Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)

One Author Made an AI Comedy, 'Evacuator', in 2 Weeks and for 150,000 Rubles in Tokens

Director Grigory Ivanets made the comedy "Evacuator" in 2 weeks for 150,000 rubles in tokens (the term for paying for image and video generations). In Albuquerque, a tow-truck driver hauls away a 1963 "Impala," which turns out to be an intergalactic robot on the run. The frames were made in Seedream, the video in Seedance 2.5, and the author developed the script with Claude. His practical advice: keep locations in one tool, record voices in advance, and show the rough cut to a cinematographer.

AI-processed from TG: @cgevent (Метаверсище и ИИще); edited by Hamidun News
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceTG: @cgevent (Метаверсище и ИИще) ↗
Editor-in-Chief
Key metric #1Locations were built in Higgsfield Soul Location and refined in Seedream
Key metric #2Lines were recorded in advance and fed in as an audio reference for every take
Key metric #3After showing the rough cut to the director of photography, some shots were regenerated

One person, two weeks, and 150,000 rubles in tokens: that is what it took to make the short comedy "Evacuator" entirely with neural networks. Tokens here means the payment for generations: every image and every video clip in Seedream (an image generator) and Seedance 2.5 (a video model) draws down part of the balance. So the 150,000 rubles covers only the generation bill, not the author's own time. Director Grigory Ivanets showed the film and described the tools and techniques that helped him stay within those numbers. It is a useful practical breakdown for anyone making video with neural networks or only planning such a project.

What the film is about and how it came together

The plot is simple and funny. In Albuquerque, New Mexico, tow-truck driver Billy Chavez tows a 1963 "Impala" to the impound lot. It turns out the car is an intergalactic robot on the run. And that is still Billy's problem.

The author started the film for the Higgsfield contest, just after Seedance 2.5 came out. The work did not make the shortlist, but Ivanets calls it the best of his AI films. He first wanted to tell a story about the Russian provinces, but friends advised him to make it understandable outside Russia too. The setting moved to Albuquerque, where "Breaking Bad" was filmed. The idea is the author's own, and he developed the script together with Claude: gags and jokes were added along the way, and the ending gradually got stronger.

The workflow: Seedream, Seedance, and Higgsfield

The pipeline had two stages. The author generated still frames in Seedream and animated them in Seedance 2.5. He built a couple of locations in Soul Location inside Higgsfield and then refined them in Seedream. The reason was practical: it made it easier to keep a recognizable place consistent from frame to frame. According to the author, other tools he tried more often changed details of the locations between takes.

Compared with conventional filming, this is an attempt to build the "set" in advance: first pin down what a street or an impound lot looks like, then shoot different scenes in it. For AI video, this is one of the biggest headaches. The hero stays the same, but the sign behind his back suddenly changes.

The voices deserve a separate mention. Ivanets recorded the lines of all the main characters in advance and fed them in as an audio reference for every take with dialogue. The film is entirely in English, and for it the author used the English lip-sync of Seedance 2.5, that is, matching mouth movements to speech. The approach is straightforward: rather than hoping the model guesses the intonation on its own, he sets it with a finished recording.

A living person in the process

The author calls the most useful stage his conversation with a friend, a director of photography. Once the film could be watched in full, they went through together how the camera decisions serve the story. After that, Ivanets regenerated several shots and even whole fragments. The takeaway here is not about the technology but about the process: a neural network quickly produces a lot of material, but judging whether it adds up to cinema is better done with a person who has experience with framing and light.

Timeline, budget, and caveats

Production took two weeks, and the token budget came to 150,000 rubles. The author does not specify how much time and effort he spent beyond that, or how many failed attempts there were. So the figure cannot be compared directly with the budget of a conventional short film.

It is also worth keeping the limits of the source in mind. All the information comes from the author's own account, with no independent verification. The film did not make the contest shortlist, although the author notes a positive reaction from Albuquerque residents on YouTube. Whether the result can be reproduced on other stories and with other tools does not follow from the account.

What to take away

For those trying to make AI video, this case yields three simple rules. First: if a recognizable location matters, build it in a single tool and finish it there rather than switching generators between shots. Second: record the lines in advance and give them to the model as a reference, rather than relying on a random voice. Third: show the rough cut to an outside person, preferably a director of photography or a director, and be ready to regenerate entire scenes. Two weeks and 150,000 rubles in tokens for a complete comedy with dialogue is a benchmark that shows full-length AI filmmaking is now within reach of a small team, and ideally even of a single person.

Story

Key Questions Answered

What the film is about and how it came together?

The plot is simple and funny. In Albuquerque, New Mexico, tow-truck driver Billy Chavez tows a 1963 "Impala" to the impound lot. It turns out the car is an intergalactic robot on the run. And that is still Billy's problem. The author started the film for the Higgsfield contest, just after Seedance 2.5 came out. The work did not make the shortlist, but Ivanets calls it the best of his AI films.

What are the key details on The workflow: Seedream, Seedance, and Higgsfield?

The pipeline had two stages. The author generated still frames in Seedream and animated them in Seedance 2.5. He built a couple of locations in Soul Location inside Higgsfield and then refined them in Seedream. The reason was practical: it made it easier to keep a recognizable place consistent from frame to frame. According to the author, other tools he tried more often changed details of the locations between takes.

What are the key details on A living person in the process?

The author calls the most useful stage his conversation with a friend, a director of photography. Once the film could be watched in full, they went through together how the camera decisions serve the story. After that, Ivanets regenerated several shots and even whole fragments. The takeaway here is not about the technology but about the process: a neural network quickly produces a lot of material, but judging whether it adds up to cinema is better done with a person who has experience with framing and light.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…