How I Used AI VFX on Real Footage to Make a Superhero Warzone Shot

I wanted an Avengers-style scene: camera orbiting around a superhero who just landed in a destroyed city. So before I picked up the camera, I thought through what that would actually look like as a shot. Wide orbit, low angle, the subject at center. I went to an open park specifically because I needed the space to pull off a wide orbit without running into anything.

Think about the scene you want first, then shoot the camera move that serves it. If the final scene needs an orbit, go shoot an orbit. If it needs a low push-in, shoot that. The AI preserves exactly what you do with the camera, so shoot it the way you want it to look.

You can also just use existing footage you already have. The workflow works with anything.

Before opening any AI tool, cut down to the exact take you want to work with. My main clip was around 13 seconds. I also did a version with a 4-second clip, which gives a tighter, punchier result. Both worked well.

Seedance 2.5 takes video inputs up to 15 seconds, so keep it short. The shorter and cleaner the clip, the easier everything after this gets.

Before writing a single word of the prompt, send Claude the actual video. Let it watch the clip and pull frames on its own. It reads the pose, the framing, where the camera settles, all of it. That context is what lets it write the prompt like a director who actually watched the footage, not someone guessing from a description.

This step is what makes Step 4 work. Don't skip it.

Once Claude has seen the footage, type /seedance-footage-vfx in the message box to activate the skill. Then just describe what you want in plain language. Here's exactly what I wrote:

"Write a prompt to replace the environment with a broken down city with rubble and an upside down yellow taxi behind the man. Also make the man into a superhero wearing superhero gear. He is crouched down and slowly stands up as he witnesses the city's destruction."

Claude takes that, combines it with the frame analysis from the previous step, and writes out the full structured generation prompt. It locks down the camera move and the performance, and only changes what you asked for. Read the full prompt before generating. If anything looks off, tell Claude to fix it right there in the same thread.

Take the prompt into Higgsfield and select Seedance 2.5 Edit mode. Not the create-from-text mode. Upload your source clip, set the duration to match the length of your raw footage, drop in the prompt, and generate.

If the first pass doesn't look right, share it with Claude in the same thread and tell it exactly what's off. Camera drifting? Suit doesn't look right? Just say so and Claude will adjust the prompt. This is the trial and error part where you start learning what prompt details actually move the output.

Take your time and actually read through the full prompt. There's usually something in there you'd want to tweak before running it again. The more passes you do, the better you get at writing these.

Same real move. Same real person. Completely different world.

Mixing AI with real footage is genuinely one of the most fun things I've done as a filmmaker. The ability to shoot something simple and transform it into a big-budget-looking scene is incredible, especially for anyone working with a lean budget. Want a superhero landing in a warzone? Go shoot an orbit in a park. That's where we are right now.

Honestly, I'm also a bit nervous about where all of this is going. I do hope there will be real conversation around regulation and how AI-generated media gets used. But in the meantime, we're here, having a great time, learning the tools, and figuring out what's actually possible. This is just the beginning.