LTX-2.5 on RTX 5060 Ti

Worst case of a 15-run session measured by the site operator on his own card, 2026-08-13, zero OOM. This run peaked high for one avoidable reason: the prompt text changed, so the 14.3 GiB Gemma-4 text encoder re-entered VRAM on top of the resident transformer. Peak across all 15 runs was 14155-15912 MiB (13.82-15.54 GiB) of 16311 MiB, and it did NOT scale with frame size: the largest job of the session (1664x960, 193 frames, 294.06 s) produced the LOWEST peak, and 1920x1088 x97 frames with synchronised audio ran in 183.12 s at 14512 MiB. Two-pass schedule (8 steps + 3). ComfyUI b323a345 + ComfyUI-GGUF 6ea2651e, torch 2.12.0.dev+cu128, driver 591.86. Transformer realrebelai/LTX-2.5_GGUFs Q3_K_M (11,525,623,808 B), conv video VAE + audio VAE, VAEDecodeTiled 512, prompt enhancer off. Raw 5,288-sample nvidia-smi trace in this session folder. Unreplicated: one rig, one operator.

recorded from huggingface.co ↗ · verified 2026-09-14