How to make an AI film

Build the location

Start with the place, not the shot. Generate the environment in Seedream 5.0 Pro or GPT Image 2 and describe it like a location scout would: architecture, time of day, weather, light. Leave the camera out for now.

Make several angles while you're there. You'll need them later, and doing them in one go keeps them consistent.

• Image. Pick the image tool from the sidebar.
• Model. Click the model row to change what you're generating with.
• Seedream 5.0 Pro. What I use for locations and character stills. GPT Image 2 is the other one worth a look.
• Prompt. Write the location here. References are optional at this stage, because you're creating the look rather than matching one.
• Output. Set your aspect ratio and resolution. I work 21:9 at 2K for anything that ends up on screen.
• Generate. Run a batch rather than one at a time. You want options.

Here's the shape of the prompt behind the location below. Label the parts, so you can change one without disturbing the rest.

That's the plate the rest of the scene gets built on. Notice how much empty deck sits in the middle. That space is deliberate.

There's a second route if you want to move a camera around inside a location instead of generating flat stills.

• Create World. Builds an explorable 3D environment instead of a single image.
• Create World tool. From text, or from an image you've already generated.
• Cast in Scene. Drops a locked character into a world you've already built.

• Create from image. Feed it the still you already approved so the world matches your film.
• Describe your world. Start from text if you've got nothing yet.

Design the characters

Same tool, same models, same approach. Describe build, age, costume, hair, scars, the lot. Be specific rather than poetic. Get one image you're genuinely happy with before going any further, because everything downstream inherits it.

Don't settle for a face you're only half sold on. You'll be looking at it for the whole film.

Design them against the world you've already built. This one's dressed to read against that neon.

Lock the character

Most people skip this step, and it's the one that decides whether your film holds together. From your hero image, generate a headshot at several angles and a full body turnaround, then save the lot as a character you can reuse.

• Create Character. Start a reusable character rather than another one-off image.
• Character tools. Create, image, video, talking video and motion sync all hang off the same saved character.

• Start from an image. The one I use. Feed it the hero image you approved in step two.
• Describe your character. Text only, if you're starting cold.
• Build your character. Guided options for look, age and so on.

• Front-facing image. Upload the approved hero still. Front on, clean, well lit.
• Name. Name them properly. You'll be picking them off a list for the rest of the production.
• Voice. Optional here, but attaching it now saves doing it shot by shot later.
• Create. Saved into your assets and available to every tool.

A character sheet prompt works the same way as a location, with the layout spelled out first so the model puts the panels where you want them.

That set is your character. Feed it back on every shot they appear in and they stop drifting between cuts.

Write the scene, then attach the assets

Now write what happens. Describe the action and the performance in plain language: what they do, how they move, what they feel. Then attach your location and character references so the model builds from what you've already approved instead of inventing something new.

Generate the shot

Take the approved frame into video. Work shot by shot, not scene by scene, and expect to run a prompt several times before one lands. The keeper rate is low. That's normal, and it's still faster than a crew.

• Video. Switch to the video tool.
• Mode and model. Text to video if you're starting cold. Image to video if you're animating an approved still, which is what you should be doing by now.
• The model list. Every model shows what it supports: reference images, start and end frames, native audio, resolution and clip length. Pick on the feature you actually need.
• Prompt and references. Your scene description, with the character and location attached.
• Audio and output. Native audio on or off, plus aspect ratio and resolution.
• Generate. Note the credit cost before you commit. Video is where the credits go.

Voice and music

Both live in the audio tool, so you never leave the platform to score a scene. Voices run on ElevenLabs: pick one from the library, design one from a description, or clone your own.

• Audio. The audio tool in the sidebar.
• Clone your voice. From a recording, if you're narrating it yourself.
• Design a voice. Generated from a text description. This is what you want for a cast of characters.

Music sits in the same place, under Create Music.

• Describe your song. Write it like a brief to a composer: mood, instruments, texture. You can attach stills so the track is written against what's on screen.
• Genre. Quick starting points if you'd rather not write from scratch.
• Instrumental. Turn this on for score. You almost never want vocals under dialogue.
• Song control. BPM, intensity and key, all on auto unless you need to hit something specific.
• Generate. Cheap compared to video, so make several and pick.

Cut, grade, deliver

This part happens outside OpenArt. Bring your clips, voice and music into a proper editor and cut it there. CapCut if you want fast and simple, Premiere Pro or DaVinci Resolve if you want control. CapCut and Resolve both have free versions that are more than enough to finish a film.

The film is made in the edit, same as it always was. Pacing, sound design and a score will do more for a scene than another round of generations. Grade last, so everything sits in the same world.