L. Oscar Plains

5.2K posts

L. Oscar Plains banner
L. Oscar Plains

L. Oscar Plains

@leuk_o

Bring Back Dislikes |⟀̠͌| Imagine going through life if you were never allowed to express what you don't like/enjoy. Algorithms are useless without both inputs.

California 🇺🇸 Katılım Mayıs 2016
216 Takip Edilen30 Takipçiler
Sabitlenmiş Tweet
L. Oscar Plains
L. Oscar Plains@leuk_o·
I've gotten some requests to share my LTX 2.3 @LTXStudio workflow. So I made an easy to understand and stable text2video GGUF workflow for ComfyUI. With this, I can generate 15 second clips using my RTX 3060 in about ~12 minuets. pastebin.com/JQFs13q4
L. Oscar Plains tweet media
English
1
0
5
199
L. Oscar Plains
L. Oscar Plains@leuk_o·
@Timcast It's quite politically advantageous that you took the last few days off from IRL. Very convenient timing.
English
0
0
0
5
Tim Pool
Tim Pool@Timcast·
There are some fair critiques here
Sean Davis@seanmdav

Why did Massie lose tonight? Massie went from principled libertarian during COVID, to GOP leadership lapdog under McCarthy, to anti-Trump Epstein obsessive in 2025 after tweeting about that issue a whopping three (3) times in the decade prior. The nail in the coffin for him was voting against OBBB in 2025 because, according to Massie, it did too much to secure the border. Trump mercilessly trashed Massie in 2020–calling him a “disaster” for America and Kentucky and saying he should be thrown out of the GOP entirely—but Massie easily swatted that away and won 81-19, so you can’t say he only lost because of Trump. He went toe-to-toe with Trump on COVID in 2020 and won overwhelmingly. Massie lost because he went from being perceived as a quirky but lovable nerd who seemed to genuinely believe everything he said, to looking like a clout-chasing influencer who cared more about getting TV time with Democrats on an issue he clearly never cared about until five minutes ago than he did about representing his voters. We’ll never know what caused the apparent personality change—maybe it was the death of his wife, maybe it was the McCarthy race followed by McCarthy’s ouster, or maybe it was a desire for notoriety or media acclaim and a lucrative podcasting career outside of Congress—but the drastic change was undeniable, as was the seeming lack of interest in much of anything happening in Kentucky. Blame Trump, blame Israel, blame Epstein, blame the tragic death of a spouse, I don’t care. But you cannot just wave away 2020 Massie going face-to-face with the Trump machine and winning in a rout only to get smoked six years later. Massie’s voters didn’t really change all that much, but he did, and they noticed.

English
360
104
1.6K
481.9K
L. Oscar Plains
L. Oscar Plains@leuk_o·
Stacking multiple LoRA's (at various strengths) is where things tend to get a bit complicated. One useful tool in Comfy I've found for testing/troubleshooting is Kijai's LTX2 LoRA Loader Advanced node. With this I can decide how much a LoRA can modify the video, audio or 'other' categories of the model. It's very useful for customizing how I want a specific LoRA to be applied. For example, some LoRA's really mess with the audio in a negative way but they are great with video. Solution? Use this node to nuke all audio influence of the LoRA while keeping all of the video aspects.
L. Oscar Plains tweet media
English
0
0
1
30
L. Oscar Plains
L. Oscar Plains@leuk_o·
@elonmusk With Earth being a perfectly balanced closed system, I wonder what happens when over a megaton of important material is removed from it year after year after year to never return?
English
0
0
1
88
L. Oscar Plains
L. Oscar Plains@leuk_o·
The only non-algorithmic feed is Following -> Recent. Which is newest to oldest for everybody you follow. Everything else is algo. If that is the only thing somebody wants, it's best they probably look at when the posts they are reading were made. Pretty sure everybody does this already. For example, I saw this post and noticed it was 19 hours old, which means between 0 and 1 real people will see my reply. But I wanted to reply anyway, perhaps I like vomiting my thoughts into the void where it will never be seen by human eyes.
English
0
0
1
85
Tim Pool
Tim Pool@Timcast·
The average follower doesn't know this and its super frustrating for me that periodically it switches to the algorithmic feed I need real time for news Some days i see a post and im like wtf this story is super old then duh it switched me to algorithm
Elon Musk@elonmusk

Critique of the 𝕏 algorithm is welcome. There will be monthly updates of the latest algorithm to GitHub with release notes. As reminder, you can always choose no algorithm via the Following tab.

English
54
24
515
59.1K
L. Oscar Plains
L. Oscar Plains@leuk_o·
As much as I like what both of them are doing right now, I honestly don't trust either of them. Vance has the stink of an opportunistic politician playing it safe, and Rubio is not much different. At one point, not all that long ago, they were both anti-Trumpers. If it came down to a 3-way choice between Vance, Rubio and DeSantis, I would vote for DeSantis.
English
0
0
0
32
L. Oscar Plains
L. Oscar Plains@leuk_o·
@toyxyz3 You really should provide some kind of context in your posts other than a link and an image of a reddit post. At least give a short description... or something. I'm honestly surprised the X algorithm is even putting your posts on my timeline.
English
1
0
2
61
L. Oscar Plains
L. Oscar Plains@leuk_o·
@EricLDaugh Remember when Boebert had her "Don't you know who I am?" moment when she was kicked out of the movie theater for being drunk, vaping and getting groped by somebody who was not her husband? Pepperidge Farm remembers. She showed her true colors that day and absolutely needs to go.
English
0
0
1
563
Eric Daugherty
Eric Daugherty@EricLDaugh·
🚨 NOW: President Trump announces he's looking for a GOP primary challenger to Rep. Lauren Boebert in Colorado, because she's campaigning for Thomas Massie "Anybody who can be that dumb deserves a good Primary fight!" "Even though I long ago endorsed Boebert, if the right person came along, it would be my Honor to withdraw that Endorsement, and endorse a good and proper alternative. Just let me know, or announce your Candidacy, and I will be there for you!" "Is anyone interested in running against Weak Minded Lauren Boebert in Colorado’s Fourth Congressional District? You remember Lauren moved to the District when it became obvious that she couldn’t win in her original Congressional District (The Third!) — A Carpetbagger, indeed!" "Boebert is campaigning for the Worst “Republican” Congressman in the History of our Country, Thomas Massie, of the Great Commonwealth of Kentucky."
English
913
2.7K
10.7K
605.3K
L. Oscar Plains
L. Oscar Plains@leuk_o·
@Shawn_Farash An efficient way to help wait for 20 minutes is to go outside and have a cigarette. It works wonders.
English
3
0
23
893
Shawn Farash
Shawn Farash@Shawn_Farash·
I've heard about the life hack that if you wait 20 min after waking up to drink coffee it's better for your body and it makes the coffee more effective... I've been doing that all week and I can say it has worked wonders.
English
94
17
328
34.6K
L. Oscar Plains
L. Oscar Plains@leuk_o·
Ah, is that what this LoRA does? Nice! I saw it yesterday on HF and added it to the (rapidly growing) list of LoRA's for me to try in LTX 2.3. About a year and half ago I played around with the WAN 2.1 5B funInpaint model where it could blend together two completely different pictures together in very wacky and crazy transitions (without just doing a fade or cut like LTX 2.3 does now with two completely different images). It would basically just build the scene in any way it could to make the two images work together (it would do things that wouldn't make any realistic sense, but it looked cool). It was very fun (but terrible quality, as the WAN 2.1 5B model was just really really bad looking). Since then I've always wanted to do that with the better video models (WAN 2.2 and now LTX 2.3), but these newer models just never do those crazy transitions when I do a first-last I2V workflow, they just fade or do a direct cut if the two images are too different from one another. It's more "cinematically correct", but less entertaining. And that's no fun! But it seems like this LoRA does exactly what I have been wanting to do, blend two (or more?) completely difference pictures/scenes together even if it makes no logical sense as to how they are blended. Example: If picture #1 is a woman walking on the beach, and picture #2 is a woman walking in a city with high-rise buildings, I want to see those building rise up out of the beach sand and cars magically appear as it makes this bizarre transition. This seems like it may only be for video clip transitions, but I'm going to see if I can get it to work with I2V.
English
0
0
0
101
L. Oscar Plains
L. Oscar Plains@leuk_o·
The potential of this LoRA is pretty crazy. It has some issues if the dialog isn't timed right, but is still surprisingly flexible. The folks at @ltx_model @LTXStudio are kicking down the doors of closed-source video models, and I love to see it!
English
0
0
0
17
L. Oscar Plains
L. Oscar Plains@leuk_o·
Should I share my workflow? I got this working perfectly with an older ComfyUI build (no update needed).
L. Oscar Plains@leuk_o

I managed to fix this on my own. These two nodes were the missing players in my workflow: 1. LTXV Crop Guides (this removes the original video from appearing after the generated clip) 2. LTXV Audio VAE Encode (instead of the VAE Encode Audio) - This makes the OG audio usable without errors. With these changes, the LoRA works great. It even clones the voice from the original video as best it can. Original (T2V) Prompt: A close-up of a smiling young woman's face, she has long blonde hair in a voluminous loose half-down pony tail, she says "Red... Green... Blue... Hello World!" Modified (w/LoRA - V2V) Prompt: A close-up of a smiling young woman's face, she has long blonde hair in a voluminous loose half-down pony tail, she says "Blue... Red... Green... Hello World!"

English
1
0
1
17
L. Oscar Plains
L. Oscar Plains@leuk_o·
I managed to fix this on my own. These two nodes were the missing players in my workflow: 1. LTXV Crop Guides (this removes the original video from appearing after the generated clip) 2. LTXV Audio VAE Encode (instead of the VAE Encode Audio) - This makes the OG audio usable without errors. With these changes, the LoRA works great. It even clones the voice from the original video as best it can. Original (T2V) Prompt: A close-up of a smiling young woman's face, she has long blonde hair in a voluminous loose half-down pony tail, she says "Red... Green... Blue... Hello World!" Modified (w/LoRA - V2V) Prompt: A close-up of a smiling young woman's face, she has long blonde hair in a voluminous loose half-down pony tail, she says "Blue... Red... Green... Hello World!"
English
1
0
0
35
L. Oscar Plains
L. Oscar Plains@leuk_o·
It's a very basic V2V workflow that I cobbled together, and I did actually manage to get it working (see video below). It's impressive and it matches the original video very faithfully and even gets the dialog cadence spot on (the character doesn't just blurt out all of the new dialog, it matches the flow of the original). Very cool! But I can't seem to figure out why it is including the original video at the end of the clip (the part with no audio). It's supposed to be a 121 frame video, but LTX is rendering it as a 241 frame with the new dialog + the old video (this also doubles the render time). Also I'm not sure how to use the original audio without getting an error (too many indices for tensor of dimension 3). Maybe it has something to do with LTX treating this as a 241 frame video instead of the original 121 frame. The only way for me to get it to work is by using an empty audio latent, but this means none of the original audio will be used (background noises or other non-dialog sounds that may be in the original video) I'm sure these are probably really simple fixes/adjustments.
L. Oscar Plains tweet media
English
1
0
0
40
L. Oscar Plains
L. Oscar Plains@leuk_o·
I see you have a new LoRA on HuggingFace called LipDub, I'm assuming this is what we would use on a V2V workflow to change/alter what somebody is saying in a video? This seems similar to what some other people are doing with LTX 2.3 recently, such as the Edit Anything LoRA that can modify scenes/characters/clothes etc in pre-existing videos by automatically masking off the right areas using just the LoRA and a standard V2V workflow (with a basic node such as LTXV AddGuide). Very neat! It seems every day I have something new to try with LTX 2.3, it's almost overwhelming (In a good way of course). Keep this momentum up and keep letting the Open Source community cook, as it is win-win for everybody.
English
1
0
1
816
L. Oscar Plains retweetledi
Phil Labonte 🇺🇸
Phil Labonte 🇺🇸@philthatremains·
maybe women athletes should be paid MORE than men because the point of sports is entertainment and this is entertaining as hell. 😂
English
472
3.8K
39.6K
3.1M
L. Oscar Plains retweetledi
Nick Sortor
Nick Sortor@nicksortor·
🚨 RINO BL00DBATH IN INDIANA TONIGHT! RINO Indiana Sen. Greg Walker, who voted AGAINST redistricting, has been HANDILY defeated by Trump-endorsed Michelle Davis President Trump is 6 FOR 7 so far tonight! LOTS of FAFO happening in Indiana!
English
517
6.8K
32.8K
552.2K
L. Oscar Plains
L. Oscar Plains@leuk_o·
Is there an easy way to manually set and use our own downloaded HF models for LoRA training with your toolkit? I spent hours the other night with Grok trying to do this and couldn't figure it out. The only way to use the trainer is to have it download and use the massive bf16 files from Huggingface. I would much rather use one of the smaller (and more up to date) fp8 models I've already downloaded from kijai. Same goes for the text encoders, your toolkit just defaults to the largest possible Gemma instead of using smaller fp8 variants. I must be using up around 150GB for all the model files, and it is crazy slow since it's using the bf16 instead of fp8 (which my card can run much faster on, especially with LTX 2.3)
English
1
0
0
288
Ostris
Ostris@ostrisai·
Just pushed a new version of my Automagic optimizer to AI Toolkit, Automagic v2. It automatically adjusts the learning rate per layer and has a fused backwards pass. It is extremely memory efficient, and uses less VRAM than any other optimizer AI Toolkit supports. 🧵
Ostris tweet media
English
15
10
187
8.2K

Keşfet

@Timcast @elonmusk @toyxyz3 @EricLDaugh @philthatremains @Shawn_Farash @BarackObama @taylorswift13