From Gimbal Move to Hero Shot

This is the foundation everything else depends on. The operator's physical motion, the crouch, the rise, the camera's arc and tilt, was all real, timed together on location. If this part isn't good, no amount of AI afterward fixes it.

The AI can repaint the world around a performance. It cannot fix a bad take. Start with something worth keeping.

Cut the footage down to the exact take worth working with before opening Higgsfield. A clean, short source clip makes every step after it easier.

Seedance 2.5's video input caps at 15 seconds total. Trimming isn't optional prep. It's a hard constraint. If you hand it a raw 45-second take, you're already working against yourself.

Instead of describing the clip from memory, open it in Claude and let it pull frames directly. The pose at the crouch. The framing mid-rise. The exact point the camera settles.

That frame-by-frame read is what makes the next step accurate instead of guessed. Describing your own footage from memory introduces errors. Claude reading it directly removes them.

This is the core of the whole workflow. The instruction to Claude was simple in shape and specific in detail: change the environment, change the wardrobe, lock everything else.

The skill's whole premise is that a transformation prompt has two jobs pulling against each other. Lock down what makes the shot recognizable (the performance, the framing, the camera's exact path) and change only the one or two things you actually want different. Under-specify the lock and the model quietly re-rolls the camera or the face along with the environment.

In practice that meant getting specific: a destroyed city block, rubble, a flipped taxi, a tactical suit swapped in for his real clothing, while the crouch-to-rise motion and the camera's arc stayed identical to the source.

Once the direction was locked in, Claude wrote the full generation prompt. The source description, the lock-down clause at the start and end, the specific physical details of the new environment and wardrobe, and an SFX line describing the ambient sound.

The skill handles the structure. You just describe what you want changed and what needs to stay. That separation is what makes the output consistent instead of random.

The prompt goes into Higgsfield running Seedance 2.5, set to video-to-video against the trimmed source clip.

Two settings matter more than they look like they should: aspect ratio and duration need to be set explicitly as parameters in the platform, not just mentioned in the prompt text. If you leave them at default, the platform override wins and you end up fighting yourself on both.

Honest version of this step: the first few passes drifted. Heavier environment rewrites pulled the model's attention away from holding the camera steady, even with lock language already in the prompt.

The fix wasn't a new tool. It was going back into the prompt and describing the camera's actual physical path (starting distance, height, pull-back, tilt angle) instead of just saying 'don't change the camera.' Specificity beats repetition here. Every time.

Same real gimbal move. Same real performance. Now standing in a completely different world, in a completely different outfit, without losing a frame of what was actually shot.

That's the whole point of starting with real footage. The AI has something worth preserving. Your job as a director and operator doesn't disappear. It becomes the foundation everything else is built on.