LTX-2.3 Paste Squeeze
Squeeze your characters from a tube like liquid paste!
BIG NOTE: This is a Last frame to video LoRA and you need to use frame "guides". They are in native Comfy, so it won't add load.
The workflow is in the videos metadata or inside kabapaste-workflow.json at my Huggingface page.
As you may notice, the videos are in 50 FPS. Otherwise, the workflow is fully default, but with 30 steps and distill lora strength at 0.6 for reportedly higher quality. It uses KJ nodes for image resize and VHS for metadata saving.
Trigger phrase is:
KABAPASTE The girl is squeezed from a tube like paste.
If the phrase is said by a narrator, add something like The video is silent with the sound of rain and squeezing sounds.
For training details, see my other recent 2.3 LoRAs on Huggingface.
Description
FAQ
Comments (15)
Okay yes but... But why?
I'm not shitting on it just Bwhuh
Because the can. Because they want to watch the world burn. Because they have the ability to make anything be squeezed out of a tube of toothpaste.
@playtime_ai_ I'm not shitting on it it's just.
Bwuh
@Valentinedesigns it's far more creative than the 3 models you've posted... Honestly, it far more creative than anything I've ever posted. I love it.
How dare you criticize this new fetish I just learned about. 😈
how do you even get training data for this
let me ponder my orb
the orb says it might be a series of commercials? but yes, where the heck you got that from XD
So, let's discuss it.
First of all, at least one classic cartoon where I live does have a squeeze from toothpaste scene - it's guaranteed this concept is in pretrain in all models (not even counting the landscape of reddit, twitter, etc.). Even if it wasn't, they all have capability to extrude objects: 3d printers, factory extruding machines, artistic pastes, etc. One simply needs to nudge it.
When experimenting how far I can push image2video into weird concept territory, I found out you can do much wider themes if you push a model towards cartoons (not only squeeze, but transformations, being stretched). And my Eat Style LoRA (strength upped to 2.0), being trained on cartoons, did just that. I used its Wan version together with also my previous Handsquish LoRA (concept of squishing objects with one hand, essential for the toothpaste) on Wan2.2. I had to disable all lightning LoRAs, set high steps and an insane cfg value of 8.0 (countered with TCFG).
Unfortunately, this unchaining magic deteriorated severely with the clip length and resolution in Wan (higher resolutions have seen less diversity), so I had to produce only 65 frames 640x640px (or lower) of the concept there. But this was enough for distillation. For this LoRA I did a lot of dataset cherrypicking, selecting only the animations where the characters have fully flown out of the tube, with the hair flowing the last and the paste fully solidifying into them with no traces. Ironially, LTX's forced body integrity helped this LoRA much to have the characters being consistent practically in all generations, because I initially tried to make a compact Wan LoRA to post here last fall and it failed to train to generate well-formed content. For LTX train, I produced around 60 such successful clips, multiple squeezes for each character for robust distribution, and then doubled them with caption dropout.
Concluding this, thanks to LTX, Wan and my cartoon / squish LoRAs, I can now produce 50 fps satisfying high resolution paste squish clips in no time 🙂
@kabachuha thanks for your shared experience.
hi, which tool do you use to create your video dataset ? (25fps, 121 frames, cropping) ? every single one I tried is not optimal. I tried shutter encoder but it doesn't calculate final frame count, but only the one before fps conversion.
Hi! I (vibe) coded my own video handling tool: mouse-based region and timeline cropping, frame count subsampling, fps change. You can find it on my Github: https://github.com/kabachuha/simple-video-cropper
@kabachuha it's looks like perfect. thank I gonna give it a try
@kabachuha mmh the fact I can't drag and drop the cropped zone, lock it to 1:1 when resizing and move frame by frame with arrows makes it very difficult to use for me
@argotic Thanks for your experience! This application is usually fine for my needs of cropping/rescaling/downsampling videos. It is a one-python-file-code and you can perfectly customize it through an LLM, I don't mind :)
@kabachuha of course no problem :). Besides of that, Ai toolkit made a new feature to "auto frame count" datasets so the only thing needed (or not !) now is cropping or resizing. Framerate and duration are automatically handled. I updated and I no more need fps/duration control. It's well better I think.. See in time if the feature is stable/become a standard.