kabachuha / ltx2-squish-one-hand

LTX-2 Squish – One Hand Only!

Model description

Squish anything now with only one hand!

This LoRA is a twist on the Squish-style LoRAs where the object is taken from two sides and then squished inside. In contrast, the Hand-squish LoRA does it with one hand . It reaches the faraway object and... well.. squishes it. Into a mess. Figurines, characters, vehicles, granite – anything you want!

The classic recipe is used – Huber, FFN, CREPA. But... NO PRODIGY . The Prodigy run had diverged after a few hours. This optimizer was replaced with high starting LR AdamW Bf16 and polynomial schedule. In addition to this, I saw the grayish aspect of the videos slightly reduced.

The training converged in 3000 steps, 5 hours 7 minutes. (musubi swap 4 because of occasional long 121 frame videos). A 5090 was used.

This is my sixth LTX-2 LoRA and for the training recipes see my previous ones.

I'm quite glad the hands are quite anatomically correct and the occasions of six+ fingers and multiple hands are rare for me here :). However, the arms are still weird.

The dataset consisted of some organic items grabbing and then continuing them in Wan with the (two hands) Squish LoRA enabled just as the hand touches the object and a large number of the Wan version of this LoRA synthesized clips. Also, for regularization organic videos of hands opening and closed without items were added.

The ComfyUI workflows are inside the .mp4 video files.

For the Wan2.2 version of this LoRA see https://civitai.com/models/1957482.

Actually, I think this LTX-2 version is now better, spatially correct and more diverse than the Wan one. Maybe it's because of the larger dataset and self-boosting thanks to Wan's distillation.

I will release the updated LoRA on Civitai very soon.

Trigger words

You should use handsquish to trigger the image generation.

Download model

Download them in the Files & versions tab.