AI City Challenge 2026 — Track 5, Generative Traffic Video Forecasting

2nd place · 76.0385 on the full test set
Team SSUPER · Team lead · ECCV 2026 AI City Challenge Workshop · Jul 2026

Pipeline — observed history and behavior descriptions into a LoRA-adapted Cosmos3-Super, producing N future frames

▲ The submitted pipeline. The 64B Cosmos3-Super base stays frozen; only a rank-16 LoRA is trained, on 15,673 phase-centered clips built from WTS and BDD_PC_5K.

Task

From K observed frames plus paired pedestrian- and vehicle-behavior descriptions, render exactly N future 1280×720 frames — dashcam and fixed-overhead cameras, history 10–224 frames, horizon 51–120.

Approach

The public 64B Cosmos3-Super backbone stays frozen.
Rank-16 LoRA on the generation-path attention projections — 39.85M trainable params, 0.062% of the backbone.
15,673 phase-level clips from the official WTS and BDD_PC_5K only.

Generated future frames for a fixed-overhead and a dashcam scene, conditioned on pedestrian and vehicle descriptions

▲ Left column is the last observed frame; the rest are generated under the paired descriptions. One sample per case, eight denoising steps.

Result

76.0385 full test / 74.0750 public subset — 2nd place.
Oral at the ECCV 2026 AI City Challenge Workshop (Malmö, Sept 2026).

Links

Official challenge site — 10th AI City Challenge, ECCV 2026 · Code