Training LTX-2.5 Video Control LoRA Locally: 47 Mins on a 48GB Card
ashishsanu · reddit · 2026-08-13
A developer shared benchmarks and a full workflow for training IC-LoRA (In-Context LoRA) locally on the LTX-2.5 model. Unlike traditional Clip LoRA that learns a single look, IC-LoRA learns a transform by training on paired clips (e.g., original video and its edge map).
Benchmarks (L40S 46GB, 90 paired clips, 500 steps):
- IC-LoRA peaks at 42GB VRAM, taking 1.03s per step and 47 mins total.
- For comparison, Clip LoRA takes 0.67s per step and 19 mins total.
Key Findings & Tips:
- High Startup Cost: Encoding paired datasets (180 video clips + 91 captions) takes 38 minutes before actual training (9 mins) begins. It is advised to run long sessions to avoid re-paying the startup cost.
- Strict VRAM Floor: With no 4-bit path currently available, 48GB VRAM is the minimum requirement; 24GB cards cannot run it.
The author also open-sourced a node-based trainer tool, Inline-Studio, featuring automatic dataset pairing and smart captioning.
More from coding & agent
- HolaOS Tops GitHub Trending as Open-Source Desktop Agent with Model Flexibility — alifcoder · 2026-08-14
- Building 3D Action Games with Claude: 'Dark Souls' in 7 Hours — vrdrift · 2026-08-14
- DeepSeek V4 Flash Makes AI Agents Affordable for Everyone — Teknium · 2026-08-14
- Workflow for Building iOS Apps Entirely via Claude Mobile — EricBuess · 2026-08-14
- DeepSeek Ships V4 Pro, Open-Sources Agent Software, Raises API Prices — The Decoder · 2026-08-14
- Are Agent Harnesses the Boring Way to Continual Learning? — scaling01 · 2026-08-14