The photo film: how it works
A photograph, traveled through for ten seconds, pulled back from for six, and hung on a wall for six. Twenty-two seconds. The photograph is never changed. The render costs nothing.
What the agent did
In September I asked an agent for a 3D walkthrough of a photograph. It had two of my standing instructions in front of it: do not touch the photograph, and find the least expensive path. It did not use a video model. It installed Blender and wrote the code in this kit.
The pieces, from zero
Blender. Blender is free, open-source 3D software. Animated films and game scenes are built in it. It can run from the command line with no window open, which is how the agent uses it.
A depth map. A depth map is a grayscale picture of your photograph where bright means near and dark means far. A small free model, Depth Anything V2 Small, estimates one from a single photograph in a few seconds on a laptop. It is a guess, and for a camera move of this size it is a good one.
The build. The photograph is laid over a surface shaped by the depth map. A camera moves through it along four points you choose. Over the pullback the surface flattens, a mat and a walnut frame appear around the whole photograph, and the camera settles in a room. The room is a generated still. The photograph is your file, pixel for pixel; the script checks its hash.
The render. Blender draws each of the 1,320 frames. On an Apple silicon Mac it takes about five minutes at 1080 x 1920, longer for 4K. FFmpeg joins the frames into an mp4, with music if you give it a track.
What you do
You don't run this yourself. Put the kit folder and one picture in front of a coding agent and direct the camera through the scene, the way you would direct a camera operator: where to start, how to move, what to pass, where to end. For the film above, the whole prompt was:
Make the film from this photograph. Start close and low on the cabins, as if I walked up the path. Move forward and lift through the trees. Keep rising until the sky takes most of the frame. Then pull back until the whole photograph is in front of me.
The agent turns that into four camera positions, builds the scene, renders and checks. The camera path is the one creative decision. Everything else is the recipe. Any still image works: a photograph, a scan, a painting, a generated picture.
What it needs
- A coding agent that can run commands on your computer. Claude Code and Codex are the two I have used. Either one installs the rest for you.
- Blender 5, free.
- Python 3.12 or later and FFmpeg.
- The Depth Anything V2 Small weights, free; the link is in the kit.
- A Mac with Apple silicon or a machine with a GPU. A CPU works, slower.
Limits
- One photograph per film.
- One room, mine. Use your own: any still of an empty wall at the right scale.
- Landscape photographs in this recipe, shown whole on the wall.
- A sequence of several photographs is not solved.
- A wide photograph at 4K is untested.
What it is for
I made it so a photograph can go on a vertical feed whole, never cropped, and so a client can see the print on a wall before it exists. Since then it has turned out to fit real estate, tours, and transitions in generated video.
Take it
Free. Take it, use it, change it. The kit's scripts and instructions are under the MIT License, which allows any use, commercial included, as long as the copyright line stays with the code. The depth model is its authors' and comes with its own Apache license, noted in the kit's README. Directed and tested by Noah Gallagher. Written by a coding agent under his instructions. First published October 11, 2026 at ntgconsultingllc.com/photo-film.