[00:00] You kept me waiting. The sea always gives it back.
[00:14] Hello, hello, people of culture. It's your boy Stoobz. Yet again, I ain't going anywhere. Recently I created a new workflow called Stoobelius Ultimate 8.3.
[00:30] for Minimax. And don't worry all of you LTX 2.5 lovers, I did not forget about you, because I personally love LTX 2.5. There's a lot of content now coming out for Minimax,
[00:46] and I felt, hey, let's even the tables a little bit and do something for LTX 2.5, because it does deserve the love. If you have previously used my previous Stubelius LTX workflow,
[01:00] you might remember it looks a little bit something like this. It has the seed hunting where you can select the sampler and scheduler, and then also it has the MSR, LoRa, ingredients, LoRa, everything enabled.
[01:16] But honestly, looking at this workflow, I felt it is ugly. And I decided to do something about it and also copy some of the great things that I did for Minimax, but put them into LTX format.
[01:30] Hence, we have the Stubelis Ultimate LTX 2.5. Let's go through the workflow, and you'll be seeing different kinds of videos that have been generated with this exact workflow.
[01:45] So you can kind of like see the differences that each of these settings do and how the quality turns out to be. There are quite a few changes that have been made, especially when comparing with the What Dreams Cost LTX Director workflow.
[02:03] Let's go through this workflow. And as always, there is a how-to-use note. After you get the workflow, you can read exactly what it does, how it works.
[02:15] But for now, the gold and black theme is gone. You are able to select your own theme, just like in the Ultimate A3, or you can rock the ConfiUI default if you so wish.
[02:31] However, I like the neon theme, so I will be showing you exactly that. Seed hunting has changed its position. It is no longer selected later in the workflow, rather it is here.
[02:45] you select how many seeds you want to do and as you can see that enables you different samplers and schedulers down below for this example we are going to
[02:59] be going maybe two seeds just so you see they are available there then you select your first path how many steps will you actually be running and the good
[03:11] practice is 8 however you can go a little bit lower oh sorry a little bit higher if you want then cst and first pass scale now you might know that oftentimes ltx 2.5 first pass uses 0.5 scale
[03:33] You can use that if you want, but I figured that I want to do the seeds essentially completely as their full videos. I want the full scale already there, and then if I want to upscale even more, then that's fine.
[03:50] there is one reason why I decided to do this and not 0.5 upscaling because you do lose some details when you are upscaling from 0.5 to 1
[04:05] and right now that is the part where I feel LTX 2.5 struggles as did LTX 2.3 hence I decided to do the 1.0 first pass instead of the usual 0.5
[04:20] But hey, that is up to you. You can change it if you want to go with the lower res there. For the checkpoints, a shameless promo. I recently released the LTX2Velius Remix Beta 2.
[04:34] It a not safe for work version which now has a baked in distil LoRa So you can run this without any distil LoRa It can do both safe for work and not safe for work If you want check that out But anyways usually you would be running here your normal LTX 2
[04:55] distilled model. Make sure to select the right model and if you need the distiller goes right here and the disty lora strength below it. Then you have the opportunity to do two
[05:10] additional loras. Now why not three or four? I've been testing quite a bit with LTX 2.5 and I noticed especially if you run loras with higher strength after the third lora it starts to get a little bit mushy and personally I just
[05:25] don't like it. So what you can do here is run these loras. If you feel that you need more then you know what to do you connect the models into a LoRa loader
[05:37] and off you go now you have more LoRa's available but in this node there are only two additional LoRa slots you have your usual IC LoRa this is where you
[05:50] would put your ingredients LoRa for example and you have a specific slot for MSR the Lycon MSR LTX version 1 and the MSR strength this is for the first pass
[06:06] usual text encoder VAEs and then as attention there is configuration the default and sage attention I personally use configuration feel free to change
[06:23] to Sage if you want. There are a couple VRAM speedups here, memory, efficient attention and chunk feed forward. However those are still quite experimental so I can't guarantee results if you decide to use these but these will help you with
[06:39] your lower VRAM. And also forgot to mention that if you want to load a GTUF this loader loads both normal diffusion models and ggufs so you can freely
[06:53] change ggufs here and start running them however there is no mmproj slot which means that your text encoder should be still not a gguf
[07:08] finally we decide already here what we want our final output to be. Refine strength you can think of this as denoising to a big degree. Here's an
[07:22] explanation also what it does. But refined strength anywhere between the usual 0.3 0.42 is where you would most likely start and how many refined steps or second pass steps you do. 3 is default, sorry 4 is default, 3 is good.
[07:40] also you can experiment and decide what you want to do. Refine sampler is where you select which sampler do you want to use as your refiner it will use the
[07:53] default scheduler as recommended by Lightrix which means that here you simply select from the list of samplers what you want to run. For the second path
[08:09] you need another MSR lower strength and refine MSR reference. If you don't know exactly how to use these you can go to Hugging Face, Lycon Studio
[08:22] MSR LTX 2.5 and check their guidance on how to run the second pass and even the first pass. However these are the usual settings that you would be running them with so I put them as default. For audio this is another part where both actually
[08:41] Minimax and LTS struggle with a sometimes when you are upscaling and refining the audio gets distorted and it's not as accurate as in the first pass hence I decided to keep the sieves audio locked meaning that whatever is in
[08:59] your first generations for the first pass they get passed on as is for the refiner the refiner second pass does not invent new sounds then you have some interesting things here you have final
[09:14] resolution you can decide here what you want your final resolution to be 720 1080 2k 4k you select already here Also you can select what your final SPS will be This is tied to rise interpolation which means that you can go 25 30 35 48 60
[09:41] I mean, even 120 if you want to. I can't guarantee that works too well, though. Anyways, here you can select what you want your final video output FPS to be. And as a reminder, if you are running MSR LoRa, then I think what you should not do is put anything here, but rather select the proper FPS right here in frame rate as the director.
[10:10] MSR LoRa wants to natively be done with 50 FPS. So here you would select 50 FPS if you're running MSR. That is just their recommended setting. And that means there will not be a frame interpolation
[10:28] at the end. And you keep this value at zero. Meaning that if you want to do your FPS selection here in the director node you keep it to zero here and that disables rice interpolation
[10:44] and for upscaler we have our vsr as well as dlss5 with a color lock because dlss5 tones down the colors quite a bit at least in my test so i created a color lock which uses the exact colors doesn't
[11:04] do anything to them from the first path or final output before actually then upscaling. And we are using mostly DLSS5. I've been testing it a lot against VSR and I feel it's
[11:19] quite a bit superior. But it is slower. I have to give it to VSR, it's freaking fast. DLSS5 is fast as well, but it's nowhere near the speed of VSR.
[11:31] You can decide to use the DLSS file or BSR depending if you want to do upscaling or not.
[11:43] It's completely up to you how you put these settings. The director node I'm not going to go through anymore because I believe that if you have used director so far you know exactly what to do here, how to do your prompts, global prompts, etc.
[11:58] Just remember if you are using MSR toggle the MSR mode on if you are not using MSR toggle it off as simple as that your reference sheet would go here and
[12:13] Just as a reminder if you want to refer to your reference sheet Then you would type here at ref1 and it would refer Exactly to your first reference slot at ref2 refers to your second reference slot
[12:27] I've been testing MSR quite a lot with this workflow and it works. For all of you doubters who say MSR doesn't work, well it works and you will be seeing some
[12:41] examples of that in this video. Here we have our seeds, the seed candidates, meaning that if you selected seeds here they will be then shown here. Naturally
[12:58] if you select one then you would only see one seed popping up here. If you select up to four then you see all of these four being generated. Live preview
[13:11] as always and the finish part. So you start from zero meaning that it does not automatically do the second path and upscaling until you have checked the
[13:26] seeds and you decide which one goes for the final path. That's why you start from zero. If you are only doing one generation you don't want to do the seeds, then you select simply from here you only put one seed which
[13:45] means that this will only be generated and then here you select one. Now it automatically goes through the entire process without stopping to wait for
[13:57] your selection after generating the seed one. Then the workflow produces your absolute last frame if you want to later on use the last frame as an image and and you get your final video And the FPS is synced to whatever you chose either at the director level or if you decided to upscale with Rife
[14:22] then it will take it from here. Overall, I feel this workflow is quite a damn good one. It's definitely something that when I run LTX 2.5, I run.
[14:37] And I guess it's a little bit of a personal question. Do you like the director? Do you not like the director? Personally, I love it. I will never go back to normal prompting anymore. And I hope you also enjoy the director.
[14:50] If you have not used my previous workflow, then I just want to mention that all of the IC videos' motion transfers, they all work. Just remember to use the LORAs if you are doing something like a motion transfer.
[15:04] be sure to put the right LoRa in and of course do the right kind of prompt.
[15:21] Other than that the workflow is pretty self-explanatory. Select your presets from here or put your custom resolutions and off you go to the races
[15:39] nothing too scientifically impossible I hope you enjoy the workflow and let me know below in the comments what do you want me to change and of course one
[15:56] One thing that I didn't remember to mention is how to get this workflow, and I think that's quite important to actually mention. There are two ways to get this workflow. One is to go to my Patreon.
[16:10] It's for my paid patrons right there at the studio tier. If you are part of the Patreon, you get it immediately after this video. However, because I do have a goal of doing YouTube and helping the community a lot more seriously, I really want to make a bit of a profession out of this.
[16:33] I do have a goal for subscribers and how I want to grow my YouTube channel. So if you are as kind as to like the video, subscribe and then comment below STOOBS,
[16:45] I will reply to you with a link where you can download this workflow and all of the custom notes directly from. And it doesn't cost you anything.
[16:58] So you can get this workflow completely for free. But of course, I do really appreciate if you want to support me also via Patreon, even as a free member. I really, really greatly appreciate all of my Patreons.
[17:12] We recently crossed, I believe, 130 patrons. So really good stuff. I really want to here also say a huge thank you to my patrons for supporting me.
[17:25] For the final thing, I want to mention that I have just finished recording my ConfUI course. It goes from absolute beginner all the way to advanced.
[17:38] and I'm currently in the middle of recording a mini course that is attached to the official ConfUI course for MCP usage which is why I've been running lots of tests with MCP also on my youtube channel you see some videos that
[17:54] have been completely created with Claude and ConfUI MCP that one I hope that you take a look at it should be dropping by the end of this week and I really
[18:07] hope you enjoy the course and get a lot of value for it. Just a little bit of a teaser for what's to come later on. But yeah, I'll be leaving you with a couple final videos just to see for you to
[18:22] see some proof that this workflow works and you can get some pretty damn good results with it. Other than that, have a great week and weekend ahead. See you in the next video. Bye bye!
[18:37] The sea always gives it back.