that can help with prompts?

>>20061900
are you sure you re-selected all models and loras when you opened the workflow for the first time?

and again....post the the whole workflow image so i can see the settings. I know you said you didn't change anything, but by default it shouldn't look like mutilated body horror scene.

I think i know what you did..... You just copied some prompt somewhere and didn't look what loras are needed to make that prompt work

>>20061910
>are you sure you re-selected all models and loras when you opened the workflow for the first time?
I did, but since you mentioned it again I went back to check it and I messed up by selecting the low noise model for both the high and low ones... Man, I'm surprised it even did anything close to resembling a person while setup like that.

Thank you for not giving up on a dumbass like me

File: images (97).jpg (22 KB, 554x554)

22 KB JPG

>>20061778

So i'm trying out your wan 2.2 WF, and I decided to up the settings where I can on it, such as the following.

using the fp16 diffusion models, switching the clip node to umt5_xxl_fp16 and also wan2.2 vae, but i notice you say to use 2.1 for the vae node. or else i end up getting this error

"Given groups=1, weight of size [5120, 36, 1, 2, 2], expected input[1, 64, 25, 54, 44] to have 36 channels, but got 64 channels instead"

so im having to switch it to 36 channels, is there a way to swap in a wan2.2 vae node instead, and if yes, is it worth it?

i got a ryzen 9 9950x3D 16-Core/32-Thread
64gb of DDR5
and a RTX 5090

what would I need to do?

>>20062438
Wan2.2 vae is for 5B models..... Wan 2.2 14B uses old Wan 2.1 vae

>>20062499
so id have to do a lot of butchering to the workflow then since it wouldnt be used in a low and high diffusion model setup then?

if thats the case, then can you recommend any really good wan2.2 5b workflows that's as extensive as yours?

up

>>20062506
Well 5b model is worse than 14b.....your rig is good to go 14b models.... Plus there is hardly any good loras for 5b

>>20063130
okay. is there anything more i can do to increase the quality to be noticeably better? ive enabled diffusion already instead of using the GGUF. will switching to fp16 for the high and low diffusions and for the CLIP node make things higher quality?

Im getting SamplerCustom PassManager::run failed, any advice? Sorry I'm new to this

>>20063227
well those will make thing looks better for sure. If the time is not the issue for you, then you could run the gens without lightx loras and increase the steps into 25-30 area

>>20063485
didn't install sage-attention? disable sage nodes and try again

>>20063504
does the lightning loras not work with fp16 diffusion models?

>>20063804
they do....but they also take bit of the quality away for the cost of speed. Other good thing running the gens without lightx = prompt coherence is better, movement is better....but you would need lots of more steps (aka time)

but once you've get used to run gens with lightx 4-8steps....going back to old ways ain't that tempting thought

>>20063870
no wonder.....you have two high noise diffusion models loaded

also.....what is the thing with you guys writing a novels into the prompt.....You got 5-6 secs. keep it simple,make stuff in parts

>>20063884
ah yeah i noticed last second. and desu, i feel i get much better accurate scenes when typing a paragraph especially when stacking multiple loras to make them multitask during a scene such as "saliva dripping, handjob while sucking, taking off their shirt to reveal boobs, facial expressions etc etc" versus typing out simple basic stuff. if im trying to go from one certain motion to a whole other motion, then sure. ill break it up into parts.

also, i can do 13 second blowjob/cum videos just fine with consistency staying stable with interpolation at 5 and seconds at 10.

File: 157194589621123194 00001.mp4 (1.57 MB, 560x864)

1.57 MB MP4

>>20063884

>>20063884
also idk why...but my comfyui seems to crash more often when trying to do fp16, it just randomly "reconnects" and disrupts the workflow entirely. is this a known bug or something?

>>20063905
well because fp16 models and clips are much heavier on your rig....if you ask me i find Q8s and fp8 clip good enough, just add some steps and reso for bit better quality on them

>>20063919
damn, even for a 16-core processor and 64gb of ddr5 and a 5090? that's crazy. but alright then.

>>20063892
interpolation at 5?....huh is you fps set to something like 120?

>>20063923
something like that. i fiddled with it until it began to look somewhat smooth and have that '60fps' look to it.

>>20063921
well....those are 30gb model files each.....then you need to load the clip 10gb, then the loras. thats loads of loading stuff into ram, then partially loads stuff into vram, clean them up do the same shit again....even good rigs run out of space

>>20063946
well...learn something new every day. thanks for the help though, i appreciate it.

>>20063894
so is this fp16 output? post the input pic you used

File: 20251104_060719.jpg (854 KB, 1079x1660)

854 KB JPG

>>20063962
Nah just fp8. And this is the starting image

up

File: 157501110646704412_00001.mp4 (2.85 MB, 496x784)

2.85 MB MP4

>>20064534
if you are using image quality like this....it's kinda pointless to use heavier better models. They won't make your output look sharp hd quality. There is saying "Your painting will never be more detailed than your scenery is"

made one with 6 steps 4xinterpolation 60fps using Q8s

im having trouble running PissWizard_WAN2.2.json
i pretty much changed nothing about the workflow, i input the image and try to run it.
it gives me this error:
Power Lora Loader (rgthree)
Error while deserializing header: header too small

File "K:\comfy\ComfyUI_windows_portable\ComfyUI\execution.py", line 289, in _async_map_node_over_list
await process_inputs(input_dict, i)
File "K:\comfy\ComfyUI_windows_portable\ComfyUI\execution.py", line 277, in process_inputs
result = f(**inputs)
File "K:\comfy\ComfyUI_windows_portable\ComfyUI\custom_nodes\rgthree-comfy\py\power_lora_loader.py", line 56, in load_loras
model, clip = LoraLoader().load_lora(model, clip, lora, strength_model, strength_clip)
~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "K:\comfy\ComfyUI_windows_portable\ComfyUI\nodes.py", line 675, in load_lora
lora = comfy.utils.load_torch_file(lora_path, safe_load=True)
File "K:\comfy\ComfyUI_windows_portable\ComfyUI\comfy\utils.py", line 75, in load_torch_file
raise e
File "K:\comfy\ComfyUI_windows_portable\ComfyUI\comfy\utils.py", line 59, in load_torch_file
with safetensors.safe_open(ckpt, framework="pt", device=device.type) as f:
~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
safetensors_rust.SafetensorError: Error while deserializing header: header too small

>>20065447
download the lightx loras manually into /loras/ folder.....see >>20060158

oh the level of nitpick ive gotten to in the last month of doing this
why are we running rife vs gimm? Speed, or easy of accessibility? Under the influence that gimm is better interp
also, very good workflow, probably producing the best results vs time out of anything ive used from reddit or civit

>>20065784
actually i swapped to gimm as well....seems faster than film or rife, should be on the next version

>>20065837
i threw on reactor as well after color match, running 4x interp 60fps, i must say, this looks pretty solid

>>20065837
also i found the moe ksampler that loads both models into the same sampler and then uses a degrading cfg over the steps gave pretty good results, but i did find it to be slower compared to two sampler workflowers for some reason, although it might just be the high steps taking longer as they appear to do on this wf as well.

>>20065890
well models are not loaded at the same time per say...what moe does, it auto-calculates where it should split the steps from high noise --> low noise (it still needs to swap'n'load the models)

i've been playing with triple-sampler past week, it has 3 model inputs,
1.high without lightx so you can run it with higher cfg (like 3-5) few steps, then
2.high with lightx with cfg 1 (moe calculates the split point where......
3.low with lightx starts and do the rest

about 30% slower process but sometimes much better motion on gens

>>20065420
What kind of prompt did you use for this? And is it the default loras from this workflow? Sorry, this is all new to me. The videos I'm generating don't even know what genitalia looks like and ignores a lot of my prompts, i.e. a girl will just smile instead of do something sexual.

Workflow: PissWizard_WAN2.2
LoRA - FIRST SAMPLER -> Wan2.2-Lightning-I2V-A14B-4steps-lora_HIGH_fp16.safetensors
LoRA - SECOND SAMPLE -> Wan2.2-Lightning-I2V-A14B-4steps-lora_LOW_fp16.safetensors.

I haven't changed any of the other settings. The default prompt seems to work well, but as soon as I do anything sexual it seems to fail pretty badly.

Thanks!

File: 1000024430.jpg (54 KB, 720x960)

54 KB JPG

When I use a prompt like 'Woman in turquoise bikini on beach removes top and bottom, licks and pinches hard nipples, fingers deep in shaved vagina, intense masturbation, wet fluids, close-up, photorealistic, 8k, explicit, no censorship', the result doesn't even do half of what the prompt says. Could it be limited? How can I change that? I can't get it to do exactly what I want...

up

>>20066009
wan doesnt do nudity and nsfw by default on its own....you need to download some loras from civitai for example (you need to make account to view nsfw content)

>>20066067
what loras are you using?

Hello, I have local stable diffusion and animatediff installed. Could everything posted on this also work in stable diffusion or is all of this just for comfyui? Sorry I’m new to Ai video generating.

>>20066067
>Woman in turquoise bikini on beach removes top and bottom
removes top what?

>licks and pinches hard nipples
need breast-play lora

>fingers deep in shaved vagina, intense masturbation, wet fluids
need some female masturbation/fingering lora

>close-up
=camera zooms in to [something]

photorealistic, 8k
>kinda pointless, wan can see the difference in styles, like photo, anime, lineart etc. Resolution you set in other settings, not by typing it into prompt

xplicit, no censorship
>useless prompt words


>I can't get it to do exactly what I want...
It very rarely does....more shit you add into your prompt, more confusing it get. Sometime you need quite few runs until you get good results. But thats the name of the game atm

>>20067183
>local stable diffusion
what is that? stable diffusion is a model type, just like flux, wan, qwen, etc.. You are most likely using some gradio interface for it? comfyui is also an interface to use ai models...

anyways....you need is on the op post, read the notes'n'go

>>20067183
once you grab a workflow just search the stuff in there, go grab it, plug it in and you're in business.
Turn off sage if you can't figure out pytorch. Go into settings and enable the memory overflow/sharing whatever it's called.
Also, read. A lot.

>>20067238
I installed a local stable diffusion to my pc which requires installing python and this other thing I forgot, which then after putting some code in my pc’s command center created a webui in my file explorer which by clicking opens up stable diffusion on explorer. You’re saying instead of using stable diffusion, I should just install comfyui to my pc?

>>20067335
short answer; yes,,,,you can use stable diffusion models in comfyui

what's a good titty drop lora??

>>20065898
ever since switching to the 1022s i really havent noticed that many motion issues, i tried triple ksampler too but found it again slow and unneccessary. could be wrong.

>>20067210
Does one of the existing workflow in this work? I downloaded a few loras, but I'm unsure which workflow I can use as a starting point. If I use one like PissWizard, there is a high and low lora set up here and I only get one safetensor file from Civit. Just trying to get a basic example working that I can learn from. Everything is either hardly NSFW or incredibly blurry.

Most the loras I see to download from the link provided are for 2.1 and PissWizard is 2.2.

>>20067210
NTA I was there for the first thread three months ago and god damn. Some of it looks like common sense at this point but then you need more loras.
I'm in love with the most recent anal insertion.

Up

>>20068679

what's a good titty drop lora

File: 1744398448704128.mp4 (1.45 MB, 480x832)

1.45 MB MP4

>>20069048
wan 2.2 doesn't seem to need one imo

>>20069352

whoaaaa. what you are using for lora?

>>20069352

prompt?

>>20069352
Jesus those tits fit in naturally. I've been messing without using any loras as well. It's hit or miss depending on what you're going for.
Struggles a bit to pull the top of a say tube top or dress down. That's the only case I've been forced to lora up in.

File: 1731219707609824.mp4 (1.13 MB, 480x832)

1.13 MB MP4

>>20069385
high noise: Wan_2_2_I2V_A14B_HIGH_lightx2v_MoE_distill_lora_rank_64_bf16

low noise: lightx2v_I2V_14B_480p_cfg_step_distill_rank64_bf16.safetensors

both at 1.0 strength

>>20069387
Amateur handheld footage captures a 30-year-old African American girl with tattoos, standing in a gym, taking a selfie in front of a mirror. Her phone and the recording are tilted at an angle. She is slightly overweight and has large saggy-tits breasts. Her neck, arms, chest, and right leg are tattooed. She is wearing a black sports bra, grey shorts, black fingerless gloves, black sneakers, and has a black hoodie tied around her waist.

Both of her hands rest on her phone held near her stomach, and her left leg is slightly raised. Remaining in place, she turns the camera upright, correcting the angle of the video, and moves the camera slightly so that her torso is clearly visible.

Keeping the phone in one hand, she uses the other to grab the bottom of her sports bra, and lifts it such that both of her breasts to fall out. They jiggle as they fall from her top. Her dark nipples and areolae are visible.

She holds her top up and pushes her chest out slightly, leaving her breasts clearly in view for the rest of the video.

Detailed skin, skin imperfections.

>>20069407
thanks, yeah a lora would make things more consistent. this picrel as you say is a little more wonky with the tube top. not bad but it's sloppier

>>20069423
>Wan_2_2_I2V_A14B_HIGH_lightx2v_MoE_distill_lora_rank_64_bf16


does that even work with bag's workflow??

>>20069613
no idea

>>20067345
Wait what? Stable diffusion models? I thought stable diffusion was just a webui where I can install models in there from civitai? I use civitai models. I feel more lost than ever. Can some one just tell me what I’m supposed to do so I can make a boob drop using stable diffusion as my local webui?

>>20069912
without screencap of your webui i would guess your webui is called automatic1111, it uses stable diffusion models (SD/SDXL) Those are checkpoint models for making IMAGES. it makes images....not boob drop videos, nor any other videos

animatediff is just to addon to make your IMAGE move a bit, its very very limited what you can do with it. You need video models and webui to make videos, like comfy. All you need to know is posted in first 3 posts in this thread, if that goes over your head, stick making images

Up

>>20065420
your vid quality is nicer than >>20063894
but your motions arent nearly as good as his.

up

>>20069964
Ok yes that’s what I was forgetting, I was using automstic1111. I understand now. Thank you very much!

>>20070499
that is one downside of gguf and lightx, motions take some hit

Hello, I installed comfyui and I have ComfyUI_IPAdapter_plus, ComfyUI-AnimateDiff-Evolved, ComfyUI-VideoHelperSuite as custom nodes. Should I add more custom nodes from github or is this fine? Also what are the necessary models for nsfw ai videos?

>>20071845
instructions are on op post. pack includes all the needed nodes work my workflows, all you have to do is to download the models, and there is a downloader in the pack.

but since you didn't get that.....open the workflow you want to use and it tells you what nodes you are missing, use manager to install them, and download the model you want to use manually

up

Up

Up

bumbtsibum

bump - harness the power of wan, lads!

>>20053516
how tf do you guys do it? I feel so retarded looking at this shit. it's empty but when I dragged one of the workflows to the screen its just a web of shit I don't understand. I clicked run to rest it but ended up with an error saying:

"UnetLoaderGGUF expected str, bytes or os.PathLike object, not NoneType"

no clue wtf I'm doing here wish this was simpler man I just want to make grok generated images I made getting fucked or sucking dick lol this is too much for me but any tutorials for retards would be helpful.

>>20073554
it's not that hard once you get a hang of it. Likely the issue is that if you're the pack that 1st Aid Bag made, you didn't run the installer for the models., a nonetype error indicates that the model is missing. you can try to reselect the file on the loader, if you did download it already.

hey guys newbie here hoping to get some help.

i keep getting the error :
Error(s) in loading state_dict for T5: size mismatch for shared.weight: copying a param with shape torch.Size([256384, 4096]) from checkpoint, the shape in current model is torch.Size([32128, 4096]).

not sure what to do

So I finally got this all working and I've gotten some results that look pretty good, but I'm having major issues with the subject's face staying consistent. It seems to be changing faces in drastic ways whether or not the subject is actually moving in the video. Any ideas?

File: 1750372859462743.gif (24 KB, 212x219)

24 KB GIF

I do not apologize for bringing fledglings here. More people should learn.
Troubleshoot the shit out of it if it doesn't work. Google is your friend.
>>20074009
Maybe lower the weight of the lora a bit. Denoise of about .97 seems to be a sweet spot. Sometimes you have to mess with weights and run it a few times to get the results you want. And sometimes a lora will force faces in you just never asked for so don't use that one if it keeps doing that.

The issue I'm having is that all of the gens with the WF come out in absolute slow motion even with interpolation lmao

>>20074122
increase the framerate. I personally use 32 but the higher you go the longer the gen time I think.

Is there any way to do Image to Image with this package? I basically wanna add a nudify option without the video

>>20074329
No, Wan is a model for video generation.

ComfyUI > Workflow > Browse Templates > Basics > Image to Image

You can also get plenty of image to image workflows on Civitai

>>20074349

Can i do nudify with just the basic templates in comfyui?

>>20073931
Sounds like clip mismatch.... Post a screenshot of whole workflow

>>20058650
For some reason I cannot find the most basic Lora I need. Where do I get the Lora to make them remove their clothes? Can't find it on CivitAI

>>20074509

you need to turn off the SFW settings in your account, sometimes you have to do it multiple times as it doesn't save. There's like 5 options

>>20074509
2.2 alone should work pretty well. some good loras have been purged as well so that doesn't help.
https://civitaiarchive.com/models/1423479?modelVersionId=1608935
https://civitaiarchive.com/models/1387238?modelVersionId=1567740
https://civitaiarchive.com/models/1597473?modelVersionId=1807726

>>20074387
yes but you'll need a model capable of NSFW gens and the example workflow will be bare bones, it's just to show you the concept really. so don't expect great quality but it'll be the simplest form of a nudifying workflow you can see.

i think getting comfyui set up and reverse engineering the template workflows is a good way to start.

File: image-0064.jpg (193 KB, 960x1440)

193 KB JPG

Thanks for the poison you have brought into my life. ive been playing with this for 5 weeks now and ive created everything i ever wanted, thanks.

my wife also left me but its fine

File: pissman.mp4 (264 KB, 320x480)

264 KB MP4

>>20074793
sample

>>20074793
>my wife also left me but its fine

there are always hookers and the doordash one click away.

>>20074793
>my wife also left me but its fine
welcome to the brotherhood

Anybody try this app called Pinokio? I got frustrated with Grok trying to troubleshoot this shit and it suggested Pinokio as a way to get img2vid nsfw shit but easier. Was curious if anybody uses it with the comfyui plugin it has and if that simplifies this for retards like me to understand.

>>20075157
As I understand it, Pinokio is like Docker where it simulates a venv, so it makes installing and running easier, but Comfy Portable is already shit simple. Definitely easier than local comfy, I think I tried it to get that low resource/I havea shitty old card stable diffusion one going...the one made by the github user with like roadrunner as a name.
But I could be wrong in my understanding.

up

So it's telling me I need to install something called nvidia-ml-py or update. Is that something I should worry about or did I just fuck up the program somehow?

I did update everything and was downloading a lora for dick sucking prior.

>>20075677

Update: Said I needed Git so I installed it, that seems to be working fine now.

Now it's telling "The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you."

This is probably an insanely easy fix, but I'm

1. Retarded
2. Probably Autistic.

So someone tell me how to fix this like I'm a 5 year old.

>>20075724
No need to worry that msg at this point

>>20075752
ComfyUI doesn't start up because of it. Guess I just wait until the people who made the pack update it?

>>20075757
did you try the first result in google? It is for Wan2gp, but might work

https://github.com/deepbeepmeep/Wan2GP/issues/925

Can't break it more, and worse case scenario reinstall and copy over your models folder. (providing you're on portable.)

>>20075790
Got it working again. Just needed to reinstall Triton and Sage through 1stAldBag.bat.

I still get the deprecated thing but it at least works now.

Need some help, getting this error message

SamplerCustom
Input tensors must be in dtype of torch.float16 or torch.bfloat16

>>20075883
your model might be corrupt (0 bytes or partial)

re-download the model

>>20076016
Re-downloaded, same error

File "D:\Torrents\ComfyUI_windows_portable\python_embeded\Lib\site-packages\sageattention\core.py", line 242, in sageattn_qk_int8_pv_fp16_triton
assert dtype in [torch.float16, torch.bfloat16], "Input tensors must be in dtype of torch.float16 or torch.bfloat16"
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
AssertionError: Input tensors must be in dtype of torch.float16 or torch.bfloat16

>>20076081
again would be 2000x times easier to help you if i could see the workflow and what you selected on it, always post screencap of the workflow.

But it seems like sage related issues. Either it is not installed or you change it from auto to something else......disable sage from the options panel and try again. If it runs, we found your problem

Can someone help me with 2 things? I got it all to work, but wondering:

1-is there a way to make the video not so slow-motion? The characters move so slow they sometimes don’t have time to act out the action (like removing clothes)

2-is there a way to make the face not change as much, and stay consistent. Even if it’s just like keeping the head from moving so much.

>>20076243
both are lora related issues. Some loras change the faces a lot because the training data they were build on had faces in them. Some do little extra effort to remove the faces before training --> your input face has better consistency. Try different lora that does the same thing or try lowering the weight of the loras, that usually helps. The issue not completing the action could be also from your prompting

On speed...kinda same cause as above, training clips might have different framerates --> sometimes cause slowmotion. You could try to prompting "normal speed" "fast movement". Might work. But the biggest reason for slow motions / less motion in general is lightX lora. That is kinda the downside of using it....but if you don't you would need to run 20-30 slow steps to get the same quality. Some say using triple-samplers helps to keep motion fast, i'm testing that currently

also interpolation/fps settings

interpolation 2 + fps 32 = pretty normal speed
interpolation 3 + fps 32 = slowmotion (higher the interpolation value = slower it plays)
interpolation 2 + fps 60 = faster motion (higher the fps value = faster it plays)

up

>>20053516
Anyone got tips for generating "pan down to see missionary sex" vids? 85% of the time the pussy just looks weird. I'm using pov_missionary and pusfix right now, doesn't seem to be much help.

>>20076600
try to use Smash Cut

Up

You can do it anons! roast that gpu!

anyone have much experience with mmaudio? found an nsfw model for it and it can be finicky but ive been getting decently good results. any chance we could see it in the next 1staidbag workflow?

>>20077766
care to post some samples and workflow so i can see how it works. Never used it, but i assume its text2video2lipsync kinda thing?

>>20077766

post the model, workflow, etc.

anyone know a good lora for fingering pussy?

i tried

https://civitai.com/models/1444759/fingering-for-wan-t2vi2v

but it doesn't seem to work for me. this is what i want to try and emulate:

https://tensor.art/template/929598193845450754

File: merged-1762924117755.mp4 (3.87 MB, 576x720)

3.87 MB MP4

>>20078102
well that is older wan 2.1 lora, that should work thou....could you open what you mean "not working for you"?

there are couple fingering loras for 2.2

https://civitai.com/models/2007360?modelVersionId=2272024 (last 5secs of the sample)

https://civitai.com/models/1952032/perfect-fingering-wan-22-i2v (first 5sec of the sample)

>>20078488

might be the prompt but my subject just stares at the screen lol. i will try those others one. thanks bag.

>>20078488

what prompt did u use for first 5 sec?

>>20078557
>Girl is nude. She crouch down, legs spread open. sensual_fingering, The beautiful girl is fingering her vagina with her two middle fingers, her fingers are slowly sliding in and out of her vagina. She lifts up her shirt with one hand to reveal her huge nude breasts.

File: Joined_00002.mp4 (3.13 MB, 928x768)

3.13 MB MP4

for those who are interested. And some1 asked about slow-motion/movement in general. Made compare vid using same fixed seed and prompt and loras (fingering_pussy and big_boobs for wan 2.2 at 1.0 weight)

>camera zooms out. girl wearing white skirt without panties and white thigh high stockings, girl moves her left hand on her nude vagina. Spreads legs open. Then girl starts pushing her fingers into pussy, two fingers, solo, targeting her vaginal entrance, slow deep thrusts, her expression is of intense pleasure with mouth open and brow frowning. She lifts up her shirt to reveal her big nude breasts with her right hand.

as you can see it make quite a diffrence if you add few steps of "lightX-free" noise into your gen

If I give u a picture could generate a video for me? I’d like to learn but I don’t have a pc.

>>20078575
these threads are not for requests. make a new thread or post into some existing one and maybe some1 helps ya out

File: Screenshot (1).png (66 KB, 422x715)

66 KB PNG

>>20078488

is this what u have?

im using ur same prompt and getting bad results

>>20078582
post your result and whole workflow

>>20078607

ill have to do it later today.

up

>>20076642
the pussy and penetration look very good with this. the owner also have this fov slider lora where it zooms out instead. I'll take that.

>>20079124

yeah smash cut is great

up and sleep

File: 157439192016489996_00001.mp4 (650 KB, 480x640)

650 KB MP4

>>20078488

here is result not sure what wrong. i will post WF in next post

File: WF.png (1.39 MB, 2470x1219)

1.39 MB PNG

>>20079849
>>20078607

>>20079850
settings-wise it looks ok to me. Only difference i had, pussy was visible on input image. I think that might make the lora work bit better. So try doing stuff in parts in prompt;

"camera pans down/zooms out revealing her whole nude body. She spreads her legs open. THEN she starts fingering her vagina. Fingers slowly sliding in and out of her vagina. sensual_fingering"

or just do 2 clips, first to reveal the body and the pussy (maybe use some pussy lora for this), second clip enable the fingering lora

>>20080052

appreciate you bag :)

>>20053516
I really appreciate that your workflow is sophisticated enough that I've been able to learn several things, but isn't the spaghetti puke 99%-custom-nodes bullshit you'll get from civitai.

So great job making a quality workflow without all the bullshit.

up and sleep

up

Anyone know the lora or how someone on a different thread was able to post those forced/aggressive cumshots where they grab the person down?

>>20081112
maybe this?
https://civitai.com/models/2023407?modelVersionId=2383114

>>20081112
if not what bag posted, maybe this one
https://civitai.com/models/2090098/cumhater-facial-cumshot-wan-22-i2v

up

File: Screenshot_20251114_00000(...).jpg (558 KB, 720x1600)

558 KB JPG

Make this whore expose her huge tits and suck cock please

I know my way around a computer pretty good and think i followed the setup pretty well but every time i try to run something i keep getting SamplerCustom PassManager::run failed and im not sure what to do to fix it

>>20081917
You didn't install sage/Triton.... Disable the sage nodes from workflow options

>>20082015
I’ve tried disabling that and it just turns it into an endless loop that heats up my gpu extremely fast until I force close it

>>20082035
Loop? Take screenshot of your whole workflow with console window visible

up

File: error.png (62 KB, 979x1024)

62 KB PNG

>>20082140
the top is where is the last thing that shows before it constantly loops the bottom about 3-4 times a second until i force close it

>>20082604
Show the workflow

File: Screenshot 2025-11-14 120111.png (439 KB, 1920x951)

439 KB PNG

>>20082620

>>20082646

Did you disable the ENABLE PATCH SAGE ATTENTION (HIGH) and (LOW)

in the Options node on the left side of screen?? check it to NO?

>>20082646
Seems new error to me....but seems it still try to load Triton/sage stuff. try disabling the nodes completely via options from the left side, or delete the sage nodes.

Just crossed my mind. Are you on rtx or amd GPU?

File: Screenshot 2025-11-14 124941.png (42 KB, 979x512)

42 KB PNG

>>20082691
did that and got this slightly different
i also did a fresh install to a different drive so see if anything change
>>20082702
rtx 2060

>>20082707
i dunno if sage 2.x is comp with rtx20xx cards....fast googling show that it might have some issue.You could try using older sageattention (v1.0) by going into your python_embeded folder, right-click to open in console and type

>python.exe -m pip install -U triton-windows
then
>python.exe -m pip install -U sageattention

or just delete the nodes....and edit the launch bat remove "--use-sageattention" flag

bumps before bed

I have issues that faces change radically. Any prompt to help prevent this or is it tied to models?

>>20083300
Usually caused by loras. I'd suspect anything related to faces first, but any lora with faces in its training data will do that.

up

>>20083211
Appreciate all your insights and helpful advice in these type of threads. I've been meaning to try out the triple sampler for a while.

There is a more refined version of the high lightx2v lora (1030) out there that some are getting good results with. I guess you're supposed to run the refined 1030 on high with the original 1022 on the low. Wondering if you've tested it out yet

>>20083810
no haven't tested, as i didn't even know new versions was out. Seems to promise better prompt coherence and and motion. As for low lightx lora, you can even use original wan2.1 lightx lora as low one

there seems to be some node painteri2v that claims to fix the low motion issues on when on lightx loras, need to test that too, seems promising, thou no flf support on it yet.....so much stuff to check out, so little free time.

guys, what are the loras you use that causes most slowmotions / face changes?

If I want to spend <5 minutes for a 5 second clip, what's the minimum GPU I should need?

>>20084349
Wish the leg aside pose transition didn't change face so much

>>20084558
40xx range...but it also depends the models you use, the resolution you want, the lora load etc...and if you installed sage, and if you use lightx loras.

>>20084559
what lora is that? need some for testing purposes

>>20084574
Not him but I have a 1080Ti
Am I cooked?

>>20084574
>40xx range
I'll wait until my next upgrade then. thanks

>>20084576
id say yes....you can make gens with lower end cards, just with lower models and resolutions (quality) if you have enough ram/vram to load stuff in.....one step might take at 5-10mins, but thx to accelerations loras like lightx, you only need 4steps to get decent results

>>20084578
but extra hundred in to get low-end 50xx with more vram, and get at least 64gb of ram on your next rig

>>20084574
It's a good one
https://civitai.com/models/2118407/leg-aside-pose-transition

Up

>General Wan 2: rentry.org/wan21kjguide
I've been trying to set this up for 3 hours and pytorch version never reads 2.7.1 it's always fucking 2.9.1

>>20085096
why you need to get it to 2.7?

>>20085105
The guide says
>Run ComfyUI with run_nvidia_gpu_optimizations.bat. Look in the cmd.exe console window and make sure pytorch version: 2.7.1 or pytorch version: 2.8.0dev is shown during startup. You should also see Enabled fp16 accumulation and Using sage attention. Make sure that every time you start Comfy, pytorch version reads either 2.7.1 or 2.8.0dev, otherwise fp16_fast / fp16 accumulation won't work.

>>20085109
>>20085105
>>20085096
I think chatgpt actually fixed the .bat though because it says 2.7.1 now

>>20085096
this thread is for Bag's workflow... but in any case, that guide is outdated.

https://rentry.org/wan22ldgguide

up

again

up

>>20084767
action looks good, but, its changes the face quite a lot even on low weights. Hope you trains it again

>>20086344
*he trains it again

File: compare_00011(1).mp4 (3.67 MB, 640x1536)

3.67 MB MP4

>>20083810
did some test runs...there is subtle difference 1022 vs 1030


>50 years old lady opens her white shirt, revealing her nude breasts:2).Then young muscular man with many tattoos appears in frame next to her from the right side. Lady looks surprised and happy as she grabs his penis in her hand, she yells "wow". Then she sucks the penis. blowjob deepthroat. Camera zooms in her sucking the penis. She looks at the viewer.

fixed seed: 1234
loras: oral insertion, jfj-deepthroat all at 0.9. LightX loras (high&low) at 1.0 (low lightx is 1022 version)
euler/simple
steps: 10

>>20086636
love that you pulled her image from the other thread, that middle one looks the best and most like her. love to see more if you're doing it

>>20054098

That's Ratajkowski's dance in Blurred lines!

https://www.youtube.com/watch?v=zwT6DZCQi9k

Nice!

Up

Bumps

up

Any suggestions or links to tweaking lora weights? All the info out there is conflicting. On civit there's zero info. Any links for wan 2.2 lora results for lightx weights not nsfw ones. Also maybe more importantly what thread had the audio discussion?

Would you suggest some good loras for facial cumshots please ?
There must be some tweaks that I didn't understand well since I'm getting cgi rendering or lights bleeding on my vids

Up

Page 10?! Bump.

>>20088304
I dunno what you mean. Generally the lora maker states the good weight range for their loras on civit. If not, default assumption is to run it at 1.0

>Any links for wan 2.2 lora results for lightx weights not nsfw ones
you can test this out yourself, make a prompt and set the seed (noise) to fixed. Make a gen, then lower or raise the lightx lora weights (don't change anything else) and run it again.

>Also maybe more importantly what thread had the audio discussion?
i dont recall if there was any big discussions...one anon said he found some mmaudio workflow and nsfw models for it, but never posted them.

>>20088437
f4c3spl45h and cum aesthetics is good combo. They tend to go bit "over the board". There are some more "realistic" cumshot loras, but i haven't used them that much to say an opinion.

>There must be some tweaks that I didn't understand well since I'm getting cgi rendering or lights bleeding on my vids

post screenshot of your whole workflow so i can see the settings. Sounds like you've set cfg too high or using some weird sampler/scheduler

up

Knowledge bump

is there a way to make the high/low noise samplers show the render being made in real time like a ksampler? that way i can cancel the render midway if its not turning out correct?

Bump

How are we with generative sound creation these days? Seeing grok have more or less sound by default is somewhat depressing.

>>20089865
i'm pretty sure is a way...i will take a look

>>20090339
for wan there is mmaudio (to add audio to pre-existing video) and then whole wan s2v model to make video using the existing audio file.

>>20089865
Go to Comfy settings
Go to "VHS" tab
Enable "Display animated previews when sampling"

May also need to enable previews in Comfy Manager? Not sure.

File: 3450989080980.png (76 KB, 653x396)

76 KB PNG

>>20089865
i did overthink myself....those are ksamplers (or based on it, in my workflows. And they have preview build in....SO

1) Go to manager and enable preview, either rgb or taesd

2) then go to comfyui setting and use search "sampling" and enable animated preview

that give you pretty rough idea whats to come....but like said, its very rough and blurry on the first steps. I usually just reduce the resolution to 240p....if it looks good, i rerun it with same seed and correct resolution

File: 3243455543.png (10 KB, 947x159)

10 KB PNG

>>20090551
>2)

>>20090555
alright thank you.

>>20090551
Good idea to reduce to 240 for testing, I know how to change seed in stable diffusion a1111 but on comfy the seed randomizer is greyed out and reloading the videod wf doesn't reload the seed. After using grok from the other r thread it's like n64 vs nes. I'm trying to get wan 2.2 working like grok. I hope that audio s2v works with the apprentice app

Oh I'm checking the workflow notes again it says when using 2.1 loras on low steps to raise the high noise, I'll try that because I'm getting slow movement on 24 fps

>>20091188
slow movement is caused mainly because lightx loras. there is pretty good fix for it. You replace the wan i2v node with painter iv2 node.

https://github.com/princepainter/ComfyUI-PainterI2V

I will be adding this for next version but im waiting if him will add last frame option into it

>>20091514

looking forward to it baggy

>>20091514
nta but this node is great, cheers

File: Joined_00015.mp4 (3.63 MB, 1536x560)

3.63 MB MP4

some sampler tests... 6 secs 8steps. These are random noise, so that affects the action of course. Main point is to compare quality

>camera zoom out. then girl goes down on her knees. She wears white skirt, white thigh high socks. Then girl lifts up her shirt, small nude breasts. Then nude man enters in the frame from left. Girl sucks his penis. Then camera zooms in on her face. Girl looks at the viewer while sucking the penis. nude breasts stay visible, she does not lower the shirt

oral_insertion lora @ 0.75, lightx @ 1.0

File: Joined_00016.mp4 (3.68 MB, 1536x560)

3.68 MB MP4

>>20091898

File: Joined_00017.mp4 (3.68 MB, 1536x560)

3.68 MB MP4

>>20091899

>>20091898
These are great. I always wanted to test them all but it's a pain in the ass figuring out what won't break it.

i think i bjorked mine.

i got drunk one night and started fiddling with pisswizards workflow. tried deleting and reinstalling but getting the same issues either failure to generate with error codes like this

"UnetLoaderGGUF
cannot reshape array of size 5642780 into shape (5120,5440)"

or it outputs something thats blurry and no clear definition.

have i missed a file on the reinstall or did i just miss a step like a moron? can send a screenshot of the workflow for you 4chan just wont let me atm

File: Untitled.jpg (230 KB, 1294x791)

230 KB JPG

>>20092827

anyone know or get freaked out by the legality of making this stuff?

>>20092827
could it be that your model files are incomplete? did you delete them and the re-downloaded them? does it run the high noise part ok?

>>20092960
yeah seems to run high noise ok but will get stuck somewhere along the way, sometimes k sampler, sometimes later in the flow. im thinking models are corrupted ive tried redownloading them again with the bat file but maybe thats made things worse?

is there files stored anywhere but the comfy ui folder? maybe its reading from that which is why a redownload isnt working?

ill just run another one and pull up the console for you

Wish there was a good i2v pussy reveal Lora. I keep getting these weird flaps of meat whenever she spreads her legs

>>20092960
i was wrong, it will run high until 50% then will ramp up my gpu and temps like mad but it hilights the low pass when the error pops up, also here is the comand line stuff from the prompt i just ran

rompt executed in 205.98 seconds
got prompt
WARNING: PlaySound.IS_CHANGED() missing 1 required positional argument: 'self'
Requested to load WanVAE
loaded completely 801.7088623046875 242.02829551696777 True
0 models unloaded.
loaded partially 1784.7911377906798 1784.7911376953125 0
Patching comfy attention to use sageattn
100%|| 2/2 [04:36<00:00, 138.33s/it]
Restoring initial comfy attention
!!! Exception during processing !!! cannot reshape array of size 5642780 into shape (5120,5440)
Traceback (most recent call last):
File "D:\comfyai\ComfyUI_windows_portable\ComfyUI\execution.py", line 496, in execute
output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, hidden_inputs=hidden_inputs)

>>20092960

^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "D:\comfyai\ComfyUI_windows_portable\ComfyUI\execution.py", line 315, in get_output_data
return_values = await _async_map_node_over_list(prompt_id, unique_id, obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, hidden_inputs=hidden_inputs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "D:\comfyai\ComfyUI_windows_portable\ComfyUI\execution.py", line 289, in _async_map_node_over_list
await process_inputs(input_dict, i)
File "D:\comfyai\ComfyUI_windows_portable\ComfyUI\execution.py", line 277, in process_inputs
result = f(**inputs)
File "D:\comfyai\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI-GGUF\nodes.py", line 152, in load_unet
sd = gguf_sd_loader(unet_path)
File "D:\comfyai\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI-GGUF\loader.py", line 55, in gguf_sd_loader
reader = gguf.GGUFReader(path)
File "D:\comfyai\ComfyUI_windows_portable\python_embeded\Lib\site-packages\gguf\gguf_reader.py", line 182, in __init__
self._build_tensors(offs, tensors_fields)
~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^
File "D:\comfyai\ComfyUI_windows_portable\python_embeded\Lib\site-packages\gguf\gguf_reader.py", line 364, in _build_tensors
data = self._get(data_offs, item_type, item_count).reshape(np_dims),
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^
ValueError: cannot reshape array of size 5642780 into shape (5120,5440)

Prompt executed in 291.73 seconds

>>20092979
go to comfyui/models/unet

delete the low noise q8 model.....download it manually (or use the bat)

https://huggingface.co/bullerwins/Wan2.2-I2V-A14B-GGUF/resolve/main/wan2.2_i2v_low_noise_14B_Q8_0.gguf?download=true

that error seems to come up when it tries to load the low model, so its incomplete/corrupt....i think.

>>20092967
all files comfyui portable are store in comfyui folder.....my bat should auto-resume broken dowloads, but maybe something went wrong at some point for you

>>20091898
Working good for me, trying 8 steps instead of 3 or 5. I do like to use old loras from 2.1 that I've been using since febuary. AI seems to understand the fluid dynamics of milk spilling out of her mouth without puking or a cum lora.

I'm having issues with skin and nipple color. AI told labeled skin and nipple colors as medium pink and peach complexion. Not working well without a lora or maybe conflicting. Also pussies that look like they drop a dangling tumour. I found this

>Key Findings
1. Clarity and Specificity Are Crucial
Precise Anatomical Terminology: Using correct and specific anatomical terms significantly improves the model's ability to render accurate depictions. For example, terms like "labia majora," "labia minora," "clitoral hood," "vulva," "erect penis," and "vagina" provide clear guidance.

Explicit Role Identification: Clearly specifying the genders and roles of participants (e.g., "an adult man and woman") helps prevent confusion and misrepresentation.

Detailed Descriptions of Actions: Providing step-by-step descriptions of actions and interactions guides the model in depicting dynamic scenes accurately.
https://www.reddit.com/r/StableDiffusion/comments/1hvw8mo/user_discretion_advised_enhancing_content/

File: peach complexion medium p(...).jpg (193 KB, 784x1168)

193 KB JPG

This is Groks image sample, the videos look just as good. Hard to get these results in wan and a more complicated prompt seems to stifle the model. Might be lighting issues I'll try that. Just don't want to have to make my own loras seems daunting. Although I might try those rent a data center computer sites for making loras or even wan 2.2.

>>20093166
groks output look really good and consistency is superb on level for sure....but the lack of the nsfw =/

It's like like inventing a new tech that has somekind of screen attach to it, 1st question is "can it run doom"
same goes with software...1st question in "can i make porn with it". If yes, instant success

File: Screenshot_20251121-155740.png (572 KB, 1080x784)

572 KB PNG

https://youtu.be/VXz0KgNjK-0

>>20091898
how do you change it?
is it via the noise thingy?