How AI Video Generation Works — And Why Most Apps Get It Wrong

Video generation is the hardest problem in AI companion apps. Most platforms either do not offer it at all, or offer it with high failure rates and inconsistent character identity.

The core challenge is character consistency. Generating a still image of a character is relatively straightforward — you describe what you want and the model produces it. Video is different. You need the character to look the same in motion as they do in the still image. Cheap approaches fail here constantly.

VelourCrush uses WAN 2.7 — one of the most advanced image-to-video models available. We use an anchor-based approach: we take the character's portrait image and use it as the anchor frame for video generation. This means the character in the video looks like the character in the portrait. Identity is preserved.

Our pipeline queues video generation jobs, polls for completion, uploads the resulting MP4 to cloud storage, and delivers it directly in chat. The whole process takes 30-90 seconds.

The result is a companion who not only looks consistent across photos but also moves in video — something no other platform delivers reliably.