Here is the only description I will provide:
Following an overwhelming amount of DMs requesting the return of my models, I decided to train this to settle the matter with Wan 2.2 once and for all.
This model is not a sloppy job like the merges currently available.
It is not one of those hastily thrown-together mixes by incompetents blending Wan 2.1 and Wan 2.2.
This model far surpasses the competition because it is not a merge; it is trained.
This will be the one and only model I release to you.
The choice is yours: Quality or Mediocrity.
It's up to you to judge !
This message is for the hateful, defamatory, and malicious people who come here to spew their venom.
You will not receive a response from me. I will ignore and block you, and I will report you to the moderators. Have a good day :)
Used: umt5_xxl_fp16.safetensors
Using the provided link and workflow is not mandatory.
If you have no configuration experience, use the standard workflows.
Not recommended for beginners
To get the most out of the model: https://github.com/XT-404/XT-404_SKYNET
Description
FAQ
Comments (208)
Hmmm...
Looks like X4ctl! what I need :P
Merry Christmas - BTW :)
Takes pretty long with the SKYNET Master ^^ Are you sure that 2.5 CFG is fine, and 1 Step is fine? Need some assurance here :D
[10:17<00:00, 617.55s/it]
yes it's ok !
Oooof, Refine is even heavier xD
4090 at 22.250 MB used, so that's fine, but it takes even longer...
Posting full Console when done.
@FrozenGT Is your computer an Atari 2600?
Even with 12GB it's faster!
Now you are really shitting me with the OLED Chroma Mimi xD
Ain't no one got time for that! 15min for Color grading... I don't think so!
128GB RAM FULLY used for the Chroma Mimic btw... What might it be now??? 25 to 30min on that Step???
Running it again at 2x upscale only - As I see the Output at 3x is too big to up on Civit Ai anyway.
100%|███████████████████████████████████████████████████████████████████████████████████| 1/1 [10:17<00:00, 617.55s/it]
[XT-404 OMEGA] SYNC: Locked to Master Clock.
[XT-404 OMEGA] EXEC: CHAIN | Steps: 1 -> 2
0 models unloaded.
loaded partially; 0.00 MB usable, 0.00 MB loaded, 13629.08 MB offloaded, 877.51 MB buffer reserved, lowvram patches: 0
0%| | 0/1 [00:00<?, ?it/s]ST:00 │ INJECT │ BOOT/RESET │ T:------ M:------ │ Δ:------/0.025 [..........] │ SIG:.......... │ 🔒 LOCKED
100%|███████████████████████████████████████████████████████████████████████████████████| 1/1 [04:50<00:00, 290.97s/it]
⚡ [Wan Accel] TF32 ENGINE STARTED (High Speed).
🛡️ [Wan Accel] Installing Contextual Switches on Norm Layers...
✅ [Wan Accel] Anti-Burn Protection Active: Sensitive layers will run in Native FP32.
[XT-404 OMEGA] SYNC: Locked to Master Clock.
[XT-404 OMEGA] EXEC: REFINER | Steps: 2 -> 6
Requested to load WAN21
0 models unloaded.
loaded partially; 0.00 MB usable, 0.00 MB loaded, 13629.08 MB offloaded, 150.03 MB buffer reserved, lowvram patches: 0
0%| | 0/4 [00:00<?, ?it/s]ST:00 │ INJECT │ BOOT/RESET │ T:------ M:------ │ Δ:------/0.025 [..........] │ SIG:.......... │ 🔒 LOCKED
25%|████████████████████▊ | 1/4 [04:50<14:30, 290.16s/it]ST:01 │ INJECT │ WARMUP SEQ │ T:0.0765 M:0.1181 │ Δ:0.0890/0.025 [▮▮▮▮▮▮▮▮▮▮] │ SIG:▰▰▱▱▱▱▱▱▱▱ │ 🔒 LOCKED
50%|█████████████████████████████████████████▌ | 2/4 [09:34<09:32, 286.48s/it]ST:02 │ INJECT │ WARMUP SEQ │ T:0.1181 M:0.2074 │ Δ:0.1449/0.025 [▮▮▮▮▮▮▮▮▮▮] │ SIG:▱▱▱▱▱▱▱▱▱▱ │ 🔒 LOCKED
75%|██████████████████████████████████████████████████████████████▎ | 3/4 [14:27<04:49, 289.49s/it]ST:03 │ INJECT │ WARMUP SEQ │ T:0.1916 M:0.4243 │ Δ:0.2614/0.025 [▮▮▮▮▮▮▮▮▮▮] │ SIG:▱▱▱▱▱▱▱▱▱▱ │ 🔒 LOCKED
100%|███████████████████████████████████████████████████████████████████████████████████| 4/4 [19:24<00:00, 291.19s/it]
35min for everything up to/before the RIFE and MIMIC... puh
Quality is great... but
@FrozenGT : What graphics card are you using? It's surprising that it's taking so long.
I tested it on 12GB, 16GB, and 24GB cards, and it didn't take nearly as long.
If you're using an AMD card, the code isn't optimized for AMD graphics cards at all. If you're using an Nvidia card, then I'm surprised and astonished that you're using 6-step processing. You can modify the ULTRA HD setting to reduce the precision from FP32 to FP16/BF16.
Prompt executed in 00:46:45
with 2x lanczos upscale
3x my normal generations :(
@XT_404 4090 - just powerlimited to 45% (220Watts)
3x the generation time over my other vids
@FrozenGT : I also have a 4090 and I generate videos between 100 and 210 seconds max
If you are "surprised and astonished" that I use 6-step processing on my 4090 - AS YOU CONFIGURED IN YOUR WORKFLOW - WHILST YOU ARE USING YOUR 4090 - THEN WTF!!!
I can tell you that I am "surprised and astonished" that you only manage to put out a handful of Videos with your "100 and 210 seconds" Workflow.
@FrozenGT Defuse and appreciate , people... Tweak the settings till you get the desired effect and maybe try temporarily disabling your undervolt to see how that goes.
@ScientistPlus9579481 No need to do that. The difference is NEVER 300%
Model running on my Workflow
Prompt executed in 00:16:12
Looking like expected!
I give yours one more try without changing any settings - Which should be great, because you also use a 4090
And btw, I posted my Master time at "[10:17<00:00, 617.55s/it]".
And I was told "XT_404: yes it's ok !"
How can that be "ok" when now the whole run shall take 100 and 210 seconds?
Reliable Information is to be expected.
SageAttention, Triton, other requisits?
@FrozenGT You can use it, yes, but you have to install it.
@XT_404 If there are no requisits, I really have no Idea what breaks it.
Must be something in the SkyNet Nodes.
Model runs as expected on my normal Workflow/Nodes.
Maybe SkyNet steals time from my PC and adds that time to yours xD
I played around with the CFG and the Steps... nope... Master alone takes over 200sec at 1CFG and 2 Steps ^^
@FrozenGT : I think it's best if you use your usual work workflow. If the provided workflow and the Skynet suite are causing you problems, then you won't have any further issues. Using my workflow is not mandatory.
@XT_404 I disabled every node before the skynet master and added new model loaders.
result = the same speed
So if they are supposed to speed it up, they dont work.
And if they are not supposes to speed things up, the issue begins with the master node
That's it. Can't check more. Going 2 Bed
@FrozenGT
Where is this from?
✅ [Wan Accel] Anti-Burn Protection Active: Sensitive layers will run in Native FP32.
@dillion1920 ComfyUi Console
I think your input Image is very large! XT404 uses a simple Image convert node that divides the image resolution by 2 and use it as desired input resolution for the video. Try a smaller image or change the desired video resolution to a normal size
@Scirocco1981 Good call! I tried it, and it instantly gives faster results. I am still in winter-hibernation ^^ So I likely won't test much more, but 512x768 was done at a start-up run in 220sec, and a 768x1344 2nd run was done in 525sec. Still no clue how he claims 100-200sec per Video. And no clue why one would release a Workflow/Nodes without ANY instructions and Zero QoL Features... At least I can try to start an attempt at a work-around, like every damn time some "Professional" releases his Workflow. Pfff. There is not one time it ain't broken from the get go or just a "Look how orderly it all is located" or "Why don't you know what DKaBz is and what it does?!" useless, unfixable, convoluted, mess that puts out Videos that look like a true Atari Output and can't even remotely keep up with my "Are you running a Atari" outputs. I am so sick of all these "Professionals" keeping everything as a secret, reflecting all blame, and posing as the Messiah. This is Bethesda levels of not giving a Cr*p about your Consumers xD I can just speak for me, but I have not let anyone's Message about my Videos and Workflow be unanswered, and I would rather die with the Face in the Toilet than give out faulty advice!
Chroma Mimic got kicked out with no regret! Why the F shall anyone spend HOWEVER MUCH time to make the Video look worse?! The Auto Image Half Node got the Boot too WHY?! Just WHY would you add that WITHOUT ANY INSTRUCTION on what Size the Input has to be. PURE NONSENSE! Generation time is still garbage, but at least the Video looks better. This is YOUR NodePack, XT404, what Resolution is the damn sweetspot for the WAN Video Node?! You tested on "12GB, 16GB, and 24GB cards" so that is very useful if you keep any results to your own...
@FrozenGT : Stop talking, my little cream puff, we're not on GitHub here, lol
@FrozenGT Chroma worked not for mee too... First! Then i updated comfy incl Dependencies to the stable Version and from now on the Videos doesnt overexpose with chroma. Nevermind i kicked this node too. But why you are so salty? Its for free mate... when you are not happy use another workflow. With a little personal tune its my daily WAN workflow. Thanks for your work XT404
@Scirocco1981 thx :) ♥
@Scirocco1981 I'm "salty" because he is belittling me and others before me. As with the last Model he published (don't look for it, it's gone!), he did not want to listen to the criticism, took all the praise, but then had to take the Model down because it was clearly and so damn obviously faulty. No clue how that one made it to release at all! I volunteered to test his next Model afterwards for him, but he did not ask for anything to be tested, even after asking him what he needed to be tested. I did some testing on my own and provided the Results how I saw fit. Still, it was like shouting at a Wall! No meaningful response was sent back ever. And now you can read this Thread about why I am "salty". I mean, it should be clear that you, as an Author, provide AT LEAST a rough outline of what your "Audience" is getting into (many still don't do that, and I know that and usually just shake my Head before moving on). But no! He keeps on praising him self, claims that he does all these things, but there is not the slightest indicator that would affirm that he truly did what he claims to have done. He wonders why people are "Haters" when there can not be any confidence in his Statements. Everything gets brushed off as some kind of "Haters are out to get me. Don't trust them."
Sorry, but it does not work like that.
Saying "My Workflow is just for Professionals and people who know what they are doing" whilst recommending it as a necessary baseline in the first place, and claiming some magical 100-200sec results later on.
Yeah, ok, I can claim 5sec results with a 10x10p Image.
Surely figuring out what "eta" or other Values stand for in his Nodes is HIGHLY PROFESSIONAL - FROM THE USER. What do you mean?! And add to that that he does things in French... yeah, you might be asking too much from 99.999% of users at that point, and when you can't figure out that you "price out" all your Users, like you have not been able to figure out that 95% of Users do not own an HDR Monitor over 600HDR Rating (95% have not a useable HDR Monitor), and so do not need a HDR Node in their Workflow, then clearly you do not deserve the trust or praise.
@FrozenGT :
You tested a template that was removed because it was faulty.
I removed my templates because some people were spreading misinformation, insulting other users, and causing chaos.
I don't claim to be using this workflow at all.
Even worse, you're the only one complaining about this workflow.
If you can't use it: use another one on Civitai or elsewhere; there are plenty available, more optimized, functional, and so on.
If you can't use the code I developed: delete it and use standard code. No one is forcing you to use it.
If your computer's performance is lower than mine with the same graphics card, it's possible that your Comfyui is experiencing a problem (possibly due to SageAttention, Flash, or other software being installed).
In any case, we're not on GitHub to come and complain and throw a fit about my code.
And you're absolutely not obligated to use it, I repeat.
I created this code suite, and it's been downloaded countless times. You're the only one complaining. On the contrary, some people thank me because, thanks to this code suite, they can generate videos on WAN 2.2 more quickly.
Instead of posting comments on the code, go to [GITHUB] and post a request.
I suggest you adopt a working method that suits you better, like your initial method where you create your female characters in a loop, for example.
@XT_404 You must be an absolute morron if you think AND PUBLICLY state "I don't claim to be using this workflow at all."
"Even worse, you're the only one complaining about this workflow."
You should know that 90% of Downloaders are not using the in the Description buried Workflow, and of those 10% that use it, 97.5% won't complain but just discard it because of incompalibility, unwillingness to DL your Nodes...
So let me do the Math with last I looked 700 Downloads... 2 Persons that would complain
TWO PERSONS WOULD EVEN CONSIDER complaining
"Thanks to this code suite, they can generate videos on WAN 2.2 more quickly."
Not seen once. In Fact, you did not even mention that out of this Thread.
And even if, WOW, the Person who had a Workflow that took 10min, now creates in 8min. Proof anything different. Show any results that are worth a damn, not like the Videos you claim to have created in 100-210sec, which look pretty damn low-res for the most part.
But why start putting out CONCRETE NUMBERS for your !!!NOT EVEN USED!!! Workflow.
This is a fraud from A-Z, and there is only doubt to be had about any of your Claims. I damn sure am not convinced of one Word of yours. This is one Leftist Goal-Post shifting, making Stuff up as we go, if I have ever seen it.
To prove my point, you can tell me how many people volunteered to Beta Test your "new Version", and how that went without you giving out any instructions...
Let me guess, something between 0 and 1, how ironic xD
@XT_404 "I suggest you adopt a working method that suits you better, like your initial method where you create your female characters in a loop, for example."
I suggest creating anything at all for your "Libidous" called Model that you so carefully outfitted with as little as possible, and excluding anything 18+
What the F are you even doing?! Complaining that people create Stuff on your Model??? Ouf yeah, let's cherish all those Creators that create Amateur Porn that can be found in Heaps on any 18+ Website, but scorn the Guy that found, used, and perfected the bug in your Model to the fullest extent.
@XT_404 18 Days since Release... 18 Video Posts
slow clap*
There you go. Happy New Year :)
@FrozenGT https://www.swisstransfer.com/d/a24224e0-7e0c-4917-a4ea-4e5ce72a1995
What better proof than video?
So your PC or your Comfyui has a serious problem!
Looks great. Can we get the GGUF quants
@OzzyOsman : There will be no GGUF version.
This model will be as is.
It is quite lightweight; switching to Q6/Q8 thread will have little impact on silk.
@XT_404 are the light 4-8 step loras baked in? what CFG do you recommend?
@OzzyOsman 6 step or 8 ,
@XT_404 its a shame not Q8 because that Quant have the best output quality so far but I respect that and thank you for your work! and It's okay to block those fake moralists and envious people who are good at criticizing but not at contributing quality stuff, many of them make very poor quality Loras and think they're a big deal.
@Ronaf99 : Thank you so much for your message and yes you're right ^^, but we only see the bad in others ;)
It works... It does what I want to see... The question is, why is there so little explanation? For example: Does this model work without other LoRa "Acceleration" options? - Yes, it does. What video sizes are best? - Anything starting at 512 x 512. And many similar minor questions for people who aren't as well versed in WAN 2.2.
@DevourerOfSins : 640 x 640 or 720 x 720. For cfg: 6 to 8. No resolution problems were observed.
If you would like more information on the workflow or custom nodes, simply read the readme.
@XT_404 Thank you very much for noticing my questions. I understand everything.
no info. github page doesn't exist.
why should anyone use this? why bother???
@kronos1959777 : You don't have to use it :)
What do you mean, the page doesn't exist? Everything opens fine.
@noisedreamart : https://github.com/XT-404/XT-404_SKYNET
@noisedreamart it was error 404 when I posted.
whats the best sampler settings for realistic videos?
2.5/1/1 and prompting : realistic,
@XT_404 how many steps and what samplers are better?
@Realmodels : 6 and Euler
Supper excellent base model, effective motion and high understanding to the prompt. thank you very much for sharing a such good model!!!
Love it! Thank you for sharing!
Your Wan models are the absolute best—nothing else is even close. Thank you for your efforts and Merry Christmas!
I noticed the workflow provided uses a clip_vision node named "model.safetensors". Wan2.2 doesn't use clip vision, so I found one with the same name model.safetensors at huggingface, but it's for SD15. Is this correct?
for you : https://www.swisstransfer.com/d/7087d1bb-7895-4fbf-a6ba-4bb0b1aff94f
model.safe
"Model" isn't a very descriptive name. What is it, specifically?
@ss9999 : It's simply a model :) nothing more, nothing less, which allows the whole system to function.
@XT_404 Predictable troll response. Using that troll reasoning, every model in safetensors format would be called model.safetensors. Obviously, as anyone can see, that isn't practical.
So, maybe consider answering the question instead of trying to be cute.
I'm guessing model.safetensor is his own clip vision model, its a comparable size to other clip visions used with wan2.2, but yes, clip vision is used in other workflows, it helps with consistency in i2v workflows. You can use the authors model.safetensor or clip_vision_h or clip_vision_l... it works with all of them.
@brendacaro We shouldn't have to guess.
@XT_404 Is there another link to this file?
@lolo780 : for you : https://www.swisstransfer.com/d/a0cf7561-bcec-4457-a55b-d0777413fff7
hi when i load the workflow i get "WanVideoDualInterpolator" missing but i've already used ur github code thingy in comfyui?
@seanhan19911990198 : I suggest you uninstall the Skynet suite from your nodes and reinstall it; it's strange.
@XT_404 just tried that to no avail, even did a complete update of all nodes. poop :(
@seanhan19911990198 It's really strange that I just did it and I have no problems, all the nodes are present and everything works.
@seanhan19911990198
Microsoft Windows [version 10.0.26200.7462]
(c) Microsoft Corporation. Tous droits réservés.
N:\ComfyUI\custom_nodes>git clone https://github.com/XT-404/XT-404_SKYNET.git
Cloning into 'XT-404_SKYNET'...
remote: Enumerating objects: 239, done.
remote: Counting objects: 100% (85/85), done.
remote: Compressing objects: 100% (85/85), done.
remote: Total 239 (delta 44), reused 0 (delta 0), pack-reused 154 (from 1)
Receiving objects: 100% (239/239), 1.58 MiB | 13.24 MiB/s, done.
Resolving deltas: 100% (125/125), done.
N:\ComfyUI\custom_nodes>
╔══════════════════════════════════════════════════════════════════════════════╗
║ CYBERDYNE SYSTEMS CORP. | SERIES T-800 - MODEL 101 | V3.7 ║
╠══════════════════════════════════════════════════════════════════════════════╣
║ ║
║ █ █ ▀█▀ █ █ █▀█ █ █ █▀▀ █ █ █ █ █▄ █ █▀▀ ▀█▀ ║
║ ▄▀▄ █ ▀▀█ █ █ ▀▀█ ▀▀█ █▀▄ █ █ ▀█ █▀▀ █ ║
║ ▀ ▀ ▀ ▀ ▀▀▀ ▀ ▀▀▀ ▀ ▀ ▀ ▀ ▀ ▀▀▀ ▀ ║
║ ║
║ SYSTEM BOOT SEQUENCE: INITIALIZED ║
╠══════════════════════════════════════════════════════════════════════════════╣
║ NEURAL NET CORE (XT-404) .............. [ONLINE] Prompt-Lock: ACTIVE ║
║ INFILTRATION UNIT (GGUF) .............. [DETECTED] ║
║ T-3000 PHASE CONTROLLER ............... [ONLINE] Omniscient HUD: ACTIVE ║
║ OPTICAL SENSORS (VISION) .............. [CALIBRATED] ║
║ AUTOMATION SUBROUTINES ................ [ONLINE] Smart-Scale: READY ║
║ TARGETING SYSTEMS (ACCEL) ............. [LOCKED] SDPA: READY ║
║ MIMETIC RENDERING (GEN) ............... [ONLINE] FP32 Core: ACTIVE ║
║ POLYMETRIC ALLOY (T-X) ................ [ONLINE] Start/End Engine: READY ║
║ DATA COMPRESSION (H.265) .............. [ACTIVE] ║
║ CAMOUFLAGE UNIT (COLOR) ............... [ONLINE] ║
╠══════════════════════════════════════════════════════════════════════════════╣
║ DIAGNOSTIC COMPLETE. ║
║ >> 18 COMBAT MODULES INITIALIZED. ║
║ Check : [V] XT-404 Core ║
║ Check : [V] Cyberdyne Hub ║
║ Check : [V] Cyberdyne Genisys (T-3000) ║
║ Check : [V] Wan Vision Cache ║
║ Check : [V] Wan Resolution ║
║ Check : [V] Wan Text Cache ║
║ Check : [V] Wan Auto Helper ║
║ Check : [V] Auto Half Resizer ║
║ Check : [V] Wan Accelerator ║
║ Check : [V] Wan Terminator ║
║ Check : [V] Wan Fidelity Gen ║
║ Check : [V] Wan Ultra Gen ║
║ Check : [V] T-X Interpolator ║
║ Check : [V] Wan Compressor ║
║ Check : [V] Wan Chroma Mimic ║
╚══════════════════════════════════════════════════════════════════════════════╝
[T-800] I'll be back... waiting for prompt.
Import times for custom nodes:
0.0 seconds: N:\ComfyUI\custom_nodes\websocket_image_save.py
0.0 seconds: N:\ComfyUI\custom_nodes\Wan_Model_Healer.py
0.0 seconds: N:\ComfyUI\custom_nodes\ComfyUi-Scale-Image-to-Total-Pixels-Advanced
0.0 seconds: N:\ComfyUI\custom_nodes\ComfyLiterals
0.0 seconds: N:\ComfyUI\custom_nodes\ComfyUI-Olm-DragCrop
0.0 seconds: N:\ComfyUI\custom_nodes\ComfyUI_ColorMatch
0.0 seconds: N:\ComfyUI\custom_nodes\comfyui-frame-interpolation
0.0 seconds: N:\ComfyUI\custom_nodes\ComfyUI-GGUF
0.0 seconds: N:\ComfyUI\custom_nodes\rgthree-comfy
0.0 seconds: N:\ComfyUI\custom_nodes\ComfyUI_essentials
0.0 seconds: N:\ComfyUI\custom_nodes\comfyui-kjnodes
0.1 seconds: N:\ComfyUI\custom_nodes\comfyui-videohelpersuite
0.1 seconds: N:\ComfyUI\custom_nodes\XT-404_SKYNET
0.3 seconds: N:\ComfyUI\custom_nodes\ComfyUI-Manager
0.3 seconds: N:\ComfyUI\custom_nodes\RES4LYF
0.4 seconds: N:\ComfyUI\custom_nodes\ComfyUI-WanVideoWrapper
@XT_404 typical :( well this seems to be my only issue with my comfyui at the moment, so i'll figure out how to use the models in a slightly different manner, thanks though as always
@seanhan19911990198 The problem therefore comes from you: and not from the code.
C:\Comfyui\custom_nodes>git clone https://github.com/XT-404/XT-404_SKYNET.git
Cloning into 'XT-404_SKYNET'...
remote: Enumerating objects: 239, done.
remote: Counting objects: 100% (85/85), done.
remote: Compressing objects: 100% (85/85), done.
remote: Total 239 (delta 44), reused 0 (delta 0), pack-reused 154 (from 1)
Receiving objects: 100% (239/239), 1.58 MiB | 3.71 MiB/s, done.
Resolving deltas: 100% (125/125), done.
C:\Comfyui\custom_nodes>
redo comyfui
This workflow has missing nodes
This workflow uses custom nodes you haven't installed yet.
WanVideoDualInterpolator
Install these nodes to run this workflow, or replace them with installed alternatives. Missing nodes are highlighted in red on the canvas
bleh
@seanhan19911990198 Is it in the workflow linked to the custom version that it's missing?
@XT_404 @XT_404 @XT_404 ah my bad i tried to use one of your videos as the workflow as im always interested i the prompt, i used WORKFLOW X-T Dual Phase from the website and all is good :D if ur bothered it was this 1
https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/644b9024-826c-4b18-9a4b-a32efc21ec21/transcode=true,width=450,optimized=true/wan22_rife_00159.webm
Auguswrath is back. In the description, he tells us that merges are trash and that only he does fine-tuning, haha! This guy respects no one. He’s the same one who said the same thing about his Libidinous and Omega models he had to delete them because some people went to his model's page to prove it was just a merge. He then changed his username and of course deleted his models (proof). This time, he was smart enough to wipe the metadata, and he keeps lying. For those who still have the Libidinous and Omega models, take the safetensors files and drag them into your ComfyUI interface: you will see the merge and just how much he lies and has no respect. This gentleman who said he didn't merge has apparently managed to stop merging... Of course, you are someone trustworthy... and weirdly enough, his model has the same flaws as other merges, surprising :) Auguswrath, So you are confirming that you didn't merge any Lightning model here?
All I know is that he pointed people to a file called model.safetensors. When I asked him what model that is, he said "it's a model, nothing more, nothing less." How precious.
@ss9999 I really don't understand this guy's goal. Why lie? Why belittle others' work while stealing from them? His old models were nothing but merges of other models, and now he dares to talk down to others.
Stop criticizing him. I've already said this to other users. Whether it's a merge or not, he worked hard to create this template. The main thing is to have a template that works. I encourage you to merge as much as possible. We're here to share, and for free. I think the detractors should shut up. I see that he's not sharing with the community—there's nothing in their accounts, no images or videos, so they're the ones profiting from it and then reselling our content on other sites. Shame on you!!!
Me personally I am so very OK with his models that you "who criticizing him" call them "merges",
@hazzoom82659 ♥ thx.
@lighthorsexajz830 My comment is not a critique, I am just setting the record straight. On the other hand, this guy doesn't just criticize other people's work, he calls them incompetent while stealing from them and making merges. Do you think that's normal? Please read the description. He claimed he paid thousands of euros to make his model, yet many people checked and found they were just merges. And now this guy is insulting others... Seriously, are you defending that?
I don't have a problem with his model being a merge either, everyone does that. The problem is that this person takes us for fools and, on top of that, insults the work of others that he is stealing. There are rules on Civitai, and what he is doing goes against those rules. He doesn't know it yet, but even without metadata, we can find out how the model was made using tensors and SHA256.
@heyad ai is full of grifters and scammers. the fact theres zero info about this wan "trained" model but a long rant about "haters" tells me more than enough already.
@heyad just stfu man lmao
Stop criticizing him. I've already said this to other users. Whether it's a merge or not, he worked hard to create this template. The main thing is to have a template that works. I encourage you to merge as much as possible. We're here to share, and for free. I think the detractors should shut up. I see that he's not sharing with the community—there's nothing in their accounts, no images or videos, so they're the ones profiting from it and then reselling our content on other sites. Shame on you!!!
@lighthorsexajz830 :
Everything has been said.
Thank you ♥.
For your comment.
They can criticize all they want, their judgment is based on a metadata anomaly.
Metadata, as mentioned, that I could have removed, but I didn't because I already knew this problem existed and I didn't care at all.
But it seems that for some people with low IQs, even erroneous metadata constitutes overwhelming evidence. I don't advise them to become lawyers.
Let them waste their energy criticizing into the void.
There's a proverb that says:
Silence is the best answer to fools; their answering machine educates them.
@XT_404 No problem, I say what I think. I don't like ungrateful people. Thank you for your effort. I love your models ;)
Honestly those attacking you XT_404, they look so much like shaming & bullying, I remember what they did with your past version (Insidious) , & here after so short time of you posted this new Exitium version, the criticism is back again like being attaracted with a magnet !
@lighthorsexajz830 That's very kind of you, and thank you so much.
I love that you're using it :)
You have talent; I encourage you to continue :)
@hazzoom82659
Indeed, while we were discussing this privately,
a wave of harassment (from four people) erupted. It seems a small group has conspired to try and damage my image and my model.
I already know who's behind this: just look at the most downloaded merged models.
Among them is someone who enjoys spreading all sorts of nonsense about me, whether on the French Discord server, the Civitai Discord server, or elsewhere.
Let them say and think what they want.
After all, they're just sheep; they can't stop you from fully enjoying the model and pushing its limits.
Thank you both for your support. : @lighthorsexajz830 and @hazzoom82659
♥♥♥♥♥♥
@XT_404 I am not yet sure what do they want, do they want you to stop ? you don't unless you really want to, not because they want it, do they try to prevent others from having your model? no, you are sharing it for free like many other creators here, you offer a functioning model (few days ago someone shared a version of "Nunchaku Qwen Edit" & just costed me bucket data & time downloading it, to discover it is only working for RTX5xxx series which I don't have & just because he didn't cost himself few lines of writing this in the model's name or description & he doesn't easily reply comments to answer question) you don't do all of that, your models work & you reply people trying to ask you about whatever, no shame at all, good on you :) . I still say it, you & DaSiWa provided us very nice custom WAN models until now, you are not the only great ones for sure, but you gave me free gifts, thanks ;)
@hazzoom82659 :
Thank you very much for your message.
This model, in any case (and I'm not bragging), has reached its absolute limits for a 13.3 GB model.
It's so heavily modified and overloaded that it's become impossible to continue training it.
I wanted to fix the missing anus issue.
The model no longer supports it.
It has far too many modifications, whether in the expert or router versions, so that the slightest improvement or training completely ruins it.
I wanted to provide users with a concrete, trained, and clean model so they could clearly see the difference.
This model has flaws, yes, it doesn't create anuses laughs.
But apart from that, everything has been reviewed, corrected, optimized, and strengthened.
Even if I were paid to do better tomorrow, I couldn't achieve that with this model.
I've reached the limit, and going beyond it would be counterproductive.
Especially if I want to maintain quality, fluidity, and consistency.
I'll continue, and I'm continuing development of the [SKYNET] suite, which is progressing well.
This pushes the model even further. I'll probably try to create a Skynet workflow focused on one-minute videos.
We'll see if I succeed ;)
In any case, thank you again so much for your messages of support and your kindness.
With all due respect, XT-404
@XT_404 I'm enjoying using your Omega Uncensored libidinous V2, INSIDIOUS LIBIDINOUS DRAFT V4, and your new EXITIUM VICTRIX models. They truly impress with their wide range of capabilities. The results often exceed all expectations. Each model is unique in its own way. Now about the attacks. I read the caustic comments from all these "people" even before you deleted your model thread. They say truth is born in debate. Yes, that's true. But only with intelligent and reasonable people, not with those filled with envy, hatred, and shallowness. And here it would be appropriate to discuss jointly improving the models, rather than hurling vitriol at someone who has managed to create something good. Therefore, I support you and urge you not to be distracted by ignorance, rudeness, and envy. Continue to inspire people to be creative; you are doing an excellent job of it.
@marsele117268 I completely agree with you, that would be ideal, but I don't think it's a good idea with this guy. To understand why, you have to know what happened with his latest models someone simply told him to stop telling everyone he had spent thousands of dollars on finetuning when he was just taking models and merging them (personally, I call that very easy work) but that's not the point, it's more about the ethics. And that guy, xt_404, did nothing but hurl insults to defend himself. To work together, you have to start by being honest and, above all, respectful, and he is proving once again that he isn't you just have to look at his description.
So these are your arguments? I haven't shared anything here myself, so I guess I don't have the right to speak? I share a lot on GitHub and Hugging Face, so should I ask people who comment if they share resources, and if not, they should keep quiet? Ridiculous... What you call "hard work" is just merging other models together when some creators don't even allow it? I call that easy work and theft, but to know that, you would need to be aware of certain things or do some research beforehand unfortunately, that isn't the case for you. One last thing, tell me how you operate it bothers you when we criticize this person's actions? Then why doesn't it bother you when he does the same, or even worse, in his description?
@XT_404 Watch out sir, we are a sect plotting against you in the shadows our machiavellian plan is to harm you... Is that the best you could come up with to get out of this? It's called accusatory inversion. And your stories about metadata corruption is that really the only excuse you found to keep mocking people? You claim to do custom model, yet the metadata corruption shows "automerger" nodes, The name of the "automerger" node is also a bug, I guess? It is still quite surprising that your bug follows the exact same process everyone uses to merge models what a coincidence, isn't it? loras and checkpoints, featuring their exact names as if by magic, and then you complain about people criticizing you? Are you joking? Keep making a fool of yourself, it doesn't bother me.
If it doesn't bother you that this person steals the work of others while disparaging them and lying to the community, fine but don't come lecturing those who dislike these actions when you say nothing about him insulting people in his description. This is called intellectual dishonesty.
@fondaIA You talk sooooo much ! The way you are talking & reacting to XT 404 looks so much to me like trolling & bullying & honestly I saw you some users started attacking XT-404 in a rude way, with my own eyes even someother user here in civitai just shared a workflow for one of XT-404's model & another user just commented attacking him like (he is just grabing a random github thing & sharing it here in civitai),, you really act very suspicious & what you are doing will not stop us from using his work just like many other creators here sharing their work with us,
if you have issue with that? can you go & complain to civitai Admin team about their own policy & their website open to all of us ? instead of you attacking Augusts's posts like trolling it ?
@marsele117268 thx ♥
@fondaIA :
The monologue was very enjoyable. Is there an end to this discussion? Or a season 2? Is a Netflix series adaptation a possibility?
@XT_404 A monologue is actually someone talking to themselves here, I have only been replying to you and to other comments open a dictionary and you'll see. You say you want to end the discussion, yet you keep talking; I don't understand?
@fondaIA :
You wrote four paragraphs without getting a response.
That's what you call a monologue.
Anyway.
Have a good day. I have other fish to fry.
@hazzoom82659 I won't ask you how I'm trolling, as that would be too hard for you to answer. However, I can personally give you an example of a troll, someone, for instance, who takes other people's work while insulting them, or tells us that metadata is a bug in a model that generates videos in 4 steps, Unlike the creator here, I trust people's intelligence, and that is why I am here. Think about it carefully this person claims not to merge, yet their model generates videos in 4 steps. If you enjoy being mocked, I repeat, that's fine, but don't blame others for not liking it. Finally, tell me so I should complain to Civitai? Thanks for the advice, but I know what to do. And if I may, before giving advice, apply it to yourself, that would be smarter. Instead of complaining about my non-existent trolling, go complain to Civitai instead.
@XT_404 You call those paragraphs (open a dictionary)? Seriously? I just sent a reply to certain comments and here I've done nothing but reply to your comment again.
@fondaIA Really ! LOL, just as I said, you talk sooooo much, maybe you are enjoying wasting our time too ?!
Pretty dang good model.
@Tompte : Thx :)
@XT_404 Got really good results after a little bit of prompting. No loras needed.👍
@Tompte :
For some features, a Loras is required.
This model offers many possibilities, but it's missing a few.
For example, it doesn't allow you to create anuses.
Furthermore, sometimes, in certain positions and from certain camera angles, the act doesn't play out correctly during anal intercourse.
This depends on the camera angle.
Otherwise, it works very well without a Loras, yes :)
The example workflows can't find my models....what could I be doing wrong? I've got my models in the diffusion_models folder
@megamau5 :
Hello, for the Safetensor models, they should indeed be in the diffusion_model folder.
For the GGUF models, they should be in the unet folder.
Please check that you haven't installed your models somewhere other than these folders, or if you've modified the Comfy path, it won't work.
My models aren't populated on the CyberdyneModelHub node; only "NONE" appears when I attempt to select high and low models.
@voboyso : Hello
Please install your .safetensor models in the diffusion_model folder.
If your models are from GGUF, install them in the UNET folder.
@XT_404 Thanks for the info! Why doesn't it just look in the same places as ComfyUI, though? All of the other model loaders see all of my models as defined in the 'extra_model_paths.yaml' file. I think a model loader should look in the global directories, but please enlighten me if I'm mistaken! Also, this all seems to work quite well so far, but the coloring can get weird. Also, I keep getting OOM on first run, but then it seems to go through on second try; I even switched FP32 to FP16 on the image to video node. Thanks for your contribution!
@voboyso :
The code is still under development (I'm developing it myself little by little).
Regarding "Out-of-Mind" errors, users with less powerful graphics cards (6/8/12 GB) generally encounter this problem due to FP32, which consumes a lot of resources. I suggest you try BF16 or FP16.
Both work perfectly; there is no noticeable loss. FP32 allows for greater precision and a higher level of detail. BF16/FP16 do the same, but with a lighter processing load, resulting in a slight, but very minimal, loss of detail.
For the moment, I'm focusing on optimizing the workflow and the code so that it's compatible with all users (hence the choice of FP32/BF16/BF16).
As they say, you can't be in two places at once :)
@XT_404 Okay! Thanks so much!
This is an amazing and incredible model. I think it's truly fantastic, beyond words.
@alien55888 :
Thanks a lot
I can't start the workflow because I don't have an INT node. I've already asked the neural network for everything I need, but it doesn't work. I can't find the INT node anywhere :(
@Renessance : rename for : ComfyLiterals
Thank you, I think this comment will help other users as well.
SKYNET.experimental, which NODE allows you to control the output video resolution?
I want to get a video in 368x480 resolution, but I don't know where to set these values in the workflow.
@Renessance : It doesn't allow that type of resolution, no. It already applies upscaling at the end of the process.
The GOAT is back in business!
@markopolo791 : lol, thanks ;)
You cooked
Works perfectly with svi2 pro ;)
@jaimybattesti958 : Delighted with your feedback ^^ and yes, it works very well with SVI :)
or any other lora
Agreed.. works great with SVI, doesn't need lightning loras, and even at 4 steps and sigmasplits at 2, the results are great, and faces stay consistent. Really good job!
@artificialotaku , THX ♥ :)
Hi. The model is good, but my computer freezes for several minutes. Please make the q6 gguf model.
@nubirq : Hello, I understand your request for a GGUF Q6 version.
Unfortunately, this isn't possible for me.
I've already tried with LamaCPP, but it's impossible: the program crashes or constantly encounters errors.
I spent five days on it without getting any results, not even a working solution. Therefore, I unfortunately won't be developing a GGUF version.
That's amazing, that's amazing. After using your model, I'm even considering buying a 5090.
@syjrsan148 : Thx :)
THIS is an extremely good model! it could be easily the best ...
... but it has lightning and saturation stability issues.
does any1 experience this as well? mostly when generating past second 2.
@fbbcool701 : Hello. You're the only person experiencing this problem. The model works perfectly without any issues. Check your negative prompts or your settings.
hmm ok. i am using an 8step lora, maybe that's the issue. it mostly happens when there is a lot of pale skin in the f0 image, then the video gets overexposed from ca. sec 2 on. (btw i do not if u merge or what :D, you are doing great!).
@fbbcool701 This isn't a merged model.
The model works in 6, 8, or 10 steps.
Don't add LoRa to reduce the number of steps; that's very likely the problem.
I have the same issue as described above, and I'm using the default settings from the workflows in the official Github, and no loras. I've tried to change some settings, but I have no idea yet what to try, as I'm quite new to all of this. Do you have any suggestions, if you fixed your problem? (great model, btw! I'm having a lot of fun experimenting with it)
@dazoner :
This could be related to one or two custom knots, but the silk models don't have this issue.
Disabled certain non-essential custom knots in the final image processing stage :)
@XT_404 Do you mean that I should try to disable some of these: Wan Spectre Chroma, OLED Dynamix, and Organic Skin ?
@dazoner , yes, oled, chroma :) and look
Love this model when I want something that moves around a bit more! The only thing I can't get it to do is stay still haha. What's the max frame length it works with (can we push it past 81 frames)?
@jeremygin : Hello, yes, it's possible to exceed 81/93 without any problem.
To prevent the character from moving, you need to lock it using prompts so the model understands it must remain still.
Great model for adding more motion.
@organic10255121 : Thx :)
THE GOATTTTTTTTT
This is, hands down, the best model out there. Handles No LORAs, 2 steps high noise, 2 steps low noise, CFG 1 -- and I get the best results of any T2V or I2V out there. Handles every resolution you throw at it -- small (240x368) or large (720x1280) -- without skipping a beat. Other models fall flat on their face if you don't prescribe to their exact resolution requirements.
@primez19 : Big THX :) ♥
Hello, please make a model of NVFP4
@nubirq :
Hello, the answer will be the same as for those who asked me for a GGUF model: No.
The model is already lightweight enough; modifying it won't improve anything and is very likely to ruin its visual appeal and quality.
If you want a modified version, I suggest you do it yourself.
This model is amazing. It works with the ComfyUI templates, or some other workflows, but it truly shines using the github and the workflow from one of the sample images for the model. The funny thing is, even those workflows don't use all the tools and nodes described on the github, this thing goes deep. It's extremely fast, produces incredible quality, is capable of motion and effects that I'm honestly not sure you can get out of any other model, and the console messages are detailed and funny. This is not just another model, this thing was crafted and the work that has went into the back end support and comfy nodes is truly incredible. I really can't say enough about it. Absolutely amazing work.
@dumpy50222 : Thank you so much for your message, it means a lot to me.
Thank you. ♥.
oui c'est génial et je te rejoins sur la qualité du travail et du développement d'AugusWrapth. Souvent des pépites passent inaperçues pour certain, d'autres n'en croient pas leurs yeux et leur chronomètre. ;-)
@sdktertiaire2 : merci beaucoup :) ♥
Currently, I've used almost all I2V models, as this is the model with the best SVI support.
@Yoxhachi : THX ♥ :)
Agreed, I use it daily with SVI and I haven't felt a need to replace this model for any other so far.
without lighting loras what's the recommended steps and cfg? scheduler and sampler combo?
@obsidiancloud :
6 steps, CFG 2.5 and 1 and 1 , shift val 5.sampler linear/euler.
want some confirmation. its cgf and steps. for a regular workflow 1.0 cgf with 3 high 3 low. for Skynet 2.5 cgf with 1 high and 1 low. is that right?
@gayanholmes787 : Yes right.
@XT_404 thank you for quick reply
Miss You 😢. What do I have to pay to get a text to video model made? 😄😅
@BigSad11 : Hello, I don't create custom templates.
This is one of the most efficient and viable templates I've created and made available.
I've already published several templates and removed this one a few months ago.
The financial investment is far too high.
Use this template; it's more than sufficient. :)
Just use a black/white/gray/noise texture instead of an actual ref image, and the model will "hallucinate" the whole action, actors included if you don't want ref images to "muddy the waters". Now if you want T2V for performance reasons, then it's different.
Have a good day!
@rmnot Thanks. I discovered this but the results are well, meh, with t2v. Absolutely love this model for I2v. Maybe the best one out there. But using it in this way for t2v just didn't give me the results I wanted. Could be I'm doing something wrong. 🤷
im having a hard time with prompt guidance, it seems to ignore half the stuff i prompt, especially after 3-4 sliding windows
@NUGGZ1616 : ???? 3-4 sliding windows ???
@XT_404 continuing videos - sorry using wanGP
This is my main model is and it's kinda very touch-and-go with the prompts... best practice i've figured out, is making sure the prompt for the second video continues the action at the end of the first video, and so on.
Astonishing quality
I've tried WAN 2.2 Enhanced NSFW, Smooth mix and DaSiWa side by side, same seed and everything.
The details of a person moving is unmatched. If someone gets tugged slightly they'll prop themselves up with their hands. A woman fingering herself will thrust her hips while she does it. Other models just don't get that detail. Facial expressions amazing as well. Should be at the top of Civit AI
I hope you comeback and update to show the others how its done every once in a blue moon.
Does your test use a simple natural language prompt or a complex prompt?
@kaytransg196 :
Hello, the model supports simple prompts, advanced prompts, heavy structure prompts and point-by-point prompts with timecode.
I have the perfect workflow settings and prompt for a vs, Remix vs Dasiwa, Remix has better complex movements adherence results so far.
I will try this.
Do you plan to make a model for LTX?
@dav79mail156 :
Hello, thank you for your comment. Unfortunately, this isn't possible. LTX uses Safetensors.
Unlike WAN 2.2, which uses two separate models, High and Low.
Furthermore, LTX supports audio, which WAN 2.2 does not.
I would have loved to do it, but unfortunately, it would take a considerable amount of time.
If, in the near future, it becomes possible to use the training dataset to create an LTX 2.3 model, I will give it a try.
In any case, thank you for your comment and your interest in this model.
Do you use another platform like huggingface.co? Here is so slow download.
@TieFighterPilot : The model is hosted here, and only here. Yes, the download is slow; it's throttled by Civitai.
I've downloaded the fp16 version but it is still being the fp8 version
@TieFighterPilot It's indeed the FP16 version with the size and weight of an FP8. And be sure to use umt5_xxl_fp16.safetensors.
@XT_404 Ah thats why 😞 then it is not compatible with Mac it requires full fp16 base
@TieFighterPilot Maybe try the City96 gguf converter and make a quantized version.
The epic and dramatic model name and description is intriguing and honestly the model is very impressive. Well done.
@mjh02111964582 : Thx :) ♥
Hey my friend. Stumbled over this model and wanted to try it, but the LOW Model seems to be not downloadable anymore. Any chance to fix this? Best regards.
@Mauner :
Hello, I just checked the model and it's in place.
You probably have the same bug we had a few months ago, or for some unknown reason the civitai website is preventing the models from downloading correctly. On my end, everything is fine; the H and L models are in place. I just tested it and downloaded both.
@XT_404 I can`t download low too.High is ok
@syjrsan148 : Civitai is quite buggy, but on my end, it's fine. I've tested downloading the Low and High models eight times already, and everything downloads correctly without any problems.
I can't seem to get this character to stay still on this model. He keeps turning his head, swaying, and twitching. Is there any way to fix this?
@lordzellos : Hello, you are the only person who has encountered this problem. I suggest you review your negative prompts, which are often the cause.
Hi, I'm new to this. I've been learning about this for about a month, and this model caught my attention. I started with the most downloaded models, but when you said, "It's not a fusion; it's a trained model," I didn't understand at all, so I decided to try it, and then I understood. It's a great model; it generates very good movements and good quality with low steps. I tried using other models, but I couldn't get used to it. This is the best one I've used so far.
@velocidad147 : Hello, and thank you for your kind message. Enjoy using it and have fun! :) ♥
I have tried all kinds off WAN models now with all possible combinations of LoRAs and must say: Wow how I am impressed! Even with my actual default workflow, without any XT-404 nodes and without any LoRAs despite of the lightning_lightx LoRA, the quality is amasing!
This will be definitely be my goto WAN models for the next future.
I probably tried 8 different Wan2.2 Models... none of them surpasses this one, i ended up deleting them all and just keeping Victrix for all my work.
Do you think you'll ever test the waters creating something for LTX2.3?
@ArtiOtaku
Hello,
To be completely transparent: I don't think so. The LTX 2.3 model uses a combined training dataset—mixing video and images in the same folder (like Wan 2.2 does) along with audio—and I have to admit that this is an area where I lack expertise.
If a future version of Wan were to be released with an integrated audio system different from LTX's—where separating high and low frequencies might prove useful in certain respects—the situation would be different. However, with LTX 2.3, everything is unified within a single core that is internally segmented; that is where I am unfortunately out of my depth and lack the know-how needed to work on this model.
Thanks for the positive feedback and for your message! ♥
@XT_404 No problems, thanks for Victrix, it has been my work model for all of 2026, and it doesn't seem that any other WAN models will be open source anymore, so, having a WAN2.2 this good is a blessing.. thanks again!