CIVITAI / Workflows

EverAnimate Long-Video Consistency Generation Workflow

Watch the full video first if you want to understand how this EverAnimate long-video consistency workflow works in practice. The video shows how a reference character can be animated through a driving video, how face and pose information are extracted, and how the workflow extends the result into a longer continuous video while trying to keep identity, motion, and visual style stable. This ComfyUI workflow is designed for EverAnimate long-video consistency generation. Its main purpose is to solve a common problem in AI character animation: the first few seconds may look good, but as the video becomes longer, the face starts drifting, clothing changes, body proportions become unstable, and the motion loses continuity. This workflow uses a segmented generation structure with motion handoff, pose guidance, face reference, and loop-based continuation to make longer character videos more controllable. The workflow starts from a driving video. VHS video loading and video information nodes read the selected frames, FPS, width, height, and frame count. This gives the workflow a clear source timeline before generation begins. The driving video is then processed through pose and face detection. The graph includes ViTPose / YOLO-style body detection, PoseAndFaceDetection, DrawViTPose, SDPose keypoint extraction, and face image extraction. These preprocessing steps turn the original video into usable pose_video and face_video conditions. The EverAnimate generation section is the core of the workflow. ComfyEverAnimate receives the positive prompt, negative prompt, VAE, reference image, face video, pose video, width, height, length, pose strength, face strength, and motion handoff settings. The first EverAnimate pass generates the opening segment. After that, the workflow trims anchor latents and duplicate image frames, then uses continue_motion to pass motion information into the next segment. This is the key mechanism for long-video continuity. Instead of generating the entire long video in one pass, the workflow uses a ForLoop structure. The first segment establishes the character, motion, and visual identity. The loop then repeatedly generates continuation segments while receiving the previous motion context. Each continuation segment is sampled, decoded, trimmed, and batched back into the full sequence. This makes the workflow more practical for longer AI charact