AIVID

AI short film studio: turn a script into a coherent short film where the person in shot 30 is still the person from shot 3.

What this is

A disciplined AI filmmaking pipeline: script, character and scene reference art, per-shot prompts, and gacha-style acceptance checks. Every step has a format and a record. The source of truth for each film is three markdown files (script, production log, asset inventory), and every image can be traced shot by shot to the prompt that produced it. v2 turns it into a product: accounts, credits, crypto payments, and cloud generation.

What problem it solves

Making a flashy 5-second AI shot is no longer hard. The hard part is making a whole coherent film: characters drift, sets change, lighting jumps. The entire reason this product exists is to lock down "consistency" with a pipeline.

Use case

Tokyo Night Anomalies: On an ordinary late night in Tokyo, someone casually films the night streets, and shot after shot runs into anomalies that should not exist. 7 shots, 12 reference images, 4 of which were blocked by the platform's copyright filter (Gundam lettering, a Naruto forehead protector, a Sharingan). I didn't bypass the filter; I regenerated clean versions that were 'similar in spirit but original.' The originals are archived and marked BLOCKED, kept only as composition reference.

Tokyo Night Anomalies: the torii shot

BLOCKED-ramen: forehead protector, cloak, eye pattern — the filter caught the combination of identifying marks

ramen approved version: same shot reworked with an original character, composition preserved

Highlights

• Open-source methodology: Emily2040's Seedance 2.0 Skill OS (MIT), with 24 sub-skills and 61 evals; I'm an adopter and added a few references. The core idea in one line: direct the model, don't micromanage the frame.
• Three AI access paths: local claude with no key, bring-your-own key, and platform metered billing, all consolidated through one routing layer.
• Crypto payments use the amount as the order number: pay 20.000347 USDT; the trailing decimals identify the order, and credits are only issued after on-chain finality.

Where it is now

Three weeks, 129 commits, roughly 90,000 lines (including 13,000 lines of third-party packages). Not deployed; the pipeline code for per-shot film generation is written, but it's stuck on cloud model access, and it has not yet produced a single real mp4. A go-live review found the multi-tenant content layer wasn't isolated; it was fixed the same day and 33 security tests were added. The argument is productization speed and methodology accumulation, not live operations.