AKUSPACE
An LTX-2.5 audio LoRA that uses prompts to place sound in an acoustic space while preserving performance timing for video. I trained it on paired dry and reverberant recordings, Eurorack modular patches for SFX, and my own environmental field recordings. Winner of the Audio LoRAs category at the 2026 LTX LoRA Jam, with an NVIDIA GeForce RTX 5090 prize.
Place the sound
Choose a space and amount to see the room change and inspect the conditioning prompt. The teaser above shows recorded model outputs.
Training
I built the dataset from my own voices, instruments and electronic music, pairing dry and treated performances using digital reverbs, Eurorack modular patches for SFX and my own environmental field recordings. The dataset contains 266 paired clips from 364 distinct audio files across 19 treatment settings, with 228 pairs used for training and 38 for validation.
I trained the audio branches with the official LTX Trainer, leaving the video branch unchanged. I kept the reverb tails intact at clip boundaries so the adapter could carry the acoustic space through longer outputs.
Processing amount is built into the training data. Reverb and modular effects were rendered at three strengths, written into the captions as gentle, moderate and heavy; outdoor ambience uses gentle and heavy. The ComfyUI controls translate level selections into these trained prompt words. With the LTX-2.5 dev model, CFG scale adds control over intensity, detail and loudness during generation.
Stack
LTX-2.5 • Official LTX Trainer • PyTorch • ComfyUI • JavaScript • Three.js