Open‑source video trick: 30s in 6 mins
Ostris trained an LTX 2.3 LoRA of George Costanza on a single 5090 GPU in about one day, then generated a 30‑second video in roughly six minutes using ComfyUI — a flashy proof point for open‑source tooling. (x.com) (x.com)
The post came from Ostris, the developer behind the AI‑Toolkit training suite that hosts LoRA and video training workflows on GitHub. (github.com) That social post was re‑shared across hobbyist aggregation sites and sparked tutorial and comment threads in creator communities. (9gag.com) ComfyUI added day‑0 integration for LTX‑2.3 and published reference workflows and nodes to run LTX‑2.3 pipelines inside ComfyUI. (blog.comfy.org) LTX‑2.3 itself ships as a 22B‑parameter video engine with distilled and quantized variants plus GGUF builds aimed at lowering VRAM requirements to the mid‑teens for faster local inference. (ltx.io) Community benchmarks and forum posts show the RTX 5090 yields high throughput in these workflows — an NVIDIA forum thread logged 1.65s per sample for certain AI‑Toolkit tasks on a 5090, and creator videos demonstrate LoRA training runs that finish in under two hours with tuned settings. (forums.developer.nvidia.com) Ostris’s own repository tracks performance discrepancies and user reports; a recent AI‑Toolkit issue documented roughly a 5× slower LoRA training time in AI‑Toolkit on Windows compared with an alternative community trainer, underscoring variability across setups. (github.com) Because LTX‑2.3 rebuilt its latent space, many older LoRAs are incompatible and creators are retraining adapters, while projects like RunComfy have released LTX2Pipeline nodes to keep ComfyUI inference aligned with AI‑Toolkit training previews. (crepal.ai)
