LTX 2.5 AI rap music video: strong character consistency, mushy motion — full workflow and settings shared

Inevitable-Bus-8654 · reddit · 2026-11-08

The author generated an AI rap music video with LTX 2.5 and shared the full workflow seeking honest feedback. Character, outfit and location consistency held up surprisingly well, but the output still feels like AI slop: mushy motion, facial detail loss during movement, smearing hands, and inconsistent lip sync.

The workflow:

Six reference images made with ChatGPT Images to keep the character consistent
LTX 2.5 22B dev (BF16) on a Runpod B200 via the official two-stage audio-to-video pipeline, with the full song mix as audio conditioning
20 independent takes of 7–8 seconds each; stage 1 at 960×544 (24 steps, CFG 3, STG 0.5), spatial upscale to 1920×1088; stage 2 with the official distilled LoRA (strength 1), reference image strength 0.8, 24 fps
Tried distilled generation, native 48 fps, the alternative convolutional VAE, and cleaner references — none reliably fixed motion quality; double-exposure artifacts were mitigated by changing seeds and simplifying camera movement

Open questions: is this an LTX 2.5 limitation or a settings issue? Would shorter shots or different guidance preserve facial detail? For a believable performance, audio-to-video directly or reference performance + character replacement + separate lip sync? Is MiniMax-H3 worth investigating?

Original post →

LTX 2.5 AI rap music video: strong character consistency, mushy motion — full workflow and settings shared

More from Multimodal