Plan a shot in Blender before generating the AI video

Start with an action you can inspect

Thanks to Omnisend for sponsoring this video. The original sponsor resource is [Omnisend](https://your.omnisend.com/Luke).

The September demonstration uses Claude Opus 5.5 and Blender MCP to stage a simple fight rehearsal, then sends the reference to Seedance 2.5 through Kie.ai. The purpose is to plan the shot before paying for repeated generations that change its action. The description says the reference follows Luke’s intended movement more closely, while still looking robotic in places. It does not report a measured saving in credits.

Keep the scene and generation routes distinct

The modelling and animation happen in Blender; the cinematic transformation happens in the video-generation service. A clear reference gives the generation a specific sequence to follow, but is not itself a realistic finished shot.

Use the retained [Blender download](https://www.blender.org/download/), [Claude download](https://claude.com/download), [Blender MCP repository](https://github.com/ahujasid/mcp-for-blender) and [Seedance 2.5 destination](https://kie.ai/seedance-2-5) for the demonstrated route. These links do not establish that the same model, setup or terms remain available today.

1. Follow the MCP connection section at 0:48 before asking Claude to build the scene.
2. Follow the Kie.ai setup section at 7:00 to connect the generation route used in the demonstration.
3. Keep each Blender revision as a new version and inspect its preview before exporting the approved reference.

Send these six prompts one at a time

The following prompts are the exact condensed, copyable sequence published with the video. Preserve the earlier results so the final comparison remains understandable.

1. Build: Use Blender MCP to create a new seven-second project at 24 fps. Add a grey floor and two articulated mannequins, red and blue, facing each other two metres apart. Give them sensible joint pivots, basic lighting and a three-quarter camera showing both full bodies. Keep them still. Save a new project without overwriting existing work and show me the camera view.
2. Animate: Over five seconds within this scene, red steps forward, stops, plants both feet and throws one straight punch. Blue ducks underneath while keeping their gaze towards red. Briefly hold the punch and duck, then recover. Keep feet grounded, bodies separate and the camera fixed. Save a new version and preview it.
3. Camera: Keep the action and timing unchanged. Lower the camera to a clear three-quarter angle and add a slow, small arc towards the action. Keep both full bodies visible and the framing steady. Save a new version and preview before rendering the approved sequence as an MP4 reference.
4. Generate: Use the reference video to guide the camera, framing, positions and timing. Turn the scene into a cinematic live-action stunt rehearsal in an empty concrete warehouse. Replace red and blue with adult performers in matching jackets. Preserve the step, planted feet, punch, duck and recovery. Keep full bodies visible, consistent identities and soft lighting. No contact, injury, extra attacks, cuts, text or logos. Use Seedance 2.5 through Kie.ai and match the exported reference duration.
5. Compare: Generate another version without the input video. Describe the same action, setting, camera angle and movement in the text prompt. Show the Blender reference, guided result and text-only result side by side.
6. Improve: Inspect the reference for punch direction, foot placement and camera movement. Fix the unclear punch and camera arc in Blender, export the revised reference and rerun only the reference-guided generation. Reuse the original text-only result in the final comparison.

Compare motion and realism separately

The final revision fixes the unclear punch and camera arc in Blender, then reruns only the guided generation. The original text-only result stays in the comparison. This makes the sequence easier to follow: it is a revised guided result against the original text-only attempt, not a claim that both routes received identical follow-up work.

Look at whether the step, planted feet, punch, duck and recovery remain legible, and separately whether the performers move naturally. A closer match to the intended action can still look robotic. Treat the clip as a creative experiment with observed limits, rather than proof that a reference always saves money or makes every generation usable.

The takeaway

Plan and inspect the reference first, then judge action fidelity and realism independently. Keep the original comparison output visible when revising one route.