Reverse engineering an AI horror short

Question: a 45 second vertical horror short turned up on Reddit with no credit: a grinning woman in a nightgown climbs out of a fifth floor window of a Soviet apartment block, runs across the courtyard, appears through the peephole, then climbs in through the living room window. Was it generated, with what, and how would someone make one just like it?

Short answer: it is an episode of a Russian AI anthology, "Архив Аномалии" (Archive of Anomaly) no. 18 part 5, by the two person studio XCVmind [1]. The pixels give the method away before the studio does: five image-to-video clips of 6 to 13 seconds, a shared character reference, native 24 fps sources slowed in a video editor, a grain overlay, two stock transitions and a title card. The studio's public posts confirm the shape of that workflow and name the models it rotates through: Hailuo, Kling, Midjourney Video, Sora 2 and Veo [2], [3]. The exact model behind this episode cannot be read from the file, because Reddit's transcode strips every provenance tag; the studio sells per-scene breakdowns with prompts on a paid Telegram tier [4]. The same look is reproducible today on the Gemini API with Nano Banana Pro for keyframes and Veo 3.1 Fast for motion, for under $15 in generation per finished minute.

This is a companion to the September survey of consistent characters and sets in AI film. That doc asked whether a series could be locked; this one takes one real short apart to see what a working studio actually does.

1. Generated, not shot. A formless white shape on the windowsill resolves into a face and a body over two seconds. Candles multiply and move within one continuous take. A door's lock hardware changes count as the camera approaches. None of that happens to a camcorder.
2. Image first, then motion. The studio says so in its own words: upload finished artwork, set a prompt, let the network animate the scene [2]. The blob-to-person opening is what image-to-video does when the still shows a curled figure and the prompt says she rises.
3. One character reference across five clips. Same face, hair and nightgown with the pink bow trim in every shot, including a fisheye close-up. Five independent prompts do not do that; a reference image fed to every generation does.
4. The editor did the rest. A 30 fps timeline, four of five clips slowed to roughly 80 %, a film grain overlay, an RGB scanline glitch and a whip-blur transition, vector title text, an ambience bed with no dialogue. All of it is a preset away in CapCut or DaVinci Resolve.
5. The model name is not in the file. Reddit re-encodes to x264 and drops C2PA, XMP and every creation tag. The clips carry no watermark. The frame cadence fits a 24 fps generator, which narrows the field but does not name one.

The file itself was a Reddit download: the RDT_ file name is Reddit's own naming for a saved video, and the stream carries Reddit's 360p transcode ladder (x264, constant quality 23, a keyframe every 60 frames, scene cut detection off, an 800 kbps buffer cap), muxed by ffmpeg 8. Every creation date is zero, there is no encoder name other than ffmpeg, no C2PA or JUMBF manifest, no XMP, and no H.264 user data. That is what a Reddit copy of anything looks like, so the absence proves nothing about the original export.

A title card search found nothing. The YouTube link, once supplied, gave the real title: "Архив Аномалии №018, часть 5" on the XCVmind channel, uploaded about a week before this doc [1]. Episode 18 is a five part arc about a mimic entity that copies people, which is why the woman grins, climbs out of a window and turns up at the door [4].

The tells, strongest first. Each was checked on frames extracted at full resolution, on per-clip temporal variance maps (the standard deviation of every pixel across a clip, which exposes anything that stays still, such as a watermark), and on a frame difference trace.

Frame cadence. Counting exactly repeated consecutive frames gives the source rate under the 30 fps container:

A native 24 fps output fits Veo, Sora 2, Runway Gen-4, Seedance and Wan 2.5. Kling emits 30 fps and Hailuo 25 fps, so those would need a different slowdown to land at the same numbers; they are less likely, not excluded.

Audio. A sound design bed with no dialogue; the low band shows no speech formants. The first 4.5 seconds are a quieter drone that rolls off at 15 kHz, and everything after is broadband ambience with transients (footsteps, a door) that rolls off at 17 kHz, the AAC limit of Reddit's 128 kbps re-encode. Two different sources, laid under each other in the edit.

XCVmind is an AI film studio co-founded by "misterXCV". It runs the anthology on YouTube and TikTok, with a Telegram channel for the films and a second one for announcements and tips [1], [2], [3], [4]. Read across those channels and the studio's profiles on iFilm, vc.ru, DTF and Pikabu [5], [6], [7], [8], the stated workflow is:

• Image-to-video is the core method. Finished artwork goes in with a prompt and the video model animates the scene [2].
• Several video models run in parallel and the best take wins per shot. Named over the channel's history: Hailuo (MiniMax), Kling including its Omni reference mode and its O1 video editor, Midjourney Video, Sora 2, Veo 3, Grok Imagine and Seedance. The studio's refrain is that no perfect network exists yet; Sora is "obedient but soulless", Midjourney is "an impressionist" [2], [3].
• Cost and effort. Up to dozens of hours and around 5,000 rubles per finished minute [3].
• The exact breakdowns are for sale. The announcements channel's paid tier, "Секретное место" (Secret Place), publishes a per-scene breakdown of each episode: prompts, reference images, which model made each shot, and process notes. It is the direct route to the recipe behind this episode [4].

One caveat on the channel's banner, which reads "XCVmind | AI Video | Sora 2": OpenAI shut the Sora app down in April 2026 and retired the Sora 2 API on 24 September 2026 [9]. Any Sora based recipe in the studio's older posts is now historical, and its current shots come from Kling, Veo, Hailuo or Seedance.

The studio's method, restated as steps, with the one stack that a single API key covers.

1. Beat sheet. Five or six beats of 5 to 12 seconds, one location idea per beat, told as found footage. The anthology frames each episode as a documentary lore entry, then shows one incident.
2. Character sheet. One image model with a fixed reference: a front view, three quarter, full body and a face close-up of the same person in the same costume. Those four images ride along with every later generation.
3. A start frame per shot. Same image model, same character reference, plus a style reference for the late Soviet apartment look: yellow brick, teal stairwell paint, patterned carpet, wooden window frames, candlelight. Generate at 9:16 so nothing is cropped later.
4. Animate each frame. Image-to-video, three to five takes per shot, keep the least broken. Prompt the camera outright: "handheld POV, slow push toward the door, fisheye through a peephole".
5. Sound. An ambience bed, a couple of transients, a low drone under the opener.
6. Edit. A 30 fps, 1080 by 1920 timeline. Slow the clips to 75 to 85 %, add film grain, a VHS glitch preset and a whip-blur for transitions, end on a static title card, export H.264.

Cost on the Gemini API for a 45 second short with two takes per shot, at the September 2026 list prices [10], [11]:

Veo 3.1 generates 4, 6 or 8 second clips at 24 fps in 16:9 or 9:16 with native audio, and takes a start image plus up to three reference images for character consistency [10], [11]. The budget model for stills, Gemini 2.5 Flash Image, is deprecated with a shutdown on 2 October 2026, so the Pro model is the one to build on [11]. Expect two to four hours per finished minute at first, most of it choosing takes.

• The model behind this episode. Nothing in the Reddit copy names it. The studio's own export, or its paid breakdown, would.
• Invisible watermarking. A SynthID check needs Google's verifier run on the original export, not on a transcoded copy.
• The full text of the studio's posts. The environment that did this work could reach search results but not Telegram, YouTube, TikTok, Instagram or the Russian blog platforms directly, so the workflow statements above come from indexed excerpts of those channels, not from reading them end to end.

1. [1]
   XCVmind, "Архив Аномалии №018, часть 5," YouTube, Sep. 2026. Accessed: Sep. 26, 2026. [Online]. Available: https://www.youtube.com/shorts/dAdaAVyWxNw
2. [2]
   misterXCV, "misterXCV ART," Telegram channel. Accessed: Sep. 26, 2026. [Online]. Available: https://t.me/s/misterxcv_art
3. [3]
   XCVmind, "XCVmind," TikTok profile. Accessed: Sep. 26, 2026. [Online]. Available: https://www.tiktok.com/@xcvmind
4. [4]
   XCVmind, "XCVmind | Анонсы | Советы," Telegram channel. Accessed: Sep. 26, 2026. [Online]. Available: https://t.me/s/vmind_ai1
5. [5]
   iFilm, "XCVmind - AI Filmmaker," iFilm creator page. Accessed: Sep. 26, 2026. [Online]. Available: https://www.ifilm.com/creator/XCVmind
6. [6]
   XCVmind, "XCVmind (@id4823848)," vc.ru profile. Accessed: Sep. 26, 2026. [Online]. Available: https://vc.ru/id4823848
7. [7]
   misterXCV, "misterXCV (@id221242)," DTF profile. Accessed: Sep. 26, 2026. [Online]. Available: https://dtf.ru/id221242
8. [8]
   XCVmind, "XCVmind," Pikabu profile. Accessed: Sep. 26, 2026. [Online]. Available: https://pikabu.ru/@XCVmind
9. [9]
   OpenAI, "What to know about the Sora discontinuation," OpenAI Help Center. Accessed: Sep. 26, 2026. [Online]. Available: https://help.openai.com/en/articles/20001152-what-to-know-about-the-sora-discontinuation
10. [10]
    Google, "Gemini Developer API pricing," Google AI for Developers. Accessed: Sep. 26, 2026. [Online]. Available: https://ai.google.dev/gemini-api/docs/pricing
11. [11]
    Google, "Generate videos with Veo 3.1 in Gemini API," Google AI for Developers. Accessed: Sep. 26, 2026. [Online]. Available: https://ai.google.dev/gemini-api/docs/video