26 shots in 7 hours on a RTX 4070: full LTX 2.5 music video workflow breakdown
ART-ficial-Ignorance · reddit · 2026-08-21
A Reddit user shared a complete workflow and real-world comparison for making an experimental indietronica music video with LTX 2.5 (run via Wan2GP):
- Why LTX 2.5 over MiniMax H3: on an RTX 4070, MiniMax H3 OOMs at 720p for clips this long, while LTX 2.5 handles 1080p and is much faster—26 clips, best-of-2, 8 min/clip, roughly 7 hours total rendering
- Core workflow: first-frame/last-frame chaining plus audio conditioning with an audio-reactive LoRA. Each clip is 4 bars (10.75s) and the final frame of each shot seeds the next, keeping visual identity, geometry, and camera logic coherent across the sequence
- Custom automation tool: slices the song into segments, aligns everything to the edit grid, organizes keyframes, and packages a queue.zip for Wan2GP, turning 26 shots into a batch job—human work reduces to keyframe design, prompting, and picking results
- Pitfall: LTX doesn't support arbitrary frame counts, so exact 4-bar durations are impossible; the author rounds up to supported frame counts and plays clips at 103% speed, but tiny BPM/offset errors compound across 26 clips, causing slight drift by the end
- First frames were made with GPT-image 2.0 (swappable); prompts shared via Pastebin
More from Multimodal
- fal launches Worlds for real-time creation and interaction with AI-generated worlds — Scobleizer · 2026-10-16
- Gummy worms as a physics test: Gemini 4 Argon vs. Claude Opus 5.5 elasticity sims — DynamicWebPaige · 2026-10-16
- 260M-parameter looped diffusion transformer beats 6x larger baselines with 5x less compute — burkov · 2026-10-16
- Seedance 2.5 demo: stretching one second of an explosion into a 30-second shot — LudovicCreator · 2026-10-16
- Octopus Short Film Made with Grok Imagine, Upscaled via Topaz — creatoroff · 2026-10-16
- Open-sourced 61-second computing history film generated entirely by Muse Spark — sophiamyang · 2026-10-16