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

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

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

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

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

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

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

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

Old thread(s)
https://archived.moe/r/search/subject/ComfyUI%20portable%20for%20newbies/

File: 157740443920743740_00001.mp4 (1.52 MB, 592x800)

1.52 MB MP4

Hey AId Bag Thanks for all your work and teaching me this magic. Any suggestions about audio or talking or lip sync using comfyui?

>>20483176

I'm not Mr. first responder (1st aid)

..but some genner's use LTX, its easy and fast but very inconsistent with any type of sex gens and requires a bit of GPU power. Others use SVI, S2v and there are some tricked out workflows that can do it with Wan and possibly even add sound to existing gens.. but I don't have experience with any of those.

File: 157382568959323227_00007.mp4 (1.19 MB, 592x800)

1.19 MB MP4

>>20483225
I have heard of this but cannot find a good local download ? i keep ending up back at the website and it wants me to pay. can u share?

>>20483231

For LTX? Just go into comfy templates, click on LTX2.3, download the models and read the boxes where it tells you to put them. Least that is what I did.

I'm sure there are a zillion tutorials on reddit/utube/huggingface too.

File: 157919977539934064_00007.mp4 (838 KB, 592x800)

838 KB MP4

>>20483251
oh shit thanks didnt realize it was just a comfyui thing when people talked about it.

thanks

>>20483176
well i use ltx 2.3 (locally, if wan ggufs runs, you can run ltx as well)

https://huggingface.co/QuantStack/LTX-2.3-GGUF

like anon said above....it has pretty limited nsfw support off the bat. But let say you start with some nsfw, load some ltx nsfw lora then prompt the woman to moan or something like that, that works ok. you can make 30sec clips, quality is ok, and is much faster than wan.

but getting a nude man appear in frame and shove his his penis in talking woman mouth...that is bit tricker task

for wan 2.2 audio, the is s2v (sound2video) model....but this kinda works other way around, you first need to make the soundfile and wan will try to gen a video out from that with help from your prompr

File: fl15153153153.mp4 (897 KB, 736x544)

897 KB MP4

>>20483307
no i just want my school teacher to talk dirty while she uses the fleshlight. nothing too crazy

>>20483328
i would guess that is possible....well maybe not that fleshlight part as i doubt model nor loras understand what that is....but you could use endframes or middleframes to force that transparent fucktube in

File: Audio_00322_.mp4 (1.2 MB, 256x512)

1.2 MB MP4

>>20483328

That is definitely doable. When you have to gen a penis from scratch is where LTX usually has issues but if the penis is already in the image should not be a problem.

File: 256151515.jpg (212 KB, 880x1184)

212 KB JPG

>>20483366
Yesss i want to make her talk dirty before she gives head. ill have to work on it.

thank you

File: Audio_00128_.mp4 (1.02 MB, 512x320)

1.02 MB MP4

>>20483328

Will take a bit of finaggling with the prompts/loras and playing with CFG/multiple runs but got this in about 5 min. Also I used a cocksleeve in starting image.

LTX isn't like Wan, it needs you to desribe all object/motions like you are explaining it to a 5 year old when you prompt and didn't feel like going into all that for an example vid without the cocksleeve already in the photo (you will likely need to describe it VERY well if it's not in initial image)

File: ltx2_00159-audio-.mp4 (1.63 MB, 960x512)

1.63 MB MP4

>>20483328

>>20483169
hey, just wanted to say thanks, you're an actual fucking hero
>>20483366
lmao

File: 15714503581898498_00001.mp4 (3 MB, 592x800)

3 MB MP4

>>20483617
This is awesome!!! THANKS

File: 157305872332894872_00007.mp4 (1.43 MB, 496x688)

1.43 MB MP4

>>20483617
thought maybe the workflow was baked in but its not? care to share?

Is flash attention not supposed to work with a 3090? I'm getting multiple lines of this when generating:

Flash Attention failed, using default SDPA: schema_.has_value() INTERNAL ASSERT FAILED at "C:\\actions-runner\\_work\\pytorch\\pytorch\\pytorch\\aten\\src\\ATen/core/dispatch/OperatorEntry.h":84, please report a bug to PyTorch. Tried to access the schema for which doesn't have a schema registered yet

>>20484390

Cut and paste that line into google and gives you a few things to try

>>20484517
Looks like it was the RES4LYF node pack, I disabled it and the spam stopped. Apparently it's incompatible with Flash Attention.

Are you guys using a normal text encoder or an uncensored one?

>>20484963

For LTX? I downloaded like 100gb of various alt model/text encoder/Gemmas but have been using the default ones linked in the template with no problem just using NSFW loras.

>>20484390
stupid question but did you install it? also are you using my pack or some other comfyui desktop/portable?

flash attn should work ok almost every card. Before sage flash was the best speed up attention to use.

>>20484963
for wan/ltx default text encoders work just fine...uncersored versions have some finetunes/edits that makes it understand some nsfw phrasing better....but sometimes it makes it worse also. tldr; they are not mandatory to make nsfw content

>>20484963
>>20485307
thanks, than i'll just stick with the default one

>>20485296
Your pack. I didn't install anything extra besides your pack and a few checkpoints and loras.

>>20485479
did you install attentions via the installer?

>>20485609
Yeah, the only steps I skipped were from SDXL to Losslesscut since I don't use these. I solved the issue by disabling RES4LYF.

>>20486117
ok....did you launch with flash attention bat?
also does the sage work ok for you?

also if you have res4lyf disable, does it load flash ok? (can be seen on the console)

i dunno why res4lyf would be an issue, but more info i get, the better

Hi, trying to make this work for the first time but not having a lot of luck.
No errors, I just followed the guide and tracked all resources for the WAN_SVI2_pro_1.2 workflow (the guide said most recent but they're not ordered by date so I guessed), tried gguf 2 and 5 but it tried for an hour and then died.

I half expected something like this because I'm running on a basic 3060 12vram + 32 ram
The thing is I'm a total newbie so other than redownloading the gguf thing I have no idea how to lighten the load to make it functional on my potato.

I don't wanna inconvenience you with spoonfeeding, I can normally poke around until I get it right on my own, but I can't tinker and improve if it won't even spit out any results.
Could I get some tips or perhaps an example of a lightweight workflow to use as reference?

Oh and this LTX thing, is it just faster or potentially lighter? would that be better than WAN for vramlets such as myself?

Thank you for the high effort guide and resources btw, Aldbag you're the goat for doing all this

>>20486144
Yeah, launching with flash attention bat.

I thought sage attention was only for 40/50xx, I have a 3090, will it work?

Yes, disabling res4lyf fixes flash, there is a guide to fix it but since I'm not using it I just disabled it: https://note.com/198619891990/n/n1c7082aa1c6e

Apparently some of the stuff res4lyf installs breaks Flash Attention by default.

>>20486163
well you should get shit running with those specs...might be slow, but should get work

if you have disc, lemme know and i'll contact you, much easier to troubleshoot/help you out (and not spam the thread)

>Oh and this LTX thing, is it just faster or potentially lighter? would that be better than WAN for vramlets such as myself?

yes and no....dev version of ltx model is about 40 gigs, plus the text encoder and loras.....gguf version are doable on local machines (12gb vram / 32bg ram). They take bit to load but gen is much faster....time it takes to make 5sec in wan = 20 secs in ltx (but again, for nsfw purposes ...ltx is not there, yet)

>>20486193
sage 2.1 works ok with rtx30xx and rtx40xx cards

sage 3.x is for blackwell chpsets cards (rtx50xx)

sage 1.0 if everything fails and for the rest (rtx20xx, rtx 10xx)

>>20486193
also what version of comfyui?....heard that the latest major update fucked up some nodes. I will make a fresh pack when this thread goes beyond bump, make a fresh install myself and see what works or not

>>20486259
Got it, I will try 2.1 then, thanks.

As for the version, the one your script installed is ComfyUI 0.19.3.

>>20486269
yeah, the installer always installs the latest repo

might change that in the future.....

>>20486269
if you installed attentions via the installer, sage 2 (and 3) should be installed

>>20486210
Thanks for the help
I just need a few quick pointers though, like I noticed it stalls hard during the tripleksampler process, I lurked and read someone with a 3070 (8gb?) swapped the ksampler for a different (moe?) version. I don't know if it helps in quality or speed though.

Could you just tell me out of all the workflows in the portable pack, which one would be the lightest for tests? If I can get it running I'll figure it out.

perhaps someone with a similar shitrig has some input.

>>20486319
just disable the 3rd sampler....set output reso to 360-480, just to see if you get anything going

also post a screenshot of your workflow...i've made about dozen so far so hard to pin out what are referring to

>>20486354
Ok I'm a retard, so sorry. I didn't realize "output longest side in pixels" had to be changed to set resolution.
It managed to produce something and lo and behold the problem: it was set to 5 seconds but the workflow extends 3 times, so it was trying to gen 20 seconds at 720p, that's why it was dying.

I'm sorry I don't even know how to screenshot this massive thing in a legible way. Looking around, although I don't think it's necessary anymore.
It's pretty much the default WAN_SVI2_pro_1.2 workflow.

Now I just need to set it up to 5 seconds, or 10 seconds with 1 extension. Is it as simple as bypassing the extend nodes/delete them?

Thanks for your help, seriously, I've never even used comfy prior to this, just the other baby ones for images.

>>20486532
>so it was trying to gen 20 seconds at 720p, that's why it was dying.

not correct....you are making 5 secs clips four times (they are all separate gens)

making one 20secs clip at 720p at wan 2.2 would kill it for sure....

for screencaps....search your window...type "snipping" , that give you the tool to take a screenshot of the area of choice. use mousescroll to zoom out your workflow where most of things are visible (output stuff not important), all text readable, then snip it

File: workflow.jpg (3.45 MB, 4960x4135)

3.45 MB JPG

>>20486573
ok it was right click > workflow image > export > png
just had to resize and compress

You've done enough but is there an easy way to disable extend 1 and extend 2?
I'm assuming the purple extends at the bottom are disabled?
I'm confused about very basic things for now
I just want simpler shorter gens so I can tinker and learn a bit faster.

Would it be better to look for a different workflow or should I try adapting this?

>>20486651
OH my fucking god I just saw the nodes for "enable extend"
Sorry disregard this I suck dicks.

Thank you again for setting all this up, no more questions, ty

>>20486656
yeah...if you just want basic 5secs clips, no need to use SVI workflow.

they my wan 2.2 1.5d...or other basic wan workflow you can find on civitai...you have all the needed components now, choose a workflow you find most easy to use

and other note; looking at your prompt....there is not going to be any fingering unless you load some fingering lora

>>20486670
>looking at your prompt....there is not going to be any fingering unless you load some fingering lora

Oh that's just the default prompt, I haven't gotten to that yet. I did track the loras down anyhow, so plenty to play with

Thanks again for all you do

>>20486679
also note when you are dealing with 5 second limit....its impossible to the woman undress, spread her legs, finger herself

this where SVI workflow comes in.....or make one 5sec clip where she does X, save its last frame and use that as new input for your next gen where she does Y (...etc....etc)

>>20486720
Yeah I caught on it's broken down in timed stages and stitched together.
I disabled 1 extend and it gave me 13 seconds (12.5?) which is agreeable.

Can non SVI handle 10 seconds or so? I read somewhere the longer it goes the more it degrades.

File: 2026-04-27 042908.png (104 KB, 763x205)

104 KB PNG

>>20486856
generally wan does does 5sec (90% of loras are trained from 5sec samples)...after that no matter you prompt...wan does what it want, most commonly it starts to reverse the action (aka it starts to go back to point where it knew what to do)

example prompt "woman opens her shirt, revealing her nude tits"

on 5secs....that works fine, she will undress and scene ends

on 8secs....she undress within the first 5sec, then spend last 3secs putting the shirt back on

>I disabled 1 extend and it gave me 13 seconds (12.5?) which is agreeable.

note the settings....just one clip (aka just initial scene) on 5sec, interpolation 2, speed 1.0 should be 5sec....if speed set to 0.5 --> 10sec and so on

>>20486930
to continue (and confuse you more)

there are ways to make single long clips;

this is using the last frame option

you provide image as input and second image as the last frame.....this kinda forces wan to do the prompt within the time you provide it. It knows where to start FIRST FRAME...it knows what to do PROMPT....guiding it all toward to your LAST FRAME

one good example of this would be if you have a pic where woman is sucking the penis. use that on BOTH; first and last frame, prompt it "woman sucking penis, blowjob" (with some lora loaded ofc), set length to 10-15sec....it keeps the face mostly unchanged as it has the end frame reference

>>20486656

Alternatively think you could just do "bypass" for the extend nodes, the re-enable when you want to use them.