WorldCast — anchor-conditioned camera-control video generation

Trained external modules (a source self-attention bank + anchor channel adapter + camera projection, ~2 GB) that attach to a frozen Wan2.1-I2V-14B-720P backbone, enabling anchor-viewpoint-conditioned video generation along explicit camera trajectories — orbit_360 (full orbit ring), orbit_<N>[_cw|_ccw] (horizontal arcs of any sweep), orbit_180_push / orbit_360_push (push-pull arcs), lemniscate (figure-8), orbit_tilt_20 (tilted orbit ring), spiral (SEVA 2-turn spiral), zoom_in / zoom_out (dolly), turn_left / turn_right (lateral trucks) — with Plücker camera-ray injection and first-frame/body-view anchor conditioning.

Files

File	What it is
checkpoints/worldcast_bank.safetensors	the trained modules (source_self_attn_bank.* + anchor_channel_adapter.*), bank-only — the base DiT weights are NOT inside
checkpoints/base_dit.safetensors	the frozen base the bank was trained against (DiT-compatible weights + camera_proj + I2V channel adapter) — the two checkpoints must come from the same release

Requirements

• Code: the companion inference repository (WorldCast) — inference only, no training code
• Base model: Wan-AI/Wan2.1-I2V-14B-720P (VAE / text-image encoders come from the base model tree)
• GPU: >= 48 GB VRAM recommended for 480x832, 81 frames

Quickstart

git clone <worldcast-repo-url> worldcast && cd worldcast
pip install -r requirements.txt

# place weights
mkdir -p checkpoints pretrained
cp worldcast_bank.safetensors checkpoints/
cp base_dit.safetensors checkpoints/
# download the Wan2.1 base into pretrained/Wan2.1-I2V-14B-720P/

python demo.py \
  --manifest demo_case_traj5.json \
  --model_root pretrained/Wan2.1-I2V-14B-720P \
  --ckpt checkpoints/worldcast_bank.safetensors \
  --init_dit_from checkpoints/base_dit.safetensors

Sampling defaults

50 steps / cfg 5.0 without a step-distillation LoRA; the bank is architecture-locked to the training line's flags (per-layer proj_out bank on DiT rope, 20 bank layers) — see the companion repository's README for the full flag matrix.

License

Apache-2.0 (the Wan2.1 base is Apache-2.0; these derived weights are released under the same terms).