LoRA for LTX 2.3 puts googly eyes on people in the shot (Andrew Carr)

TL;DR

Andrew Carr tweets that someone trained a LoRA for LTX 2.3 that overlays googly eyes onto every person in a generated video. No paper, no project page, no link to a checkpoint in the visible tweet — just a one-line observation with a video attachment. Worth filing as a data point on how quickly small community LoRAs for the latest video diffusion bases are spreading.

Key claims

A community-trained LoRA for the LTX 2.3 video model adds googly eyes to every person in the shot [tweet body].

Method

Not described. A LoRA fine-tune of LTX 2.3, presumably trained on a small curated set of googly-eye composites or on a reference image, applied at inference. No training details, no author of the LoRA named in the visible tweet.

Results

A demo clip showing the effect (not retrievable as structured content). No quantitative claims.

Why it’s interesting

Connects to two things already on the wiki. First, the underlying base model: LTX-2: Efficient Joint Audio-Visual Foundation Model is the LTX-2 technical report — this is community evidence that LTX-2 has enough of a user base for one-off LoRAs to surface within weeks of release. Second, the LoRA-as-quick-personalization pattern: contrasts with the heavier “instant LoRA” infrastructure papers like Text-to-LoRA: Instant Transformer Adaption and Doc-to-LoRA: Learning to Instantly Internalize Contexts — those try to generate LoRAs on the fly from text/docs; this is the manual-training equivalent and shows the same end-state (“attribute X on every person”) is reachable with a single hobbyist fine-tune.