Generating Argentine tango with H3: key-pose reference images finally worked
Christian4243 · reddit · 2026-10-15
A user challenged H3 to generate two dancers performing authentic Argentine tango figures (ochos, giros, ganchos). Plain prompting, motion transfer, depth maps, and storyboards all fell short.
What finally worked, partially: one or two reference images of key poses, prompting the model to generate the motion between them. The full structured prompt is shared — subject definitions plus second-by-second action descriptions, e.g. constraining a gancho with prep-pose and peak-pose images, plus strict rules preventing lip-sync. Still imperfect; the author invites others' techniques for dance generation.
More from Multimodal
- MiniMax-H3 hailed as favorite open release of the year, runs fully local — AdinaYakup · 2026-10-23
- Michal Malewicz experiments with depth maps in his creative workflow — michalmalewicz · 2026-10-23
- One-Shot Tunable Style Transfer on Flux3 Via Reference Latent + Node — Winter_unmuted · 2026-10-23
- PhysVidBench accepted at EMNLP 2026: do video gen models grasp physical commonsense? — aykuterdemml · 2026-10-23
- EMNLP 2026 paper teaches multimodal LLMs to get cartoons via incongruity-resolution supervision — aykuterdemml · 2026-10-23
- A gorgeous Midjourney style: sref code 820221253 for dreamy lavender watercolor looks — umesh_ai · 2026-10-23