Skip to content
Konte
How it works
SH-03 · 8s loop · KONTE · MiniMax H3 on a local GPU

HOW IT WORKSHow Konte works

From your words to footage

The chat AI, the shot list, the render engines. What happens inside Konte, on one diagram.

01—THE LOOPFrom conversation to footage

Every conversation adds videos and assets

Planning and shooting both happen as you keep talking. Everything you make stays in the project.

  1. 01—TALK

    Talk

    Describe the video you want in one sentence. Attach cast, sets or photos too.

  2. 02—PLAN

    Get a shot list

    First a plot; if it works, a breakdown with chapters and shots. Keep talking and it gets revised.

  3. 03—SHOOT

    Press to shoot

    Once the cast and set reference images are ready, press “Generate from this shot list” (この構成表で生成する). Earlier takes stay in the history.

  4. 04—KEEP

    Keep it on the shelf

    Videos appear beside the conversation, shots go to Outputs. Cast and sets carry over to your next piece.

↺ Your next piece starts from this shelf

Konte app screen (in Japanese): Talking with AI. Talk, and it moves from plot → shot breakdown → cast and sets → rendering. Finished videos collect on the right.
Talking with AITalk, and it moves from plot → shot breakdown → cast and sets → rendering. Finished videos collect on the right.

02—THE MAPThe whole picture

Konte, the studio around the model

The model draws the picture. Konte handles the planning before and after it, and the shelf.

YOUYou

Talk in the browser

  • The video you want, in one sentence
  • Keep talking to fix it
  • Pick shots and shoot them

Claude Code / Codex on your machineOptional

Connected over MCP, it can read the shot list, add shots and edit the wording.

KonteSTUDIO

SHOT LISTThe breakdown

Video›Chapter›Shot

One shot = one row. Picture notes, camera, cast, length.

AIClaude

Breaks the conversation into shots, writes up the prompt right before rendering, reads your cast and sets.

SHELFThe shelf

CastSetsWikiOutputs

It grows with every shoot, and your next video starts here.

QUEUERender queue

To the organization's default engine

STAGEWhere it renders (one per organization)

ComfyUI

Local GPU

Your own GPU. Renders from a workflow (JSON)

MiniMax H3

Generate at 480P → ESRGAN x2 → 1080p

Seedance In preparation

A slot that becomes selectable once a workflow is added

Cast and set sheets

Turnarounds and wide / close views, from reference images

OR

fal.ai API

Cloud

No GPU. Usage is billed to your own fal key

MiniMax H3

No quantization, 5–15 s

MiniMax Hailuo

6 / 10 s

The organization's owner picks one place to render, under Default render engine (既定の生成エンジン). Shots of one video rendered on different engines wouldn't match.

03—ONE CUTHow one shot renders

From one row of the shot list to an MP4

What happens behind the scenes, in order, when you shoot one shot.

  1. 01—ROW

    One row of the shot list

    Picture notes, lines with their speakers, camera (angle, lens, move, speed), cast, length. One shot = one row.

  2. 02—CAST

    Add the cast

    The reference images you registered go into the shot as keyframes, so the face and wardrobe match.

  3. 03—PROMPT

    Write up the prompt

    Right before rendering, Claude writes the row up as an English prompt. Konte sets the structure and camera terms; the AI only writes the description. Lock it to keep your own wording.

  4. 04—QUEUE

    To the engine

    The job goes to the organization's default engine. On your own GPU, one at a time in order; in the cloud, in parallel.

  5. 05—RENDER

    Render

    MiniMax H3 generates picture and sound together. On your own GPU it renders at 480P, then brings it up to 1080p.

  6. 06—SHELF

    Onto the shelf

    The MP4 goes to Outputs. The seed, prompt and camera settings are kept, so you can go back to an earlier take any time.

04—ENGINESRender engines

Render engines, by where they run

Named the way the shot list shows them (in Japanese, in the app).

  • MiniMax H3 on a local GPU

    Local GPU

    VIA ComfyUIThe default engine. An int8 build that fits in 12GB of VRAM, rendered at 480P and then brought up to 1080p. About 139 s for a 4-second shot (measured).

  • Seedance on a local GPU

    Local GPUIn preparation

    VIA ComfyUIThe slot is in place: add a ComfyUI workflow and it becomes selectable (the workflow is still in preparation).

  • MiniMax H3 via the fal.ai API

    Cloud

    VIA fal.aiThe same model as the local one, without quantization. 5–15 s per shot, 768P by default. Usage is billed to the fal key you registered.

  • MiniMax Hailuo via the fal.ai API

    Cloud

    VIA fal.ai6 or 10 s per shot. No waiting in line for a GPU; in exchange, usage is billed to the fal key you registered.

05—UNDER THE HOODComfyUI inside

New models, same shot list

Underneath is ComfyUI. Engines are added as workflows.

  • Hand over a workflow, get footage back

    Konte hands ComfyUI a workflow in API format (JSON) and takes back the finished footage. Nodes are referred to by name, so swapping the workflow leaves Konte's side untouched.

  • The GPU can be another machine

    Files move over HTTP only. The Konte server and the GPU machine running ComfyUI can sit apart.

  • More models, same shot list

    New video models are added as ComfyUI workflows. Seedance's slot is set up the same way. Your shot list and cast carry over as they are.

GET STARTEDStart here

Your film studio, starting today

Start by creating a single cast member. We're happy to talk it through.

Contact us