Thread 19799683

Many have asked, so;

https://www.mediafire.com/file/i1pslzko0hr7hy1/1stAIdbag_v1.json/file

It's my first workflow from scratch, so bare with me, gimme tips, fixes, let's revision it better. I assume you would know something about comfy, if say there are some errors/bugs/wtfs. i'll try to help (i'm kinda comfy-noob myself) where i can. I don't wanna guide how to install comfy, those tutorials are just one google search away

it's dual workflow in many ways, you could use diffusion or Q (gguf) models, it has option to use single sampler or dual sampler setup (later needs lightx's self force lora).

Maybe keep this in general comfy-faggot-q&a thread...

THIS IS NOT A REQUEST THREAD!!!!
posting your bitch in here will make i won't do it, ever
THIS IS NOT A REQUEST THREAD!!!!

Oh lawd, he splattered Harold.


At least it's milk, though.

bump this shit time to time

florence is an llm, right?
does it really improve the gens?
will it run on a 12gb 3060 with 32 ram?

>>19800023
>does it really improve the gens?
for 4chan needs, no =) might be helpful in same cases, like multiple girls in same pic, florence can give them some reference words

>>19800499
>will it run on a 12gb 3060 with 32 ram?
one anon test ran it with 8gb rtx card, worked ok. i added blockswap node into it, so enable that if OOM

can not install comfyroll studio
not trying to be an asshole but...
do you REALLY need all those custom nodes?
are there no default nodes that do similar things?
what would be the BARE MINIMUM workflow to get your results?

anyway, glad it works for you, your gens are top notch.
and thanks for sharing.

>>19799683 (OP)
You are a legend for posting this, thank you wizard king <3

>>19801963
>open comfyui manager
>install missing nodes
>good to go
Sounds like a you problem.

bump because learning is caring.

>>19801963
you could replace some node with core ones sure, custom ones do shit better

and i'd say 8gb / 32 gb ram gets you going on on q8 model, might need to enable blockswap node

>>19799683 (OP)
Hey Ald Bag. I've read somewhere that ComfyUI can be dangerous if you install some nodes that can run malicious stuff to get passwords/accounts or virus. Did you have any problems with that or heard similar stories?

>>19803753
Not him but my antivirus program detected alot of threats after getting comfy and loras. Better to run a virtual machine.

I use Windows Defender and nothing's been flagged when I installed it

>>19803756
This message brought to you by motionmuse

>>19803753
there are no nodes in my workflow that aren't avail to install via manager (aka safe)....and yeah sure they could do that...be on your toes when when somebody asks you install "node" via git url

and even then, by default, comfyui wouldn't install it due its security system

>>19799683 (OP)
so clean.. cheers anon <3

>>19802144
salty bitch haha

bump

this is made with 20-step single sampler

>>19804520
and this is 8-step dual sampler. Quality-wise not much difference, but the movement is not there

dual sampler is ok for tittydrops, cumshots...on loras that has much movement, stay on single sampler

>>19804525
What is your CFG for these?

>>19804541
3.8 on single sampler, 1 on dual.....it has to be 1 on dual setup due acceleration lora used

>>19804568
You'll know me as the wizard no one likes because I don't put out slop, ever. Since you are trying to help people, I want to offer help to you, but only if help is wanted.

>>19804619
>>19804568
I'm sure there other reasons wizards don't like what I do, but it all boils down to jealousy as far as I am concerned. Anyway, I have some tips if you're interested.

>>19804638
If you're that one that uses Krita for nudifies, he's an artist and I love observing his work.

I don't know why anyone would hate an anonymous wizard anyways. Beggars gonna beg, and any scrap should be appreciated.

>>19804694
That's not me. Along those lines, I think I'm the only wizard dropping small tits these days. I drop one pair and suddenly requests for tiny-titty girls fill the thread, apparently its a novelty.

>>19804710
That's vaguely my irritation; people don't know what is possible and just want something. (Which cn be understandable when you can't do anything, so you beg for anything) Specific requests, people, use your imagination.

I get default prompt and reliable results, and partake of it, but I am here because I want to see new things and ideas and try things, fuck your "dealer's choice,'' I don't know this woman or you and anything I make will be followed with "That's great, but...''

IF YOU KNEW WHAT YOU WANTED WHY DIDN'T YOU ASK IN THE FIRST PLACE?

>>19799717
kek

Hey wizards, I have a different request to make, if it isn't too much of an issue. I used to mess with SD and gradio webui until the time around when Civitai implemented their currency thing, and am just coming back after seeing that i2v generations became very interesting. I got Comfy set up by following the rentry guide while never having messed around with it, and did my first generation which completed in around 50 minutes, as the guide suggested it would.

I have a few questions, do y'all that generate stuff here use the force method in the guide that cuts that generation time? The quality loss is negligible?
I'm currently running a 16/16 RAM/VRAM setup, while it generated, it's clearly at the allowed limit. Would it be best for me to go with Q6 instead? Is the quality drop very bad?

The workflow posted in this thread was very helpful to visualize how it works, I'd be interested in seeing one for deepthroats too. Any wizard feeling kind to share one of those too?

>>19804823
Post an image to work with and I can show you examples

>>19804826
I don't wanna risk anyone thinking this is an impromptu wizardry request, so if we need an image let's use this adobe stock picture for reference, already resized for 480p, the one I'm using. But I'm really interested in the answer to the questions too, any tips wizards wish they knew when first getting into it to help me navigate will be very welcome.

>>19804694
Not sure there's just one of us, but thanks anyway. The secret is relentless cherrypicking from dozens of gens; also, the Krita plugin makes it easy to fix small problems with an inpaint.

How would I be able to do this with a laptop top?

>>19805223
1) Buy a gaming laptop with a sufficientngraphics card.
2) Setup a gaming PC that you can remote access with a regular laptop.

>>19804710
Gawd dayum, do you mind sharing your workflow to create that masterpiece?

>>19804520
Can you please share your prompt?

>>19805374
Ok, so laptop by itself is pointless?

>>19799683 (OP)
send full prompt list please, the ones in the prompt stash saver node <3

>>19805599
There are laptops out there with sufficient GPUs. Depends whatyou prioritize, convenice, processing power, or cost.

Or buy processing space online, that would be another option, but I don't know how thatworks.

>>19804720
Omg this...bugs the shit out of me too. "Can you do anything on this pic?".... Do something on the the pic then the anon goes "nice but I would like to X in the pic then Y happens".... So you kinda wasted my time as you can't type in the op posts what you really want.

Note that Wan makes 5-6sec clips..... That isn't that long time.... "can you make tittydrop as she bj two black dude, then they cum on face, then one of the guys starts fucking her..." ain't going to happen in 5secs.... Has to be good image for Ai Wiz faggots merge all that together

bump

>>19799683 (OP)
Glad to see there are other workflow perfectionists who align every single node with meticulous detail.

Bump

up for safety

>>19799683 (OP)
KEK

harold got it good

Bump

>>19805932
Wiz pls: Handjob cumsht facefck blwjob etc wiz pls

ffs people, just pick someting

>>19799683 (OP)
Hi Ald Bag. I got 2 questions.
1) If I load the json template you provided, it's quite easy to change from cumshot to facebook if I change a "lora" only and the prompt? Or do I need to fiddle more?
2) It's possible to have good results with a RTX 3050 with 8GB VRAM?
Bonus) Do you think Tensor.art NSFW will come back or is nuked for good?

>>19809251
>1) If I load the json template you provided, it's quite easy to change from cumshot to facebook if I change a "lora" only and the prompt? Or do I need to fiddle more?

well, no need to more than prompt it correctly (have certain trigger words in there) and load the lora.

>2) It's possible to have good results with a RTX 3050 with 8GB VRAM?

Yes, it will be slower ofc, but that card should be able to load Q8 models. Having more ram helps on this loading process. If you use dual samplers you will get decent output even in 4-6 steps, but like i posted the examples above, movement will go down.

>Bonus) Do you think Tensor.art NSFW will come back or is nuked for good?

I dunno, i've never used tensor.art. But i assume it's the same shit like civitai faced. And what i assume what went down, partners who deals with their payment options go and went "No, we don't take part of this nudify things, take these options down or we go"

Hey, thanks for this workflow, I’ve loaded everything needed/listed in the post but it crashes whilst generating due to an issue with GGUF node, any idea how to fix it or edit the workflow to not need that node? I’ve tried all the versions available and it always has this issue
What one are you using?

post the image or the error msg and what models you have loaded

>>19809375
This was on a fresh instal with the triton/sage added

Error:
UnetLoaderGGUF
Expected str, bytes, or os.PathLike object, not NoneType

Models
Every one you list in the textbox, v8 of the text encoder, all in the correct file

Nodes
All the ones it told me to download when I loaded the workflow + the manager and the teacache from the triton/sage guide.

ComfyUI-Manager - v3.0.1
ComfyUI-Custom-Scripts - v1.2.5
ComfyUI-LayerStyle - v1.0.90
ComfyUI-GGUF - v1.1.1
rgthree-comfy - v1.0.2507112302
ComfyUI-Easy-Use - v1.3.1
ComfyUI-KJNodes - v1.1.3
ComfyUI-Florence2 - v1.0.5
ComfyUI-VideoHelperSuite - v1.6.1
Comfyroll Studio - nightly
ComfyUI_essentials - v1.1.0
ComfyUI-TeaCache - v1.9.0
ComfyUI-Frame-Interpolation - v1.0.7
ComfyUI-mxToolkit - v0.9.92
ComfyUI-wanBlockswap - v1.0.0
Prompt Stash - v1.2.1
WAS Node Suite - 1.0.2

>>19809419
well i assume from the error it doesn't find the the model in name/path gguf loader....

even it shows up right in the slot, that file points the file in workflow creator's computer, therefor every time you get new workflow file, you need to reselect the files in every loaders to the files point in your computer.

you could try using diffusion models, just click the select-switch on the left. once diffusion model nodes are loaded, gguf ones should be bypassed(pink) ....and vice versa

Bump

Is anyone having problems with NAGCFGGuider/ComfyUI-NAG not being installed with the ComfyUI Manager? Is there a manual installation or I'm doing something wrong?

>>19799683 (OP)
Is anyone having problems with NAGCFGGuider/ComfyUI-NAG not being installed with the ComfyUI Manager? Is there a manual installation or I'm doing something wrong?

Installed ComfyUI, updated, then the ComfyUI Manager, updated and all missing nodes was smoothly besides this one.

>>19809970
is there more info on logs why. You could try manual. go manager install via git url:
https://github.com/ChenDarYen/ComfyUI-NAG.git

>>19809983
Manual install worked! Thanks bag. Couple of questions about the whole workflow to work:

I've download all the files you've put on the notes inside the workflow. All besides the last item called Sage/Triton install. These ones are a absolute must have to work?

It's easy to find these loras: wan_cumshot_i2v.safetensors
wan-thiccum-v3.safetensors
Or I need to search on private trackers or something?

>>19810032
you can remove all the loras from the list and build your own lora list....there are some huggingface sites that have most of the nsfw loras backed up, also civit archive siteare worth a check

>>19799683 (OP)
First of all, thanks for this workflow, Bag. I had some trouble installing being a beginner, but now is working. I've attached the results from the two methods in the workflow. I'm using a RTX 3050 8GB VRAM with 16RAM, in case anyone wants to try it out with lower end cards.

After using it, I'm curious about some stuff:

A) It's just me or the single sampler feels more realistic (cum) than the dual one? Only downside it's that the girl open the mouth right at the last second. And single sampler takes double the time (from my tests)

B) How adaptable is this workflow is if I just swap "loras" to insert ones like this:
https://civitai.com/models/1340403/wan-21-deepthroat-blowjob-fellatio-oral-lora-t2v-i2v
or this:
https://civitai.com/models/1350447/wan-cumshot
I would need to update the "vae" too?
And in that case, swapping lora and vae, everything should go smoothly and there is no need to install more stuff, right?

C) What makes the results better besides a greater GPU? If I load like 5 loras with the same subject (cumshots), this improves a lot or I need to improve the prompt to get a sexy result (cum being thicc, cum dripping in the tongue, realistic cock appearing, etc)

Again, thanks Bag! You're definitely a great wizard!

>>19810700
what are your total gen times? I'm sitting at 10 minutes for a 7 second video on dual sampler (haven't fucked with single)

3060 12GB VRAM 32RAM

>>19810700

A) are those runs from same fixed seeds? Anyways, single sampler most likely give more realistic as its runs more step (samples stuff more) Also what sampler and scheduler you use makes difference, uni_pc/simple is safe combo, euler/beta works, i use dmmp_2/sgm_uniform on single sampler

B) pretty much. I dunno why you talk about swapping them out, you can just load all your loras into the loader, then just enable the ones you are going to use (see pic). And there is only one wan vae-file far as i know, so dont touch that. Once you have the models and nodes, no need to install anything. Now you only need loras and prompts

C) i dunno if there even is 5 cumshot loras out there, but that would be bit overkill. orig cumshot lora with thicc cum lora work most of the time. Qualitywise, better quality kinda needs more steps, more steps needs more time

.....and remember, you rarely get the output you had in mind on first try. Sometimes it takes several tries, tweaks into the prompt and lora weights, etc, to get something good

>>19810707
I didn't do many tests yet, but the window with single is like 30min and dual is like 15min or so. (5 seconds of gen)

>>19810717
Hi Bag. Do you mean on the RandomNoise seed? I didn't change that and leave as it came. I'll try looking into the combos you mentioned.

I've tried to search for all these loras you have and some of these I can't seem to find it (by G, Yandex or Bing):
wan-thiccum-v3.safetensors
wan_lezkiss.safetensors
wan_bukkakewalk.safetensors
The rest I could find it and I'll leave a file with the URLs on the next post. Would be awesome if you could provide the path to download thiccum-v3 if you could.

I'm sorry, you're right about the availability and how many you need to do the magic. Next tests I'm going to try with more steps and changing some of the prompt to see how it goes.

Loras Download Pack - URLs:
https://www.mediafire.com/file/qi3ep375k8yi6bw/WAN_Loras.txt/file

>>19804520
Also, would be amazing to know what kind of settings/prompt was used here. wan_facefuck only or combination of wan_undressing too? In my opinion, one of the best gen in the threads of /r/.

>>19810860
lol, except for the morphing face and massive color shifting, right?

>>19810825
https://civitai.com/models/1643871/wan-thiccum

https://civitai.com/models/1503991/wan-french-kissing-v1

>>19810865
I'm not saying it's the Mona Lisa, but it's one of the best gens I've seen here, in terms of how it happen and the movement.
>>19811471
Thank you. I should've searched more on civitai

>>19810865
why so salty, ran out of credits on your spam site?

>>19799683 (OP)
How do I queue multiple different photos but with the same workflow/different workflow to let it running?

>>19810825
https://civitai.com/models/1364982?modelVersionId=1542133

bukkakewalk (i rename my loras, but even orig name its hard to find nsfw loras in civitai, as some set to private). Note that this not cumshot lora but helps to keep cum on face on longer gens, keept in on lower weight thou

>>19810860
i'd say make one action/lora at time, 1st undress, then save last flame, open it up as input, disable undress/enable facefuck, run, then join the clips. You could try to prompt all stuff to happen, but 5-6 sec isn't that long time

>>19810865
purpose what to show difference on single/dual samplers, not to please your ai-preferences. But yeah, poor baits are poor

>>19811912
just change the input and press run....if one is running, next one goes into queue

up

Bumps

Bumpa

Thanks for sharing your wisdom, Grand wizard

Thanks for the replies, Wizard Bag.
>>19804520
You were right about using the single sampler with blowjobs/deepthroats. Double sampler the movement is not there at all (from one and two tests I did)

From what I've saw, it's better to have 2 loras of the same category to complement each other with different values than just one full power.

Let's say I liked the prompt and loras, to maximize the quality I need to jump from Q-6 models to Q-8 right? What about steps, going overboard like 40 will deliver a good to awesome result?

Also, how do you do quick tests with loras? Dual sampler with very low steps? I'm asking because would be cool to try it out each one to see how the effects are without waiting too long.

>>19814391
yep, jump from q6-->q8 is big step up on quality, also you could try fp8 diffusion model. On sampling steps....i usually go with 20-25, going beyond 30 steps, naah, but sure, make a test, run one with 20 step, run it again with 40 step, you decide is it worth the wait

On matter on testing out stuff; changing prompt, lora weights, cfg, or steps. ALWAYS set your seed (noise) to fixed. This way it tries to make the same output everytime -/+ change you made to some setting. And sure you could use lower steps on your tests, but i'd test lora's on single sampler, like you noticed, full potential is lost on some loras on dual sampler

bump

Bumpity bump

>>19814423
So, I think I've messed up. I tried loading q8 models and my GPU of 8 VRAM couldn't handle it and it crashed trying it (I know, silly me for giving it a try).

Worse part of that, when I'm trying to create with q6 again, its taking too long when it gets to "KSampler" stage and the second bar below doesn't start (both single and dual). It's taking like 20+mins. Half with "Attempting to release mmap" and another half with starting "Patching comfy attention to use sageattn" Something is telling me the files are corrupted after trying q8 and crashing it. Even more so after it's getting real slow after generating. Before I could even use Photoshop/Illustrator open while was creating.

If I want to start clean again and "uninstall" ComfyUI, Triton and Sage Attention completely, I just need to delete the ComfyUI folder right? I hope it doesn't store settings outside the main folder.

>>19814423
So, I think I've messed up. I tried loading q8 models and my GPU of 8 VRAM couldn't handle it (I know, silly me for giving it a try).

Worse part of that, when I'm trying to create with q6 again, its taking too long when it gets to "KSampler" stage and the second bar below doesn't start (both single/dual and multiple tries). It's taking like 20+mins. Half with "Attempting to release mmap" and another half with starting "Patching comfy attention to use sageattn" Something is telling me the files are corrupted after trying q8 and crashing it. Even more so after it's getting real slow after generating. Before I could even use Photoshop/Illustrator open while was creating.

If I want to start clean again and "uninstall" ComfyUI, Triton and Sage Attention completely, I just need to delete the ComfyUI folder right? I hope it doesn't store settings outside the main folder.

>>19817005
if you do so, back you your models folder at least somewhere, so you dont need to dload all shit again. there isn't much settings in comfy to store. getting nodes is pretty fast so dont worry about them

q8's should run on 8gb cards...if your comfy hangs up / get fucked up, dont try todo something else on the same run.....because at the time, comfy is fucked up. shut down, and restart. On lower cards, max time 4-5 sec, max width 480, also you could try enabling blockswap

when you are using bigger models like you moving from q6->q8 it will hog up a lot your ram and vram, i would they are full load. i don't know what your hardware is, but maybe not launch any heavy programs while running

>>19799683 (OP)
Hey Bagman, this isn't from your workflow, but pretty much the original from the rentry guide. But I figured I'd try to see if you recognized it anyway.
Do you know what causes the flash and weird artifacts at the start of this vid?

>>19817042
many things....but my guess is that you are using facials lora too high (or maybe some other lora). lower it to 0.5 (this also helps to keep orig face) or something like that and try again. you could add some other cum loras with lower strength with it. You could alse try keeping all the settings same but keep it 81 frames (5secs)

I might be wrong but this has todo something with how the lora was trained, and/or what size the dataset images were used, usually they are 512x512 or 1024x1024, but if there is mismatch there somewhere that might cause it.

...thats pretty good gen btw

Bump

bumps before ZzZ

Bump

Update workflow little bit

-Dual sampler is now just single sampler using lightx self-force lora. Does the same job, same quality imho
-Replaced the NAG node with different one, since quite few had probs installing the old one
-Replaced random seed node with other one
-Replaced the clear vram node with other one, when swapping two sampling methods around, its good to clear model after run.
-Removed brigtness/contrast node
-Added blockswap to options for fast enable/disable
-Removed colormatch node from bypass list, i think there might a a bug on the node as when bypassed, it stops the process, i tried using any-switch between it and raw vae-images, but it would output input image as output video. Workaround for now is to set colormatch strenght to 0 if you dont wanna use it
-Added separate sliders for lightx/normal sampler steps, so you dont accidentally run normal sampler with low steps or lightx sampler with too high steps
-Added cfg into options for normal sampler method (lightx cfg must be "1")
-Added links for fp8 diffusion model and clip
-Updated lightX lora link. There now i2v based version. This new workflow uses it, it's important to get this lora in order self-force sampler to work correctly! Movement seems better now! So get it

Workflow 1.1: https://www.mediafire.com/file/p6atp5r6m5zjwx5/1stAIdbag_v1_1.json/file

>>19820111
Did a test and made a fresh install of comfyui_portable using this fork of it:
https://github.com/YanWenKun/ComfyUI-Windows-Portable/releases
Once all unzipped, rename folder "python_standalone" --> "python_embeded"
Then download:
https://www.patreon.com/file?h=124253103&m=441360653
Place the bat in root of comfyui_portable and run it, install triton/sage/and libs, no need for kijai nodes and wan stuff
Once done, rename python folder back to "python_standalone"
Launch comfyui, open my workflow, you will get popup window stating missing nodes, close that window, go to comfy manager instead, click install missing nodes, select them all and install / reboot comfy
All should be installed now, re-select model/clip/clipvision/vae/lightX lora file (so they point location in your computer)
Open image, load up your some loras in powerlora loader if needed, prompt it up and run
All went good, everything installed on first try without hickups

>>19820111
this is 8-step run with new i2v self-force lora, As you can see, movement is much better now even on deepthroat runs

>>19820113
>>19820111
This is awesome! Got it running

>>19820111
Amazing stuff! Gonna test the new version when it's possible and report back. Thanks!!

>>19820111
>>19820113
If you had to summarize the changes, would they primarily be:

1) Better Quality
2) Better Gen Time
3) Better ease of installation

>>19820111
this workflow works well for me, sometimes under 2 min gens with 2 loras.

bump

Bump

Bumps

bumps before bed

>>19820113
This works mostly great, but every time I run, I get "KSampler RuntimeError: PassManager::run failed".
I've tried switching to Q4 unet/cliploader, enabling blockswap, restarting comfyUI, reinstalling triton/sage/libs, but no luck. Could just be my card, running a pretty old NVIDIA GeForce RTX 2070.

>>19820111
Thank you, almighty aid bag. Was able to get the generations going, each takes around 10 minutes but they're coming out pretty good.

>>19822577
Could be this is sage/triton releated issue on older rtx cards....you could try

a) disable sage on the node
b) do fresh install (backup models folder before you do) DONT install sage/triton

Bumps

>>19820143
I see you got last frame. Can you share that with the metadata? Just wanna know your prompts for this and the loras strength.

Does anyone know of a proper img2img ComfyUI workflow that allows for undressing/nudifying etc.?

>>19823176
any inpainting one would work, unless you're looking for one of those ones that automatically detect clothing.

>>19823176
Go to civitai, search for intapaint workflows. I suggest sdxl ones. But you could use flux ones too but on my opinion sdxl has better NSFW models. For sdxl try biglove v3 checkpoint as base model

>>19822792
Disabling sage sorted it, banging work all over mate.

>>19823372
Post something you made up with 2070 what time it took, also what model and how many steps

Great workflow. it's 5000x better than what I was using. What is the purpose of last frame? Is there an automation that can extend the video or do you just copy and past that into the input image and run it again? Also, how do i know if I am dual sampler or single sampler

>>19823429
>What is the purpose of last frame? Is there an automation that can extend the video or do you just copy and past that into the input image and run it again?

exactly that
there is no dual sampler in v1.1 (both options are single sampler ones)...use v1.1 that with new lightx lora, much faster and better

>>19817039
Reporting in after having those problems I mentioned before. I deleted ComfyUI, uninstalled GIT and did a clean install of everything, including the GPU driver. After some tries, the slow/lagging of everything finally stopped and happens just when it's processing a lot. It's working again. Previously the step of "TeaCache: Initialized" wasn't happening everytime for some reason.

And it's great that you've updated the workflow. Gonna test in the future for sure!

I've got a couple questions if you don't mind, Bag:

1) There is some photos that take like 35min to be done. While there some others (like picrel) that took like 2 hours overall. What gives? I've noticed that there is a node Image Resize, but I took the time to resize all the photos to 480px in width to help a little. Maybe the 2hrs happened because of the quality of the photo, face features, background or the image being a bit vertical? Maybe I'm tripping and this happens to everyone and it really depends from photo to photo.

2) Your best results comes from single, double or triple combinations of loras? And also, did you noticed a difference of time having just one or multiple activated?

3) If I tried to use a different model and lora likes this one:
https://civitai.com/models/255972/gag-gagged-ballgag-spidergag-ringgag-bitgag-sdxl
I just need to swap some of the vae/unet in the workflow and it's possible to work?

Thanks Bag, much appreciated!

>>19823601
I think your biggest problem was/is: you are trying to use wrong models and wrong loras. My workflow is build for WAN (14b), you need wan models/clips/loras.....You are trying to use SDXL stuff (which is for image creation) on video generation workflow. No wonder it hangs up

1) time it takes has nothing to with image size you load up as it resizes it before it does the sampling. There is width slider on my workflow, that will determine the video size. Higher the width = longer it takes, on 480p model, keep it under, 600px, if you want better quality/bigger resolution on output, you would need to use 720p model but that kinda double the time on sampling.

in nutshell; time it takes comes from :::
model quality (480p or 720p and/or is it quantized aka Q model or diffusion model) + steps + output width
(sampler and scheduler play some part too, what those above play major part)

2) Depends what i do. Cumshots usually 2-3 loras, some thing just one lora. Sometimes i make things in smaller steps aka make 4sec clip with one lora/action, then continue by adding another and maybe removing the first lora.....its just test and play

3) This what i said above. ONLY WAN STUFF. Use the links i provided in the workflow and save them where they should go. Thats all you need to get shit going, if you need loras, get wan 14b i2v ones. and there is no way it takes hours to make a vid, even older cards can push 5sec clip out in 10mins

>>19823601
My bad, I should've mentioned in the last question that I've never used different models with the workflow. I'm strictly following the workflow as it is and just experimenting different values/prompts. I was just curious if it was possible to run other stuff besides WAN. Silly me, of course not.

About taking too much time, I don't know, maybe my GPU (RTX 3050) doesn't have that much power or 16 RAM is kinda low. But I'll gladly take most of them for 30min or so. The workflow is great. I really want to test the updated version when it's possible!

One thing I was thinking and testing, if the loras are trained on cumshots that is mostly from girls with the eyes closed, trying to put a prompt to keep the eyes open is gonna create a conflit and not always work, right?

Thanks for your time and sorry if the padawan questions was too much. Next time I'll limit a question per day, lol.

>>19823780
don't be sorry for asking questions. I gladly help when/if i can. You should get pretty good results on your card within 5-10mins if you use lightx lora sampler method, as it can run shit on 4-6 steps.

on cum loras.....yeah, that tends to happen as most of the images/videos they used in the dataset when training the lora, girls have their eyes closed. There is no conflict per say.....but lora wins your prompt. You could try;

on positive: Girl keeps her eyes open and looks the viewer.

on negative: closed eyes, cum on eyes

You might also need to lower the lora strenghts, but not too much, because then there is no cum either.....like i said it test and play. Sometimes it might work, sometimes won't. On cum loras high strength tends to change the face too. So it's up to you to find the golden line

and don't ever think you can get things right in the first runs you make.

>>19823780
Prompts can beat loras, within reasons. As per above, adding closed eyes to negative prompt and open eyes to positive prompt can help. I noticed this too, and that none of them were smiling. It takes a few to many gens to get what you want, and even then, sometimes it is just luck of the seed.

And I'm running 3060 12GBVram and 8x8gbs RAM and get 10 minute gens (for 5-10 seconds of generated video). VRAM has more to do with the actual gen speed, whereas ram is more about loading the different models and steps, unless a smarter person wants to correct me.

It taking 30 minutes is eyebrow raising, and yeah, when a gen takes 30 minutes, it can be painful to troubleshoot prompts and lora weights.

30 min

>>19823864
>VRAM has more to do with the actual gen speed, whereas ram is more about loading the different models and steps, unless a smarter person wants to correct me

this is correct....if you on low RAM, you might want to change models/loras/lora weights, as that causes shit to rebuild/reload into ram-->vram, before anything starts moving, Im having 32gb, and sometime i feel i spend more time on loading samples than actual sampling. Planning to buy other two sticks as ram isn't that expenssive

>>19823880
>you on low RAM, you might want to change models/loras/lora weights

*you might NOT want to change models/loras/lora weights after every run, as that....

Bumper

Hello all,
I'm succeeding to make successive short clips (5s) by using the last frame of the previous clip to regenerate and so on.
The issue that I have is the face is quite changing the more generation I'm doing.
Does anyone have a solution or a faceswap workflow that would resync the original face on all the frames? Thanks ! Cheers.

>>19824647
You need pretty sharp image as last frame. if its blurred/half missing, you are kinda giving the keys to ai. I'd say 2-3 i max runs before unvalley on horizon

you could run it throu facefusion....again depends on the action..wont for on blowjobs and big big no on cumshots

Any tips/template prompts for deepthroats? Mine doesn't come quite right... is the "jfj-deepthroat-v1.safetensors" on the list the recommended one? Does it need to be used in conjunction with facefuck or should have good results by itself?

>>19824884
Is there any meaningful difference between FaceFjsion amd FaceReactor? I still ise facereactor with thr nsfw mod

bump

>>19824904
is the "jfj-deepthroat-v1.safetensors" on the list the recommended one? Does it need to be used in conjunction with facefuck or should have good results by itself?

Should work fine by itself, you could add facefuck on lower strength (0.2-0.4) to aid the insertion. You can get example prompts on civitai's lora page, set filter metadata only

>>19824963
FaceFusion is seperate ai program. Create your clips, join them up, then run it throu FF using orig pic as source, your video as target. But like said wont work on cumshots or blowjobs that well

So this does NOT need pytorch?

>>19826707
no, it's disabled by default, if you have pytorch version that works, then you could try to enable it, but i think timesave is pretty minimal on sampling contra time it takes to load

Total noob here.
What are the system requirements to run this thing decently?

>>19826815
I'd say rtx 30xx+ card (8gb vram or more) , 32gb of ram (16gb would work but it will be much slower loads)... But seems like ppl got it running on lower setups

Tested on v1.0 Workflow - Prompt example from civitai lora page: "A woman performs oral sex on a man. She moves her head up and down while she sucks the shaft of the penis. A small amount of white cum drips down the penis. She maintains eye contact with the viewer. The view is POV."

Like it was mentioned before, the dual sampler is not that good for movement of the BJs. But it has the advantage on some details. The small amount of cum dripping appears on dual but not on single, which is kinda strange.

I'm trying to getting close to third example from the lora page, but something tells me you need the start the first frame with a photoshop dick already in place to have similar results. Or you just need to experiment with prompts/values to let the lora do it's magic?

>>19826918
try the 1.1, much better motion on that. And i dont think you would need penis in the frame, make your first clip short, where "pov shot, woman kneels down between man's legs. Man has big erect penis"....using just penis lora, no bj/deepthroat ones

up

I was trying to set this all up on my own but I might really need your workflows. Shit's not as easy as I thought.
I'm grateful this thread is still around.

up

>>19830179
Great result. Would be awesome if anyone who post examples/test with the loras used or prompts. This way it's easier to refine stuff together

>>19830179
Nice!

bumpity bump

>>19820113
Hey just a heads up, I followed your instructions and the python folder was already named "python_embeded" when I downloaded ComfyUI portable. When I renamed it to "python_standalone", ComfyUI refused to load after running "run_nvidia_gpu.bat"

>>19832436
Also, ComfyUI Manager doesn't seem to be installed automatically after following the instructions

>>19830896
I didn't make it desu, but here is the metadata:

{"image_mode": 0, "prompt": "blowjob, deepthroat, rough deepthroat, a sexy woman. Close up view of woman performing deepthroat on a penis. Static camera. Pov perspective, top down angle. POV, You are a naked man recording a woman in front of you, your legs, stomach, belly button, lower body appear from the bottom of the frame, your big penis prominently displayed before the woman. Woman puts your penis into a her mouth, her lips wrapped tightly around his penis. Man grabs womans head with his hands and pushes it up and down his penis. Man pushes his penis deep into her mouth. A pov view of a woman seen giving a deepthroat blowjob to a large, circumcised penis. He is pushing all the penis into her mouth. his penis disappears into her mouth. The woman is visibly gagging and choking on the penis, she recoils as the penis enters deep into her mouth. She keeps eye contact with the camera the whole time. maque, high dynamic motion", "negative_prompt": "", "resolution": "1024x1024", "video_length": 121, "batch_size": 1, "seed": 69823584, "num_inference_steps": 4, "guidance_scale": 1, "flow_shift": 5, "sample_solver": "dpm++", "repeat_generation": 1, "multi_prompts_gen_type": 0, "multi_images_gen_type": 0, "skip_steps_cache_type": "", "skip_steps_multiplier": 2.5, "skip_steps_start_step_perc": 0, "activated_loras": ["dicks_epoch_100.safetensors", "\u9ad8\u901f\u52a8\u6001.safetensors", "jfj-deepthroat-v1.safetensors", "pov_blowjob_v1.1.safetensors", "lightx2v_I2V_14B_480p_cfg_step_distill_rank32_bf16.safetensors"], "loras_multipliers": "0.3 1 0.7 0.7 1", "image_prompt_type": "SE", "audio_prompt_type": "", "sliding_window_size": 257, "sliding_window_overlap": 5, "sliding_window_overlap_noise": 0, "sliding_window_discard_last_frames": 0, "temporal_upsampling": "", "spatial_upsampling": "lanczos1.5", "film_grain_intensity": 0, "film_grain_saturation": 0.5, "RIFLEx_setting": 2,

Friends, please write me a list of comfyui download order, please help me how to do it

>>19832883
Follow the bags instructions, I got in running in 20mins

>>19804710

what lora, sir?

Bump

Hey op I got a few questions. First of all thanks for setting this all up, you da man. It works.
So what does sage/triton do exactly? Installs pytorch? And is there a downside to not using it.
And where did you learn this? I taught myself how to do quite a few other things in comfyui but couldn't grasp wan. A resource would cool unless you had to patch all this together from different corners of the web.

>>19835059
>So what does sage/triton do exactly?

Gives 30-40% speed boost

>And where did you learn this?

Getting different workflows and try to figure how they work, if they have something good, i try to adapt into my own flow

>>19835082
>Gives 30-40% speed boost
Dang seriously and I have to do a clean install of portable first right? I was all comfy on the desktop version.
Only issue is now I'm moving to an external. Will report in a day or two on that.
60 seconds for a 5 second clip with 12gb of vram. I feel like there's a setting I'm not abusing.

>>19835088
i dunno how desktop version works, if if has python_embedded folder, should work. I never had success installing sage on old comfyui thou. On clean portable it installs just fine

up

Would an older MacBook from Walmart be decent enough to run comfy on? If I learn how to do it well I’ll fill request

installed everything, added a picture, tried running and get

CLIPLoaderGGUF
'NoneType' object has no attribute 'endswith'

is there anything else that you need to set up in this workflow?

if I untick Enable GGUF, I get

ModelSamplingSD3
'NoneType' object has no attribute 'clone'

My eyes have been coming out a bit fucked up. Any advice on that?

>>19836497
reselect every model/clip/vae/loras everytime you open some new workflow....so the files point to your computer

Hi OP, Wan 2.2 just got released. Do you know if it's compatible with current LoRAs? Could the workflow be updated to support it? <3

I’m having an issue where shortly after I hit Run, comfyui stops working and it just says “reconnecting”? I have a an RTX3070 and 64GB of ram so I feel like my computer should be able to handle this..

>>19837300
Check the logs (The button should be on the top right of ComfyUI). What does it say when it tries to reconnect?

every time i do the blowjob (dreamjob) lora, the faces changes a bit... not sure how to fix. any suggestion??

>>19837359
Dream blowjob LoRA is mainly for t2v, not i2v. I would recommend the deepthroat LoRA and set the strength to 0.7

https://civitai.com/models/1497390/deepthroat-wan-21-i2v

Wow no Wan 2.2 workflow yet, what are we doing bro. (JK unless you are Kijai)

>>19837370

thank you

>>19837370
There is i2v version of dr34m
https://civitai.com/models/1395313/wan-dr34mjob-doublesinglehandy-blowjob

>>19837206
Yeah i know it was coming this month, did some fast googling and seems like wan 2.1 loras wont not work on 2.2

>>19837464
there are some workflows around already, comfy updated today also for support, native models are huge, but there is also repacked model out there that runs on 8gb cards. I haven't tested it, but seems like much more better camera control and better light-schemes from the demo vids. Give it a few days and Kijai makes this shit 10x better =)

Bump

>>19837359
facefuck works as well. It's just iffy.
https://civitai.com/models/1515071/wan-21-deepthroat-facefuck-entry-blowjob-fellatio-oral-lora-t2v-14b

>>19837882
>There is i2v version of dr34m
nta thanks this was driving me nuts. I wish they would clearly label these things more.

>>19837359
>the faces changes a bit
oh also set your light sampler to at least 6 or something. It'll take a bit longer but you'll get rid of the most morphing.
It also helps to make the women "emotionless" in the prompt at times. Good luck soldier.

Well, seems like 2.1 loras do work on 2.2....at least what i've tested, you get shitload of lora-key errors, but it outputs. Also seems like wan 2.2 isn't a single model but rather split in two, so you have to use two samplers...also meaning double the l

There is also new 5B model for low end cards, i assume loras wont work with this model.

>>19839758
How the HELL did you prevent it from making the "man" lurch over and kiss the top of her head like a fucking idiot? So many of my vids wind up with some guy lurching over like a tard lol.

>>19839769
this might help, add "man's upper body stays out of the frame" to positive prompt

>>19804520
Although this is not a request thread, you end up granting and old request of mine!!!
How lucky I am!!!!
Thank you very much! Astonishing job, as aways!

>>19820143
Aaaand I got lucky again, lol!!! Another former request of mine!
Tks again and again!!!

>>19839753
Can u share this WF?

>>19839821
more like testing ground rather than workflow, so atm no (as i don't want to try to help on something i don't have a grasp on)....i just started building it today

Any chance you could help me find some celebrity lora? And also im having a hard time doing txt2vid even with the stock comfyui setup for it i get missing memory but i got a 24gb gpu (xtx7900) thanks for help

>>19839938
>xtx7900

ATI card....i mean they work, i guess, but you need to put bit more effort into it to get it working, just installing stock comfy won't do it. I'm no expert on non-nvidia things, maybe some1 else can point you to right direction

Why did it fucking stop working. It was working fine for two days. I've changed nothing. All of a sudden a vram error on vhs the literal last part.
What the flying fuck.

>>19839950
a) because it's comfy
b) you did change something, even if something small, read a) again

pic of the workflow and error-log

>>19839959
I sincerely changed nothing. I even reloaded the workflow to ensure nothing was changed.
Nothing was added or removed. I don't get it.

>>19839976
so it just hangs?, nothing on the logs?

>>19839980
First off I disabled the logs but enabled them again just now because I realize I could come here for help.
Second, I found a workaround for anybody else who encounters this error who is working with low vram (12gb in my case).
>settings
>server config
>disable smart memory management

>>19840025
i have 12gb card as well.....sometimes it hangs up on random places, but usually rebooting the comfy will fix the issue.....how much ran you have?

>>19840040
32 and an i9 man. I feel like a damn fool for not getting a desktop two years ago.
I'm genning on a damn laptop and it's working thanks to you.

bump

I have no idea how to get these nodes installed. I've tried using the manager, installing from git link, manually downloading and pasting into the directory. I've used a bunch of different versions too and the folders will install, files and everything but they wont be recognized by comfy as installed.

>>19841125
Gotta be patient and keep restarting when it asks you to. do it one by one instead of all at once.
If one thing breaks move on to another for the time being. Keep going.

I have a quick question regarding how seeds works in WAN. I generated a 5-sec clip and liked the results, so I tried generating with the same seed again for an extended lenght (10 seconds instead of five) thinking it would generate the same clip and just extend it, but seems like I was mistaken. The final clip came out pretty differently and even with a different seed number, which is odd but ok. Am I doing something wrong, or does it work like that? In SD even if you changed the parameters the file would be named after the same seed, the WAN results just leads me to believe for some reason it isn't picking the seed I manually input from.
I know the last frame image is supposed to have exactly that purpose (extending clips) but it often has the woman where her facial features are hidden (if she's balls deep for instance, closed eyes and such) and won't get a good result out of it...

>>19841864
After doing some more testing it is safe to say it is picking up the fixed seed I'm setting, it just gives out a different one in the output for reasons. So I guess changing lenghts will inevitably end up with a different gen and the only way to extend it is by using the last frame, unless there's an alternate advanced wizardry for it.

>>19841864
Even if you keep the fixed seed but change the length from 5sec to 10sec, Wan now has 2x time to do your prompt, so surely it will look different. I wouldn't go beyond 7secs anyways, as many loras tends to to do weird shit if too long Gen

On last frames. If it's blurry or face not visible, I would not use it. You could add node that saves all the frames and then pick usable (takes more time and disc's pace) ... But then you have to take this into consern when joining the clips

up

>>19799683 (OP)
Good looking workflow, looks clean, most of my workflows are a spaghetti factory

Hi Wizards

I'm quite new at this but I want to learn how to do it, I've just started using pix.ai but I want to know how to begin with the videos like you.
I have an AMD Ryzen™ 9 6900HS Mobile Processor (8-core/16-thread, 16MB cache, up to 4.9 GHz max boost) NVIDIA® GeForce RTX™ 3060 Laptop GPU
ROG Boost: 1475MHz* at 120W 16GB DDR5 on board
16GB DDR5-4800 SO-DIMM
Could you please guide a neubie? Thanks!!!

>>19843826
mayby if you use lower models, you can get it runnin...i think the model loading times takes more than the actual generation, as you can get pretty good outputs just with few steps now

>>19839911

This is still much better than anything anyone else has presented, patiently waiting for you to post your WF! I appreciate you sharing the first one as well, have been using it religiously over any of my other WFs.

>>19841125

Did you try updating the whole comfy frontend?

Man I cannot get sage/triton to install no matter what I do. Same shit with the base installation of pytorch.
The commands just, don't, run. I do every step to a T.
I know I'm not retarded because I already ran the commands to grab a few other things but pytorch or triton will not budge.

>>19844657
What card you have? Did you do fresh install on comfyui and tried installing triton/sage via that bat file

>>19839769
LOL I know your pain brother haha. Thank hell for previews.

Is there a prompt share site or are yall making up your own? I see very little talk of prompts in general.

>>19844433
Love your guide m8, you are a legend.

Random comfyui question. I recently did a complete reinstall of my comfyui and its amazing except 1 thing. the Video Combine node has its standard video output preview, used to work fine in my last install but with my new one it stutters and only show the middle of the output vid. If I open the preview in its own tab its fine. any idea what could cause it to freak out? it does it on my other WF's too. cheers

>>19845307
i've noticed it too, i think latest comfyui front-end might've somehow broken it. Might be fixed in newer version of VHS node

>>19845294
civitai is good source, go to the lora page, view gallery, on filter set "metadata only", look at prompts ppl use

>>19845294
prompts are always common sense and ridiculously simple to handle, there isn't much to talk about.

any wizard will tell you the prompting is the easy part, it's balancing everything else that separates the skill levels.

>>19845320
Ah good, glad its not just me. thanks <3

>>19845332
lol no, people be clutching their prompts like secret sauce.

Settings matter but ya prompt is the conductor.

>>19845354

>the woman takes off her dress revealing her large natural breasts. she opens her mouth and shows tongue to camera. she tilts her head as she is grabbing and squeezing her large breasts staring at the viewer.

The prompt is literally describing the action, that should be good enough for you to replicate this then, right?

Let's see how close you can get without knowing which 7 loras were used, along with their respective strength values.

>>19845454
that's a lot of words for "no I wont share my prompt"

>>19845460
I literally shared my prompt.

>>19845460
that's a lot of words for someone that can't read lol

>>19845454
Damn, 7 loras? Max I used was like 4.
Guessing game is fun, but sharing is even more cool.

This is dope, really good thread OP.

Not sure if I should ask here or in a new thread. Does anyone here have a similar workflow to decensor existing videos? Wanting to setup a similar workflow for JAV, "comedy censorship", news footage, etc

>>19845320
>>19845307
This is likely the issue?

anyone have a decent workflow for wan 2.2 yet that they're willing to share? my current workflow that im using for 2.1 is attached on this post here. hoping to make something as close to it as possible.

https://gofile.io/d/2Po5N2

>>19846081

You would have to decensor initial frame with either SDXL or kontext and then use something like VACE to keep it consistent with the OG clip.

It took me a few days of messing with it but I got pytorch running.
Had to edit path variables to make windows actually read where python and pip was.

>>19846583
>https://gofile.io/d/2Po5N2

Seems like you are into acceleration loras yet?. Also i recommend to set your teacache to start at 0.2 / stop at 0.9. Not ideal to skip last steps, might cause some glitches

WAN2.2 needs two samplers (has 2 model, high and low noise ones). I've been playing around with it for a week now. There are some basic 2.2 workflows at civitai to get the idea how to setup the samplers

bumps

Up

Very nice tips here.

>>19848789
Are or aren't? And what are acceleration loras?

up

>>19851448
*aren't. Acceleration loras lika causvid, fusionx or lightx. Skips you steps need down to 4-8

>>19852088
so those are actual loras, or nodes? how would I incorporate that into my current workflow?

>>19852197
those are lora files yes. Short answer;
>get lightx lora for i2v (https://civitai.com/models/1585622?modelVersionId=2016546)
>load it up 0.8-1.0 weight on your lora loader
>(might want to add WanVideoNag node just before the Ksampler)
>set cfg to 1 (important)
>set steps to 4-8 (lcm/simple, euler/beta, unipc/simple combois work ok)
>go

>>19845454
most results have a 'generic tits' look. This looks unique in shape, weight etc... Is it a random result or can you control it somehow?

>>19852235
Okay, so I did what you recommended. and the reduced timeframe is amazing! the results look great. However, I bumped into a hiccup. The first render I do right after a fresh bootup, turns out fine. No issues. But any render I do after that, turns out with the results shown in the picture. Only way to get it to stop doing that, is to restart comfyui and then the first render again turns out fine, followed by this issue happening with the rest after.

>>19853475
i tested with removed the wanvideonag incase it was the cause. no changes, same results. then i tried changing the lcm/simple and euler/beta and unipc/simple combos, and once again, it does the same thing. first render after fresh bootup turns out fine, but then the rest end up weirdly pixelated like in the picture.

>>19853478
Is comfyui ui up to date? Things you could try is to add clear vram node after video is done. Other thing to try out is after your first run change one of your Lora strength up or down, even 0.05 will do, this forces model to reload next time you hit run

>>19853523
yeah after doing research i realized its my gpu vram cache thats being caused from v-ram exhaustion since i only got a 2070 super with 8gb of v-ram. think its slightly pushing my memory limits. but considering the first render after a bootup always turns out fine, ill just have to add that node at the end of the workflow that'll clear out my V-RAM cache so it can be fresh for consistent renders.

>>19853523
i did hit update all though, so imma check to see if it does it again with the rainbow pixels just incase.

>>19853539
nah sadly updating everything didnt fix it, still has the same issue after the first render. so i'll just install the memory extention folder and use a node to clear out my v-ram cache

Update all updates the nodes... What's the version on comfyui itself. And try changing one of the Lora strength after first run... See what that does for you

>>19853562
ill get back to you in the morning to see how it turns out. falling asleep here at my pc. i appreciate the help so far though.

Bumps

Hey bag, have you been trying out 2.2?

I've been having mixed results. Honestly the best ones have been without LightX, but I really don't want to go back to 1+ min per 1 sec of clip.
The main advantage I see, but haven't tested yet, is that wan 2.2 seems to force a return to the original at the end when it comes to faces. Maybe this could be a big deal for extending vids.

>>19810835
I’m tech illiterate but I’m planning to get into this. What are these exactly?

Bumper

>>19855249
Loras are concept models that work in with the main model and prompts you use. No lora, no magic output of sexy images/videos in the end. You can search them on civitai and learn more about each one.

>>19855219
Yeah, i kinda have mixed results on 2.2 as well, quality is there for sure and movement is better by default. Using old 2.1 loras...is kinda coinflip. I'm sure lightx is coming for 2.2. But i dunno how new loras are going to be, also separated into high/noise parts?

>>19855334
there are a few new Loras (all t2v based) that have low and high files... Mixed feelings on that, organization is not my strong suit. Haven't tried them though.

>>19855268
Okay, so I’m a month away from a good pc, what list of tools would I need to get to you guys’ level

>>19855396
what kinda of pc build you had in mind?

>>19855420
I have a friend Barney styling me through it and he has me window shopping for stuff. I do know a buddy is giving me a 3060 for free, and he’s gonna tell me what all I’ll need this week after he puts in his 5090. I haven’t cared about. Computers in about a decade, so I’m fumbling, but this is an aspiration of mine

bumpzi bump

Bump to keep it alive

>>19855481
With a 5090 and 3060 you should have a pretty smooth machine for gens. 5090 alone does it. Focus on that ram and drive space.

Lightx2v for Wan 2.2 seems to have just released:

https://huggingface.co/lightx2v/Wan2.2-Lightning/tree/main

>>19857225
Can you give a few recommendations? As I said it’s uncharted territory so I’m taking all the notes I can

>>19857738
I barely know what I'm doing myself. Stronger bits = faster gens.
And put like a thousand niggawatts of ram in it.

Up

>>19857436
Seems only t2v atm (but some say works if old lightx i2v version and causvid loaded too)

Anyone know why my outputs look like this ? the first frame is my original image then its just this random noise

>>19859236
if my workflow, post screenshot of it so i can help ya bit better

We are nearing the end of this thread. Big thanks to all you bumped this into knowledge to others. Maybe we got few new wizards from this.

>>19860007
I have a build setup now. Gonna put a 5070 in it to start, with an intel i9, and an msi pro a max. I pray you will be there to help brother. I’m about 1400 away from joining you in wizarding

up

Up my guy

>>19860714
Up?

>>19820111

very newbie.. (today) running into this problem... can you explain it to me like im 5

>>19861539
EVERYTIME you open any new workflow RESELECT all model/clip/vae/lora files (so they point locations in your computer)

>>19861559

does anything look wrong here?

>>19861594

looks like CLIPLoader is the specific error its referring to according to that status bar on Comfy

>>19861594
wan 2.2 model....my workflow is for 2.1

wan 2.2 need dual samplers. I make one for 2.2 once acceleration loras get updated

>>19861621
you also need gguf clip, if planning to use GGUF models

>>19861627

I didn't download anything outside of what was on your JSON from the UI manager. Are you saying that's seperate?

>>19861663
here is all the stuff you need.....

>>19861686

and to confirm, the umt5_xxl_fp8_e4m3fn_scaled.safetensors goes in like 3 of those folders? I went to each individual link to download, but that file downloaded multiple times from every link

>>19861717
>umt5_xxl_fp8_e4m3fn_scaled.safetensors
goes into text_encoders

>>19861733

Thank you. For Clip GGUF, which of these should I install? Any difference? I'm running a 7800x3d and 5070 if that makes a difference

>>19861785

or do they all need to be downloaded

>>19861784
with that card you should use diffusion models, GGUFs are good for lower end card / low ram

>>19861785
but if you wanna use ggufs, then Q8 or f16

>>19861791

I'm confused, so I shouldn't download then? Just use the diffusion models I'm downloading from the link?

>>19861804

Sorry if these are dumb questions. Just trying to understand everything

>>19861790
only one needed

>>19861808

Thank you. I'll report back later once shit is downloaded

>>19861804
yeah, get the diffusion models....you can still download the gguf models (q8 or f16) if you wish to see the quality difference

>>19861807
no worries, comfy is bit confusing at firsts

>>19861818

still getting the same error

>>19861539
>>19861886

specifically

>>19861886
care to post the whole workflow image? also what happens if you try diffusion models?

>>19861895

standby might have fixed it

>>19861895

fixed that part, have a new error i'll post after this

>>19861897

did you remember to download the lightx lora and load it? thats pretty important if you want to make fast gens

>>19861903
is that in one of the side folders? Do you have a discord? might be easier there

>>19861903

this is in my models-loras folder

>>19861907
means you didn't install sage/triton....you can disable it....but i highly recommend you install it as it give 30% speed boost

>>19861915
looks good.....lightx must be loaded in here (seems like you do)

>>19861916

Interesting... I distinctly remember doing that

>>19861920
>>19820113

this is what I followed. Maybe I messed up a step?

>>19861932

ran one without sage activated. It's hotdog time

>>19861932
seems ok....
you could try....go to python_embedded forced , make a bat and run it. Its pain in the ass to install sometimes....

>python.exe -m pip install --force-reinstall sageattention

if that not work, just disabled and run it slow

>>19861953
you have 3 loras that try to add penis into the gen at the same time =) i dunno what deepthroat lora you got, this is good one, remove penis and facefuck loras

https://civitai.com/models/1497390/deepthroat-wan-21-i2v

Also remove the lightx lora from your power lora loader (no need to load it twice)

>>19861955

haha to be honest, don't even know how to do this

>>19861975
no worries, you should get pretty good speeds on your card, few extra mins waiting wont kill ya

Different wizard here. I switched to a 5090 a couple of months ago (from an older card), and at the time the 50-series cards had a whole bunch of trouble with sage/triton and more. I don't remember exactly how I ended up solving the problem, but since then I have seen new installers that work with 50-series as well. You could try googling for that while you wait for your current gens.

>>19861981

Where do you find your loras? I've used some I've seen in this thread. I like your work specifically with the face fucking- and the scenes you created with two dudes taking off someone's clothes

If I duplicate the windows, can i load a whole other test? Or do I need to open an entirely new comfyUI

>>19862022
>If I duplicate the windows, can i load a whole other test? Or do I need to open an entirely new comfyUI,

you can just duplicate the tab

also: i found other installer for sage, maybe this one works better on 50xx cards

https://github.com/HerrDehy/SharePublic/blob/main/sage-attention-install-helper-comfyui-portable_v1.0.bat

(place it inside comfyui_portable)

>>19862029
>>19862022

like how do I reduce breast size to make it realistic with the image? Different lora? each keeps coming out much larger than the images

>>19862044
you can prompt "reveal small nude breasts"
there are also loras

https://civitai.com/models/1613190/wan-14b-boob-size-slider?modelVersionId=1825608

(^^ is for t2v, but works somewhat in i2v)

>>19862029

I just have to put it in the portable correct? If so unfortunately it did not work.

What do you usually say in your prompts? Plain language? Like blow job/suck vs fellacio? Would you mind sharing that lora/prompt with the two guys removing clothes and the females acting shocked? Or is that a secret? haha

Thank you again for all your help

Is it possible to make this work with a RTX 3050 Ti Laptop GPU? It only has 4GB in VRAM but my laptop has 32 GB of RAM. I tried to generate it using Q8 and Q4 and I always run of memory. I also enabled blockswap. What settings do I change to make it work?

>>19862103
yep in the root of comfyui_portable, you need to rename python_standalone TO python_embeded (note that comfy can't be running) then run the bat, once done, rename the python folder back to python_standalone

You need ripclothes lora: https://www.seaart.ai/models/detail/cvmbsule878c73c869cg

prompt simple as:
>two shirtless black men appear in frame. Men rip woman's clothes off revealing her nude body. Woman looks shocked

>>19862121
Q8 way too big model for your card. You could try:

>On launcher you could enable low vram mode.
>get even more lower Q model
>don't load many loras
>lower the output resolution (width)
>lower the steps (i suggest you use lightx method and 4-6steps)
>decrease the length, try 3-4sec

I wouldn't keep my hopes up thou....you might get it to run, but details are out of the window on those lower quanta models

>>19862122

I'll try that out. thanks.

Side note: for some reason my dicks keep coming out looking like hot dogs

Kind Wiz, could you recommend a LORA or a prompt to get my subject to turn around and show her ass while the camera pans down to her butt?

>>19862171
try genital helper:
https://civitai.com/models/1434650/nsfwfemale-genitals-helper-for-wan-t2vi2v?modelVersionId=2012120

>woman turns away from viewer and (bends over:1.2) to show cute butt and innie vagina camera zooms in on butt

>>19862169
what image are you trying to do, what prompt and which loras loaded?

>>19861953
>>19862182

the one referenced, and the lora you recommended to me earlier. Can you give an example prompt? Maybe i'm being too complex

>>19862187
post the full pic

>>19862191

also thread hit bump limit. Would love a new one from you so I can ask questions if I have. Tried searching you on disc but couldn't find you

>>19862182
Thank you!! That's exactly what I needed

>>19862195
This is kinda the downside of lightx, due to low amount of steps, movement and details take some hit....you can get bit better looking dick by adden erect penis lora on 0.5

alternative you could use normal sampler, takes more time, but much better movement (and dicks =))

>>19862266

yup that's basically what my dicks came out as. Unfortunately the sage thing still didn't work... just did a 10 minute render..

>>19862283
did you remember to enables in the node?

>>19862287
I would love to answer that but I have no idea what that means. Still my first day lol

>>19862287
>>19862290

there's an option in the Patch KJ section that says something about Sage but it doesn't work when I try and run it. There's also a triton option in there

>>19862290
>>19861916

You could also try different deepthroat loras, some render better penis

>>19862298

I'm guessing that's a different nora than the previous one you linked?

>>19862302
yep....
https://civitai.com/models/1694742/ultimate-deepthroat-wan-video-lora?modelVersionId=1992542

>>19862307

I'll try it out. Thank you. So far one decent video thanks to you.

>>19862311
post you disc and i get touch with ya....will make new thread once i get wan 2.2 workflow done

>>19862330

gman1990_

>>19862330

one last thing.... can you rename your lora files?

>>19862335
yes...i would even recommend that case you have many different types of loras (flux,wan,sdxl)

Anyone know any good Loras for wan 2.2 ? i tried using the 2.1 ones and those fail to load.

>>19862712
2.1 loras work just fine. They throw some lora key errors because not all aspects of 2.1 loras are used in 2.2 (which instead uses 2 models). But they do load and work

>>19862804
ohhh okay i thought the errors meant they are not working...Thank you soo much!! i had no clue

Would you mind sharing how long it takes your setup to generate a 5 second video?

so i had this workflow working nicely
but after a comfy update im getting this error
i tried updating all nodes but still getting error

Does anyone have the Penis Reveal lora to share? Seems like it's one of those that civitai nuked and it isn't on the txt...