CIVITAI / Workflows

SCAIL-2 Two-Person Biological Long-Video Driving Workflow

Watch the full video first if you want to understand how this SCAIL-2 two-person long-video driving workflow works in practice. The video shows how two characters in a reference image can follow a two-person driving video, while the workflow keeps left and right character assignments, pose structure, temporal continuity, and long-video extension more stable. This ComfyUI workflow is designed for SCAIL-2 two-person biological long-video driving. Its main purpose is to transfer global two-person motion from a driving video onto two characters in a reference image. This is not a local replacement workflow. It is a skeleton-guided animation workflow where both characters follow the full-body movement from the driving video while keeping their assigned identities and visual roles. The workflow is built around wan2.1_14B_SCAIL_2_fp8_scaled.safetensors as the main SCAIL-2 model. It also uses WAN VAE, UMT5 XXL WAN text encoding, CLIP Vision identity encoding, SAM3 subject tracking, SCAIL2ColoredMask, WanSCAILToVideo, SamplerCustom, VAEDecode, ForLoop continuation, frame trimming, color matching, video combining, and original audio restoration. The LoRA chain is also preserved for enhancement, including LightX2V, WanAnimate relight, Wan2.2 Lightning I2V, FastWan 480p, Wan21 PusaV1, Wan2.2 Fun InP, and stage-based enhancement LoRAs. The first important rule of the workflow is size alignment. Both the reference image and the driving video are aligned to 512×896 before entering SAM3, CLIPVision, and SCAIL. This avoids mismatched masks, unstable pose conditioning, and identity drift caused by inconsistent input dimensions. The second key rule is two-person tracking. SAM3 is configured with max_objects=2. SCAIL2ColoredMask uses object_indices=0,1 and sort_by=left_to_right. This means the workflow treats the left and right subjects as separate controlled identities. The goal is to reduce identity mixing, role swapping, clothing confusion, and left-right character instability during two-person motion transfer. The workflow uses replacement_mode=false, which means it focuses on two-person skeleton guidance rather than local character replacement. The reference image provides the two target characters, the driving video provides the motion, and the colored mask system links both sides together before entering WanSCAILToVideo. The long-video structure is one of the stron