Cubric Studio - Overview

Cubric Studio

Start here for the shape of the app before you dive into install steps, models, projects, prompts, queueing, history, and tools.

Pick a page from the sidebar when you need a specific answer, or use the videos below as a guided tour before reading the full docs.

What is Cubric Studio?

Cubric Studio is a desktop app for AI image, video, and sound work. It wraps ComfyUI as the generation engine, but the working surface is organized around projects, media, prompt controls, queueing, and history instead of node graphs.

The app keeps projects on your machine. Model files live separately in the engine's model folders, while project folders carry the media, history, settings, and metadata for the creative work itself.

Until version 2.0 the app was called Cubric Vision. The separate Cubric apps that were planned became one app, so it took the family name. Older videos and posts still use the old name.

New in version 2.1

A big update to 2.0, and 2.1.2 and 2.1.3 add more on top. The headlines, newest first:

  • MOSS-Image (2.1.3): a picture model that writes exact lettering on signs and posters, with a fast Turbo mode. Create boxes reads a picture you like and lays out boxes for a new one in the same layout, on MOSS-Image and FLUX.3 Klein.
  • Singularity H3 (2.1.3): a fine-tune of MiniMax H3 Reference that tears less and handles fast action better. Both now have a Fast Motion button, and Enhance describes each reference you attach so its details stay through the clip.
  • Use Tiles, redone (2.1.3): pick a tile size, and each tile is described before the upscale so grass is redrawn as grass. Use Grid is gone.
  • Stacks (2.1.3): drag pictures into a new order, turn a stack into a numbered sheet, and show only stacks from the Filter. With card info on, point at any card to see how it was made.
  • Changes to existing models (2.1.3): LTX 2.3 now comes in one version with sharper quality steps, Custom size works on Control too, and the NVIDIA PiD Upscaler now shows its non-commercial licence.
  • FLUX.3 Klein (2.1.2): Black Forest Labs' new image model makes and edits pictures with up to ten references, up to 4K, with a fast Turbo mode on by default. Draw Boxes where things go and name them in the prompt.
  • Custom size on Edit (2.1.2): an edit keeps your picture's size unless you turn on Custom size, on most image models that edit, local and cloud.
  • Cosmo can see pictures (2.1.2): on an Agent model that sees, such as MiMo-V2.6-Flash, it looks at a picture itself. Since 2.1.3 the default is DeepSeek-V4-Flash-0731 again, which is quicker but works from a description of the picture. Long chats also run longer before they sum up.
  • Ollama Cloud models (2.1.2): add one as Cosmo's model from the Remote panel, no terminal.
  • The Windows app is signed (2.1.2), so Smart App Control no longer blocks it, and a gated model's licence asks for your Hugging Face token only once.
  • Six new Flows: Video Edit swaps the person, head, outfit or background in a clip; Character Sheet from Images builds a sheet from pictures you already have; Character Sheet Editor changes one thing on a sheet, such as the clothes, hair or age; Transcribe turns a recording into text and an .srt file; 360 Panorama paints a whole world around a place; and Make seamless makes a picture repeat edge to edge.
  • Qwen-Image 2.1: makes and edits images with up to eight references, eight styles, and a No Background switch for transparent results. Its research licence means the images are not for commercial use.
  • Reference to Video on LTX 2.3: add pictures of a person, a place or a thing, describe each on its own line, and the clip keeps how they look.
  • 360 and 3D Scenes: fly around inside a 360 panorama, drop points along a path and render a 360 video along it, then look around any 360 video and take a picture of the view.
  • Mask and Extract for video: name what to keep, Detect follows it through the clip, and Extract keeps only that. Export for editing saves a ProRes 4444 clip with its transparency.
  • RunPod: what your Pod is costing shows in the status bar, the GPU picker rates each card for video as well as pictures, and you can upload your own LoRAs and upscale models to your volume.
  • Speech to text where you choose: dictation runs through DeepInfra, on your own GPU through ComfyUI, or through Ollama. Cosmo uses the same choice to hear the clips it looks at.
  • Cosmo does more: it installs a Flow after you say yes, sends cards to another project, and offers to start fresh or compact a long chat.
  • A child-safety check on every prompt, whoever wrote it: you, the Enhancer, Cosmo or a connected AI app. Adult content is untouched.
  • Smaller additions: Use Tiles in Upscale, Combine for pictures, a hover panel showing how a History entry was made, starred projects, dev updates on the home screen, every background job under Transfers in the queue, and a LoRA & Upscale button in the settings popup.

Changes to how things already worked:

  • The old CubricVision.exe is gone on Windows. If a shortcut or taskbar pin still points at it, pin CubricStudio.exe instead. See Installation.
  • The engine moves to ComfyUI 0.42.1 and updates itself on first launch. Enhance and Describe no longer need the Image Describer plugin, which is gone. See Getting Started.
  • Resize runs inside the app, with no engine needed, and always uses Lanczos, so its Method setting is gone. See Video Tools.
  • The Projects page has one Library link with a Models | Flows switch, in place of separate Models and Flows links. See Projects.
  • Character Sheet has one Headless front body button: every sheet comes with both versions, and the button switches between them.

Watch the overview

Most of these videos were recorded when the app was called Cubric Vision, before version 2.0. The main one is nearly an hour long. Use it as an alternate starting point if you prefer to watch the product flow first. The written docs remain more complete for specific controls, edge cases, screenshots, and reference details.

The video covers the app from a different angle than the docs, so it is useful for orientation even when a page has more detail.

Shorter walkthroughs cover the 1.1 and 1.2 releases, remote setup, and the agent:

New in versions 1.1 and 1.2.
How to set up the RunPod remote engine. See Remote for the in-app fields.
The in-app agent running on DeepInfra, plus how to connect another AI agent over MCP at the end. See Agent for the in-app steps.

How the app flows

The main path is Landing to Gallery to History. The Landing page is where you create or open a project. The Gallery is the default project workspace, where media cards, prompt controls, model choices, and the Cue queue come together. History opens from a Gallery card when you want to inspect one image or video group and continue from a specific result.

The Prompt Box appears in the Gallery and History workspaces. It snapshots the selected model, operation, prompt text, media inputs, and controls when you press Cue. Loop builds on the same Cue path when you want the app to keep generating from the current Prompt Box state.

  • Start with Getting Started if you need the download, first-launch setup, model folder, and Settings path.
  • Use Projects when you need to understand where project data lives and how project folders move between machines.
  • Use Prompt Box and Queue when you are learning the generation loop, Cue, Loop, model settings, media inputs, and resumable video controls.
  • Use History, Image Tools, and Video Tools when you are refining an existing result.
  • Use Models and Hotkeys for install states, operation support, and keyboard shortcuts.