Newcomer Shoots Cinematic Shootout in Seedance 2.5 in 4 Hours for $40

User @Maf1k3 has been doing AI video for only a couple of weeks. In that time he made a Guy Ritchie-style series, "Konchenye Kozyrki", starring himself and his friends. The first episodes were a joke: a set of images with almost no plot. By the finale he wanted proper cinema, so in episode five he shot a full-blown shootout. The breakdown is interesting for its process, not its result: the author ended up with a clear, inexpensive pipeline that others can repeat.

How the pipeline works

The scene is a showdown with Don in an underground parking garage. One misunderstood phrase, the hero grabs his gun, and then it all plays out in slow motion to Jim Croce's "Time in a Bottle". It had to feel epic and sad at the same time.

The author split the work between models like roles on a team:

• ChatGPT: co-writer of the plot and the characters' motivations. It also built a skill that writes Seedream prompts in the author's style.
• Claude: "technical director". It proposed the staging for each shot, and the author adjusted it to his own vision. Claude also wrote all the Seedance prompts.
• Seedream 5.0 Pro: reference frames.
• Seedance 2.5 via Hedra: the video itself and all sound effects.
• DaVinci Resolve: editing. Music: Jim Croce.

The author made the other episodes in MiniMax H3. But he wanted the finale to look good, so he moved the entire shootout to Seedance and paid for it separately.

What works in Seedance, and what doesn't

The main part of the author's notes is practical observations. Here are the ones that will help anyone making video from prompts:

• Cut into shots. The model gets confused by long action in a single take without cuts. What worked: separate shots with hard cuts, with a duration and transition type specified for each. Within a shot, the author laid out everything second by second: camera, movement, slow motion, what is out of focus.
• Always specify the camera. If you don't say where it stands and how it moves, the model copies the angle of the reference.
• Watch your references. The model reads extra still frames as the first frame. It is better to describe the characters' positions in space in text.
• Describe characters on separate lines. Give each one their own tag and role. According to the author, faces hold up better this way.
• Don't chase long generations. They came out imperfect, so the author cut the good pieces out of them.

Emotions came out well. In the snorricam shot, one of the characters had to show fear and confusion. According to the author, this worked almost on the first attempt. The model never botched the faces of real people, even at 480p. That said, the scene relies mostly on close-ups with almost no extreme wide shots, and faces usually fall apart first in those.

Money, time, and resolution

Generating and editing the scene took about four hours. It cost roughly 6–7 thousand Hedra credits, about $40.

Another finding concerns resolution. The author planned to make the finale in 720p and use 480p for drafts. But the drafts looked decent enough, and he didn't redo them. In the end, almost the whole scene stayed in 480p, and only a couple of shots were made in 720p. He compared versions 2.0 and 2.5 on the first scene and saw almost no difference, but in Hedra, 2.5 at 480p turned out to be even cheaper, so he did everything in it from then on.

What to make of this, and where the catch is

An important caveat: this is one author's personal experience, not a benchmark. He is a beginner and says openly that he can't imagine going back to other models now. There is no metric-based comparison with competitors. Phrases like "nothing to nitpick" and "almost everything on the first try" are subjective assessments. A 480p result is fine for a homemade series on Telegram, but it may not be enough for commercial production. The cost is tied to Hedra's pricing and may change.

The practical takeaway for developers and product managers is this. Video generation is starting to look like ordinary work with code: you have a shot-by-shot specification, an executor model, and a "technical director" model that writes that specification. A storyboard with timings, hard cuts instead of long takes, and character tags are techniques you can adopt in your own process right now.

Try it this way: start with a low-resolution draft, then decide whether 720p is worth the money. The author ultimately decided it wasn't, and saved.