Swing shows compute-quality frontier for motion models: gains came from data selection, not scale
andrew_n_carr · x · 2026-10-13
Swing shared a motion-quality-vs-time-per-generation chart: more compute buys better motion, with Choreo sitting on the frontier at 25%, 50%, and 100% compute budgets, ahead of a commercial model and open models. Most gains came from data selection and post-training, which improved prompt adherence and motion quality together — Choreo 100% reaches 1270 Elo vs 1010 for Swing Sketch and 930 for open model A.
More from Multimodal
- Google showcases Nano Banana 2.1's standout character consistency in new examples — egeberkina · 2026-10-14
- AI prompt drops a portrait into Montmartre as both real photo and street caricature — Aiden_Tech_Ai · 2026-10-14
- SeedAudio 1.5 teased: AI-generated audio joins the video editing workflow in CapCut — FellMentKE · 2026-10-14
- Pixel art LoRA for Krea 2 and Anima lands on Civitai with solid results — aziib · 2026-10-14
- Alexandr Wang responds to Muse Wonder Full review, admits deployment kinks remain — alexandr_wang · 2026-10-14
- The Prompt Times: one tilt-shift prompt sparks a wave of isopod-rider remixes — charis_ai · 2026-10-14