Shared plans · Selective rendering
For Lazy Storyboard V2, we documented specifications and integration directions for sharing plans and separating rendering. Video generation itself was not first introduced at this stage.
A2P / RESEARCH & DEVELOPMENT
From AI generation
to production tools.
From asking whether we can generate it, to asking whether we can create the scene we want.
Explore how AI
Lab's experiments and development grew into a production experience.
Research and development records are presented with the latest work first. Some research ran in parallel;
experimental results are distinguished from implementation and design records.
01 / A structure for repeatable production
Re-create while preserving
the story and style.
We analyzed why storyboards compressed events and directing information, and outlined a structure that separates story planning from image and video rendering.
Intended production experience
We aimed to let users compare results from the same story plan, produce only the parts they need, and repeatedly apply their own styles.
Reusable custom styles
We outlined a structure for converting user descriptions into prompts for each generation stage, registering them, and reusing them through an assigned style ID.
Two tests identified the fixed nine-shot limit and equal allocation across scenes as the main causes of compression. Even shot-heavy climaxes received the same shot budget, losing directing information.
This stage primarily consists of design, specification, and integration records. No measured improvements in cost, user satisfaction, or story retention were reported, so the plan-based structure cannot be said to have solved the compression problem.
Next direction
The work points toward comparing generated results from the same story plan and selectively revising and producing parts of it.
02 / An experience users can direct
Bring a generated scene
closer to the user's intention.
We reconstructed images as editable 3D scenes and used poses, objects, and cameras adjusted in PoseShot and 3D Proxy for image and video generation.
POSESHOT / 2026.07 — 08
Direct control, from poses to cameras.
Feature demonstrations from actual development records.
Apply poses by body part
Select a region, such as the upper or lower body, and apply a pose.
Hand Editor: control down to the fingertips
Edit hand poses directly to create the expression a scene needs.
Record camera paths
Record camera movements and pose changes based on saved frames.
Connect one pose to the next
Interpolate the transition between different poses.
Intended production experience
We aimed to let users visually specify compositions and poses that are difficult to describe in prompts, then correct the reconstruction themselves to reach the scene they want.
PoseShot · Pose and scene editing
We added primitives, move/rotate/scale controls, hand and partial-pose editing, copy and paste, and manual person detection, then expanded into camera path recording and pose interpolation.
3D Proxy · Reconstruction and manipulation
We improved character, camera, and object reconstruction and alignment, and tested reconstruction streaming and rendering performance separately.
Performance validation · Two distinct experiments
61.9% / 54.7%
Average reduction in first-object response / total completion time
Reconstruction streaming experiment: four successfully completed images, with up to four objects each
34.3 → 60.0
Median FPS
A separate rendering experiment measuring the same scene three timesUser testing indicated that visual composition editing was more intuitive than text instructions. This was qualitative feedback, not a measurement of satisfaction or usability improvement rates.
Occluded areas, complex interactions, and large viewpoint changes remained unstable. Reconstruction errors also affected generation results, so manual correction was necessary.
Next direction
Rather than focusing solely on perfect reconstruction, we emphasized a workflow users could correct and iterate on.
03 / Connecting the production workflow
Connect generated elements
into consistent scenes.
We derived characters, locations, and props from scripts, connected references with 3D previs, and passed the required information to each shot.
EXPERIMENT / 2026.07
Combine references and connect them into scenes.
We assigned roles to reference images to generate characters, then used them with locations in storyboards.
01 / Input references
02 / Generated character
03 / Add objects to a location
We combined a torii gate reference with a prompt for a Shibuya background.
04 / Connect to a storyboard
Intended production experience
The goal was to carry the desired appearance and spatial composition across scenes and reuse prepared assets.
Elementalist · Reference composition
We tested character generation combining face, top, bottom, and hairstyle references; location-object combinations; text-based prop generation; and connections to the storyboard.
Improved scene references
We reflected character visibility in generation and fixed an issue where the default background was referenced instead of the generated one. We analyzed why A-poses persisted and improved the prompts.
We verified the feasibility of character and prop replacement and reuse of existing 3D backgrounds. In a validation shot, regenerating the same input resolved the A-pose while preserving composition.
Identifying three or more people, background detail consistency, complex interactions, and contact with floors and chairs remained unstable. Prompt improvements did not solve upstream motion-application issues.
Next direction
Alongside better reference transmission, we explored editing tools that let users directly control the scene itself.
04 / Testing feasibility
How far could AI go
in video production?
We tested whether AI could create the characters, backgrounds, motion, and voices needed for video and connect them into a single production workflow.
EXPERIMENT / 2025.11
The same workflow, different results.
Shot images from experiments testing character replacement through video generation.
The unsuccessful experiment recorded character mix-ups and unwanted people appearing. These images are a collection of shots from the experiment, not individually annotated error examples.
Intended production experience
We aimed to reduce repetitive production tasks and reuse prepared characters and scenes in other content.
Character and scene generation
We experimented with background editing and sequential character replacement, validating a workflow from images to video.
Motion and audio generation
We experimented with dialogue-aligned motion and voice generation, as well as story-based background music.
87%
Success rate for dummy character removal
56.5%
Share of final clips with no defects
Results from the same 108 clips evaluated under two different success criteria.
In a test of 108 clips, success measured by dummy character removal was 87%, while the share of final videos with no defects was 56.5%. Being able to generate something and being able to use it directly in production were different standards.
Limitations remained in overlapping characters, body proportions, pose reproduction, and background consistency across camera angles. Unwanted people appearing in video could not be fully prevented.
Next direction
We expanded the research beyond individual generation functions toward a production workflow that carries the same characters and spaces across multiple shots.
FROM RESEARCH TO CREATION
Beyond generated results,
toward production that expresses intent.
We tested feasibility, analyzed limitations, and added tools for direct editing.
A2P's records trace our
exploration of both better generation technology and better tools for making things.