xiaokunfeng/miga

MIGA is a research project for generating very long AI-produced videos without retraining the underlying video model. Standard AI video models are trained to produce a fixed number of frames, generating longer videos by chaining segments tends to produce visible inconsistencies where style or motion drifts between sections. MIGA adds two techniques on top of existing pretrained models to address this. The first technique is called Two-Stage Training-Inference Alignment. During inference, noise levels are arranged in a zigzag pattern across the frame sequence in stage one, then unified in stage two. The goal is to bring the noise patterns seen at inference time closer to what the model encountered during training, reducing the quality degradation that accumulates across long sequences. The second technique is Dual Consistency Enhancement, which has two parts. Self-Reflection evaluates early high-noise frames by comparing them to later low-noise frames in the latent representation, and corrects frames that have drifted too far from the established visual style. Long-Range Frame Guidance uses distant low-noise frames from earlier in the generated sequence to steer the denoising of new frames, helping maintain consistent appearance and motion throughout. MIGA has been implemented on two existing video generation models: Wan2.1 and VideoCrafter2. The repository provides separate environment setup instructions, model download locations, and example generation commands for each. Configuration is managed through YAML files. Key parameters include the number of denoising steps, the zigzag window width, the number of frame chunks to generate, and how many long-range guidance frames to use. Generated video length scales with the chunk count and zigzag width settings. The paper describing this method was accepted at ICML 2026. Memory consumption remains constant regardless of the generated video length.