Team created an AI animation in 1.5 weeks and spent about 15,000 rubles

In an AI-generated video, the expensive part is not the first generation but the final revision. The team behind the “Hooligans at the Club” project tested this in practice: the stills were completed in less than a week and cost 3,000–5,000 rubles, while the animation required another 1.5 weeks and about 15,000 rubles. Editing, 2D effects, and sound then took roughly a week of manual work. Final replacements added another one or two weeks.

The result is a useful breakdown of hybrid production: neural networks did not replace artists and editors here, but accelerated shot discovery and movement between key scenes. Blender provided controllable geometry, Photoshop helped correct details without fully regenerating an image, and manual effects tied the material together into a coherent video.

Lock the visual design first

The work began with a mood board in Soul 2.0. Based on it, the team prepared locations, scenes, and character sheets. Individual shots were refined using Seedream 5 Pro, Nano Banana Pro, and Photoshop.

This sequence matters. If the character, clothing, environment, and composition are not locked before animation begins, every later change propagates through the entire pipeline: the source frame must be redone, the movement regenerated, the edit reviewed, and the effects repeated. That is exactly what the project’s creators encountered: final revisions took another week or two even though the main generation work had already been completed.

The practical product lesson is simple: approve a complete set of reference materials, not isolated attractive images. It should include character designs, key angles, the palette, props, and several control scenes. This package serves as a specification for the entire team.

A draft tests the prompt, not motion quality

The animation was created using Seedance 2.5 in draft mode. It initially generated video at 480p, making it cheaper to test the idea and prompt wording. However, taking a successful draft and simply upscaling it often produced artifacts.

The issue was not limited to weak textures. The low-resolution output also contained errors in physics, object trajectories, and camera movement. During subsequent enhancement, the model preserved the logic of the original clip—including flawed motion. The image became sharper, but the error remained.

The team therefore changed its process: draft mode was kept for prompt testing, while promising prompts were rerun from scratch at 1080p. This costs more than pressing a single “enhance” button, but reduces the risk of receiving a technically clean clip with incorrect camera work or broken physics.

A separate acceptance criterion is needed for this production process. A draft should be judged not on detail but on composition, direction of movement, object behavior, pacing, and camera work. If an error concerns these parameters, increasing the resolution will not fix it.

Control scenes are better built manually

For complex sequences, the team assembled simple low-poly mockups in Blender and supplied them to the model as a visual foundation. This framework defined poses and the relative placement of objects, while the generative model added appearance and movement. It provided greater control for scenes where a text description alone was insufficient.

Several lines of dialogue were voiced in ElevenLabs. Editing, sound, and 2D VFX were completed manually. This combination ultimately determined the result: AI accelerated production of the source material, but final coherence remained a human responsibility.

The main limitation of this approach is the high cost of late decisions. Generative shots may appear independent, but in the finished video they are connected through characters, editing, sound, and effects. Replacing a single foundational scene may require several specialists to redo their work.

A sensible workflow for the team is therefore to approve visual references first, use inexpensive drafts to test movement, then recreate the final shots at the required quality, and only afterward proceed to effects and sound. The generation budget remains relatively small, but savings emerge only when expensive rework is eliminated before editing.