# Two Years, 1.2 Million Images, 100 Full-Film Passes: What Bona's Account of Sanxingdui Says About Long-Form Consistency

- Published: 24 October 2026
- Updated: 24 October 2026
- Author: RenderBob team
- Category: Workflow
- Canonical: https://renderbob.ai/blog/sanxingdui-production-account-long-form-consistency

Bona Film Group has published numbers for how its all-AI feature was made. They are company figures, but they show where the effort in a feature-length AI film goes.

Bona Film Group has published numbers for how its all-AI feature was made. They are company figures, but they show where the effort in a feature-length AI film goes.

## What Bona says it did

TechNode's report of 28 September, drawn from Bona's announcement, says the creative work was done by a team of more than 100 people in Bona's AIGMS division, following a conventional film-production workflow. A system called Infinite Canvas, developed by Boka Film Cloud, put scripts, storyboards, characters, sets, lighting and other audiovisual assets into one environment. Humans kept the creative concept, the narrative direction and the final artistic decisions. The article does not name the generative models used.

The reported scale:

- Time. Two years.
- Images. More than 1.2 million source images.
- Iterations. Over 100 full-film iterations.
- Targets. 4K image quality, high dynamic range colour and multi-channel surround sound.

A separate wire report on opening day gives nearly 100 people and a cost of more than 10 million yuan, again from Bona.

## The most useful detail

To deal with consistency across a feature-length film, the team calibrated details such as characters' micro-expressions by hand, frame by frame. That is the part of the account worth remembering. A long film is not a long prompt. It is a large number of generated shots, and the cost sits in checking them against each other and fixing what drifts.

## What to treat with care

- "Source images" is not defined. It could mean reference material, generated frames or something else.
- A "full-film iteration" is not defined. We do not know whether each one regenerated every shot or reassembled existing ones.
- No tools or models are named, and there is no independent account of the process.
- Team size differs between reports, more than 100 against nearly 100.
- The cost and time figures come from the producer.

## How to use it when planning your own long piece

- Budget for passes, not for generation. If a feature took over 100 full passes, plan time for review and re-rolls on every scene.
- Keep every asset in one place. Characters, sets and lighting references should be easy to find and reuse.
- Fix the target spec early. Resolution, dynamic range and sound format shape what you generate.
- Reserve hand-fixing time for faces and continuity, which is where this team spent it.
- Write down who decided what. Bona's account separates human creative decisions from generated material, and a record like that is easy to keep if you start early.

Our 21 October checklist on ten-minute claims asks whether a long video is one coherent pass or stitched shots. Bona's account describes a heavily supervised process with many full passes, which is a more believable picture of how long AI films get made today, though it does not say whether the shots were stitched or generated in one run.

## What to watch

Whether Bona names its tools, publishes a making-of, or shares its asset-management approach.

## RenderBob 3 and this story

RenderBob 3 treats iteration as the normal state: each generated clip sits on its own lane, so a re-roll or a model swap touches one shot, and prompt logs and references stay with the studio across a hundred passes.