AI cartoon series: a 24-second episode built ten ways, the best cost 162 credits

The team that makes the vertical cartoon series "Forest Street" entirely with cloud models has published a production breakdown. Between October 29 and 31, 2026, it rebuilt a single episode, "Agitation Forbidden", in about ten ways using Wan, Kling, Veo, Seedance, GPT Image and Nano Banana Pro. The episode runs 24 seconds, with recurring characters and Russian speech. Costs are given in Higgsfield credits (the internal currency of the service where Kling runs) and in Flow bonuses (how Veo is billed). The units cannot be added together. There are no audience figures for any method yet, and each test was a single run, so the breakdown should not be read as a model ranking.

Order of steps: a cheap draft first, then animation

The pipeline: character bible, second-by-second script, human approval, reference character sheets, frames in GPT Image, ElevenLabs voices, assembly with camera movement, checks, and a human review of the draft. Only then is the scene handed to Kling 4.0: one continuous shot, 2–3 takes. That way animation money is spent only on a story that has already passed review.

The authors' main conclusion: the weak point was vague prompts, not the models themselves. The script is written out second by second: who speaks, with what delivery, what the listener does, where the pause falls, when the object is revealed. In the third Kling prompt they added a ban on echo, exact intervals between lines and a "ZINA ONLY" tag. All five lines were recognized without repeats. Before release, a fresh AI assistant retells the scene frame by frame without sound. If the meaning does not match the intent, the episode does not ship.

Ten tests on one story

All tests are from October 30, the story "Agitation Forbidden", about 24 seconds:

• Wan 3.0 Prime + Speak 2.0, shot by shot: 116 credits, about 47 minutes. The clothespin is already in hand, and the letters were drawn in on their own. Rejected by a human.
• Kling 4.0 + Avatars 2.0, shot by shot: 61.75 credits, about 17 minutes. Extra letters on the T-shirt, and Zina blinks instead of squinting. Rejected.
• Veo 3.1 Fast, shot by shot: 180 bonuses, about 29 minutes. The character sheet moved instead of the clothespin. Rejected.
• Images with camera ("comic"): no new cost. Noticeable shot changes, rejected.
• Kling 4.0 in one go: 54 credits, about 6 minutes. The vatrushka pastries are visible in advance, the board has pseudo-text, the room changes. Filtered out by the automatic check.
• Seedance 2.5 in one go: 168 credits, about 8 minutes. Two Zinas, wrong captions, no sound. Filtered out.
• Veo 3.1 Fast, three parts: 80 bonuses, about 10 minutes. Characters blended together, the vatrushkas are not revealed, loudness −18.3 LUFS. Filtered out.
• Kling 4.0 with second-by-second direction, 3 takes: 162 credits, about 16 minutes. One shot, five lines, vatrushkas revealed in the finale. Accepted.
• GPT Image + camera, 10 frames: 12 generations, 373 characters of voiceover. Poses change abruptly, text has no errors. Accepted as the cheap version.
• Nano Banana Pro 2K + camera (unlimited): poses change instantly, the pigeon is out of place. Not rated.

In a blind test of six clips (162.5 credits), the reviewer liked Wan 3.0 Prime most. But Kling did better on the scene with dialogue and emotion, and the choice did not hold. Seedance was dropped from production over quality, even though an unlimited plan covered it. Video models mangle letters, so all sheets and boards in the video are now kept clean, and the lettering is drawn in GPT Image and Nano Banana images. Video is generated at 720p and upscaled to 1080×1920. A Kling generation at 1080p cost 96 credits versus 54.

Voice and checks: what automation cannot hear

Voiceover is done with ElevenLabs eleven_v4 using emotion tags. For eight characters they ran a casting of 24 candidates (2,992 credits). Zina's first voice, Nadia, was dropped: it is tagged as young, while the character is 74. Babushka Valya was chosen, but by description, not by ear. Slow recordings are sped up to ×1.30 without changing pitch.

Kling speaks in its own voices, so the speech is replaced. Audio Isolation separates the voice, lines are cut into windows, each goes through Voice Changer (eleven_multilingual_sts_v2) and is laid over a quiet background. The video stream is unchanged; its SHA-256 hash matches before and after. Word boundaries shifted by at most 0.06–0.08 s according to Scribe, the ElevenLabs speech-recognition service. This does not confirm lip articulation.

Checking is separate from assembly. Frames are captured every 0.25 seconds, about 100 per clip, and an AI assistant reviews them. Speech is recognized by a local faster-whisper small. Loudness is measured with ffmpeg and normalized to −14 LUFS. The native track fluctuated: Kling around −11.7, Veo −18.3 with a dip to −23.3 in the third part. No AI assistant can hear audio: uploading a WAV returned "audio input not supported". Recognition returns words, not stressed syllables, so it cannot check stress. They could not connect a 31-word dictionary: the API key lacked the pronunciation_dictionaries_write permission.

What remains unsolved

Transferring a good image into video is unreliable. In "А телефон куда?" a box covers the tray, a lamp shifts with no visible action, and the phone rises before the pause. Tyoma's sheet looks younger than 16, and every frame inherits that. The image-based versions double silhouettes on five-frame transitions. There is no by-ear confirmation of the voices. Most importantly, the authors do not know how many rubles and hours of human time it takes to get a good episode regularly, rather than one lucky attempt. The practical takeaway for anyone making AI video: write the scene out second by second, test it on cheap frames, and do not trust an automatic check report without your own ears.