Make it on your own machine

• You want hundreds of tries. On your own card each try costs only time and electricity.
• The material is private: family photos, client footage, your own voice. It never leaves your computer.
• You want to see how the pieces fit together, which also makes paid tools easier to judge.

A cloud tool is still the better call when you need one polished result today, or when you do not have a PC with a recent NVIDIA graphics card, which is what our setup runs on.

Our box has one NVIDIA GeForce RTX 5070 Ti with 16 GB of memory. On it we run open models for still images (Flux), image to video (Wan2.2 and LTX-2) and music (ACE-Step). All of them run through Wan2GP, a free, open-source app built to run these models on modest graphics cards.

The limit is the card's memory. Each model needs its own slice of it, so our runner waits for enough free memory before it starts a job: about 7 GB for a Flux still, 6.5 GB for a Wan2.2 clip, 10 GB for an LTX-2 clip and 5 GB for a music cue. The largest video model we use has 22 billion parameters. It fits on a 16 GB card because we run a version compressed to about 4 bits per weight.

1. Write the bible first. Characters, wardrobe, places and the look live in one folder, so every prompt describes the same person the same way.
2. Break the story into shots. Each shot is one line in a queue file: the shot number, the model, the prompt, a fixed seed, the size and the length.
3. Make the still before the motion.Flux draws many candidate frames for a shot, and the best one becomes that shot's reference frame. On October 23, about 3,300 of the 4,000 jobs in our queue were stills.
4. Then add motion. The chosen still goes to two video models, Wan2.2 and LTX-2, and the better take wins.
5. Let it run while you sleep. A short script on the box takes the next shot that fits in free memory, renders it, and tries a failed shot up to three times before setting it aside.
6. Choose and cut by hand. Watch the takes, keep the ones that match the bible, and edit them together yourself.

Close-up, static camera, single continuous take: Cal, a young man with short wet dark hair and light stubble in a charcoal-and-rust BASE suit with black harness straps, looking off-frame. He slowly lowers his chin and eyes toward the ground in one small, deliberate nod of agreement, then lifts his head back up and holds still, looking off-frame. Mouth closed, face and features stable and unchanged, hands never enter frame. Wind moves his hair, mist drifts behind him. Exactly one person. Setting: a fictional forested Irish/Scottish-style valley below a sheer slate-grey granite wall, wet moss greens, overcast soft light, drifting mist, photoreal cinematic film still, muted palette. No cuts, no text.

The prompt asks for one camera position, one continuous take and one small action. Each clip is generated on its own, so the prompt says who Cal is every time, and it names what must stay fixed: a closed mouth, steady features, no hands, exactly one person and no cuts.

• Start with one shot that lasts a few seconds.
• Keep your prompts in a text file so you can run them again.
• Fix the seed and change one thing at a time.
• Make many stills before any video, because a still renders much faster than a clip.

Then bring what you made to a Studio gathering and show it, or go back to the guide for another place to start.