Consistent videos: the same character and the same product in every shot

What you'll see in this lesson

1. Why the character changes from shot to shot

   Every generation starts from random noise: if you don't give the model visual anchors, it reinvents the face, the clothes and the product.

2. The character sheet and the product sheet

   Before generating any shot, prepare the character from several angles and the product from every side, with the same lighting.

3. Reference images: how many each model accepts

   The fastest way to lock in a character or product is to pass its images to the video model in every shot.

4. Keyframes: first and last frame

   For maximum control, create the first and last image of each shot yourself and let the model fill in the motion between them.

5. Character LoRA: for recurring characters

   A LoRA is a small add-on trained on your character or product so the model learns it; it pays off for series and mascots.

6. Lighting, wardrobe and a fixed prompt block

   Copy the same description block into every shot and change only the action and camera: that way neither lighting nor clothes vary.

7. The pipeline: Qwen makes the frames, MiniMax the video

   An agent like Hermes or Claude Code can chain via API the frames with Qwen-Image and the shots with MiniMax H3.

8. The instructions for the agent

   Give the agent a closed brief: inputs, steps, a review checkpoint and a spending cap.

9. Summary

   Consistency isn't requested with words: it's built with images, chained frames and a prompt that doesn't change.