Big Blue Ceiling

b2c3d

Traditional 3D content creation, rebuilt with AI inside the tools.

Eddie OffermannFounder and architect

Role
Founder and architect
Areas
AI, VFX, Pipelines, Tools
Three-quarter view of a 3D night-market scene generated in b2c3d: Miso, a fox in a white apron and blue headband, beside his wooden noodle cart with a red canopy, steaming pots, stools and paper lanterns.

Generative models moved the cost of 3D from authoring to steering. A generated chair is a draft, and the real work is getting it right. We built b2c3d for that part: review, correction and iteration, with nothing destructive and nothing anonymous. The goal is Maya as it would be designed from scratch, with generative models as first-class citizens.

Miso and the noodle cart

Our demo character is Miso, a fox who cooks noodles at a night-market cart. Both Miso and his cart were made entirely inside the suite:

  1. Concept. A painting of the night market, generated in the Concept app.
  2. Isolate. Miso and the cart were each lifted out of the environment, front on, on a plain ground.
  3. Every side. Each got a turnaround: a ring of consistent views plus views from above.
  4. Build. Each was reconstructed in 3D from its views, then textured from the same views.
  5. Rig. Miso was given a skeleton and skin weights, including his tail, and the Rigging app lets an artist correct whatever the model got wrong.
  6. Animate. Motion came from the curated mocap library and from text prompts, retargeted onto Miso's own rig, then trimmed, looped and blended.

Every step is a job you can watch, cancel and repeat, and every result is an asset that remembers the step before it. Concept to Character follows Miso and his cart through each step with the images from every stage.

The applications

The suite is one web app, with applications laid out in the order of the work:

  • Concept: describe an image and generate it, then take it straight into Studio.
  • Studio: turn an image into a scene of separate objects, then texture, light and edit it.
  • Rigging: give a mesh a skeleton and skin weights, then correct what the model got wrong.
  • Animate: put motion on a rig, edit it, plant its feet and play it back.
  • Assets: everything made or imported, by where it lives and what it came from.
  • Organizations and People: shows, sequences and shots, and who can reach them (covered with Assets).

Most image-to-3D models treat a picture of a room as one object, with the sofa fused to the rug. b2c3d separates the objects first and generates them together, so they come out as separate meshes that already share a frame. It began as a fork of the MIDI-3D research model, and everything around it (the platform, the applications, the asset model, the job system and the other stages) is our own.

Artists working in Blender and Maya reach the same assets through BigBluePipeline, which treats b2c3d as its single source of truth.

How it is built

  • A capability is an operation, not an endpoint. Each one is a single typed function that registers itself and gets an HTTP route, a command-line entry, a job with progress and cancellation, and an MCP tool for agents. A person in the web app and an agent reach the same product.
  • Assets are immutable. An edit writes a new version with the old one as its parent, and content is stored by hash, so history is a graph and nothing is overwritten.
  • Progress tells the truth. A phase with no honest denominator shows an indeterminate bar rather than a number that stalls.

The stack is Python with FastAPI, Postgres, S3-compatible object storage and Redis for the job queue, with a React and TypeScript client using react-three-fiber, running in CUDA containers. Architecture goes further, including how BigBluePipeline's Exchange bundles land in the library.

Attribution by design

Provenance is part of the data model. A textured mesh knows the mesh it was baked onto, which knows the mask it was segmented with, which knows the image it started as. Each generation records its model, seed and settings.

A model may be depended on for release only if its training data is released and openly licensed. Models that don't pass yet are labeled interim or R&D only, and that label travels with every output they make, through retargeting and every edit. Bundled materials and motion come from openly licensed libraries, credited on every clip. Provenance and Attribution sets out the policy, the standing of each stage's models and every credit.

Walk and idle motion from the 100STYLE dataset (Mason, Starke and Komura), CC BY 4.0.

  • The b2c3d marketing site's opening section, headed 'A scene, not a blob.', with three panels: the painted noodle cart concept, the same concept with one box per object, and the generated 3D scene lit by its lanterns.
    A scene, not a blob: the concept, one box per object, and the scene that comes out.
  • On the left, the painted concept of Miso the fox at his noodle cart; on the right, Studio showing the generated scene of the same cart from a three-quarter angle, with Miso beside it.
    The concept beside the scene Studio built from it.
  • Painted concept art of a fox noodle cook in a white apron standing at a wooden noodle cart under a red canopy at a night market, with paper lanterns and steaming pots.
    Where it starts: concept art generated in the Concept app.
  • The Studio app at its Segment stage, showing a painted night-market concept of the fox and his noodle cart, ready for boxes to be drawn around each object.
    The concept handed to Studio, where each object gets a box of its own.
  • Miso, a cartoon fox in a white apron and blue headband, standing on a plain grey background.
    Miso isolated from the environment concept, front on, on a plain ground.
  • Miso cut out on white in an A-pose, arms out and tail to one side, ready to be built in 3D.
    Cut out and posed for reconstruction.
  • Six renders of the finished, textured Miso model from the front, three-quarter, side, back, rear three-quarter and above.
    The finished asset, from every side.
  • Four close-ups of the same furry dark-brown hand in different styles: human fingers, gloves, paws with pads, and claws.
    Four hand styles for Miso: human, gloves, paws and claws.
  • The untextured grey Miso mesh with its skeleton joints shown as yellow dots, including the tail.
    Miso rigged, tail and all.
  • The Rigging app with Miso's grey mesh and joints in the viewport, auto-rig settings on the left and a list of 52 skeleton joints.
    The rigging UI: auto-rig, inspect every joint and its skin weights, then correct what the model got wrong.
  • Miso running. This motion came from an interim text-to-motion model, so the clip is research, not shipped content.
  • Miso walking on his own rig in Animate. Walk motion from the 100STYLE dataset (Mason, Starke and Komura), CC BY 4.0.
  • The Animate app with Miso's rig loaded and a text prompt reading 'a person ladles noodles from a pot into a bowl', with a notice that the motion comes from an interim model.
    Motion from a text prompt, generated onto Miso's own rig. The interim-model notice travels with the clip.
  • The Animate app playing a forward-walk motion-capture clip on Miso's rig, with trim and retime controls and the clip's dataset credit.
    Editing a clip. Library motion carries its dataset credit. Walk motion from the 100STYLE dataset (Mason, Starke and Komura), CC BY 4.0.
  • The Animate app showing Miso standing in an idle motion with his joints visible.
    An idle loop on the same rig. Idle motion from the 100STYLE dataset (Mason, Starke and Komura), CC BY 4.0.
  • Six views of the generated noodle cart: front, three-quarter, side, back, rear three-quarter and top, showing its red canopy, wheels, pots and wooden body.
    The noodle cart, built from its own turnaround of views.
  • The textured noodle cart lit on a dark ground, with a red cloth canopy, metal pots and a wooden body on spoked wheels.
    The cart, textured.
  • Cart turntable.
  • Studio's Look panel on the noodle cart scene with Miso beside the cart, showing a warm lantern-glow point light, a cool rim light and environment settings.
    Look development: lights, materials and environment, all in the scene document.
  • Front view of the assembled night-market scene: Miso stands at the noodle cart with stools in front and lanterns hanging overhead.
    The assembled scene.
  • Side-by-side comparison of the painted concept art and the generated 3D scene of Miso at his noodle cart.
    Concept and result, side by side.
  • The b2c3d application launcher, with cards for Concept, Studio, Rigging, Animate, Organizations, Assets and Your profile.
    The suite, laid out in the order of the work.
  • The Concept app with a text prompt describing a mid-century modern living room, an inherited film-look style, and a grid of four generated variations of the room.
    Concept: styles are inherited from the show, sequence or shot.
  • Studio's Segment stage on an illustrated living room, with boxes drawn around an armchair, a dog, a side table and three plants.
    One box per object you want as its own mesh.
  • Studio's Clean up stage showing the generated untextured geometry of the armchair, dog, side table and plants as separate grey meshes.
    The objects are generated together, so they come out already placed relative to each other.
  • The same living-room scene after texturing, with the outliner and a history list where each edit is a new version.
    Restoring an old version adds a new one; nothing is removed from the chain.
  • The Assets app showing a grid of generated meshes, with a side panel tracing the selected scene back to its source image and segmentation.
    Assets: every asset walks back to what it came from and forward to what uses it.
  • Diagram of a provenance chain: a textured scene produced by texture_scene from untextured geometry, produced by generate_scene from an instance mask, produced by segment from the original uploaded image. Each step shows a content hash.
    Provenance is part of the data model, not a report written afterwards.
  • The Organizations app showing a demo studio with a show, its sequences and shots, frame ranges, and per-level access grants.
    Shows, sequences and shots. Access granted on a show reaches everything beneath it, never upward.
  • Layer diagram with imports flowing downward: the React web client, the FastAPI layer, the job queue, the operation registry, then model runtime and content-addressed assets, with vendored model code at the bottom.
    How the platform is layered.

Sources

  • BigBluePipeline

    A VFX studio pipeline, rebuilt as the artist-side front end for b2c3d.

    • VFX
    • Pipelines
    • Tools

  • Depth2STL and Obj2GLB

    Two small open-source tools for getting 3D data where it needs to go.

    • Tools
    • VFX

  • QIE Studio

    An image generation and editing studio, built by an AI agent working unattended from our specification.

    • AI
    • Tools
    • Apps
    • VFX

  • AuRA Avatar Generator

    Customizable, photorealistic presenters for AI-generated video briefings.

    • AI
    • Tools
    • VFX