Release 0.33.0

Release Date: September 2026

Previous Version: v0.32.0

Editor and workflow experience

• Batch project actions. Select multiple projects - Shift-click for a range, or select all visible - then delete them or change their Team access together, with confirmation and permission-aware handling.

  
• A consistent project actions menu. Project cards and the editor now share one menu for duplication, naming and details, deletion, Team access, public sharing and link copying, and QR-code actions.

• Single-click section and group renaming. Click a section or group title to edit it inline, with a discoverable pencil indicator, explicit save and cancel controls, and keyboard support. Newly created annotations can be titled right away.

• Animation clip selection in the 3D viewer. Models that contain multiple animation clips now have a clip selector, so you can preview each animation individually.

Workspace assets and reuse

• A workspace asset browser in the editor. Browse images, 3D models, videos, audio, and PDFs in the new Assets sidebar, preview an item, and drag it onto the canvas to create the right input node. Images include raster, SVG, and EXR files, and the sidebar can be resized.

• Filters for finding existing work. Narrow the browser by source project, creator, uploaded or generated origin, browser or API upload source, and image subtype, and sort by creation date or file size. The asset viewer shows provenance and file details — source, creation date, size, and relevant dimensions or mesh statistics. The same discovery filters are available through MCP.

• Search assets by meaning. Describe what you are looking for in the selected media tab and get results ranked by semantic relevance. Semantic search also works through MCP and the project assistant; results form a single relevance-ranked page, while ordinary browsing keeps its sorting and pagination.

• Reuse assets across browser and API workflows. Uploads and generated outputs from projects with Team access become available to the whole workspace, including direct API uploads and exported-API execution assets. Workspace-authorized files can be reused by file ID in API workflows and downloaded through the authenticated result-download endpoint.

• Clear publication rules for shared assets. Turning on Team access publishes a project's existing files and any files added while sharing stays enabled. Published assets stay available to all workspace members even if the source project later becomes private or is deleted. Files in projects that have always stayed private never appear in workspace discovery, and the sharing and deletion dialogs explain this.

• SVG input & rasterize nodes. A dedicated SVG input lets workflows accept vector files directly, including assets brought in from the workspace browser.

• Expanded asset previews. The shared media view now supports asset browsing alongside node outputs, with on-demand thumbnails for supported 3D assets that lack a preview, so you can open an asset in the viewer before reusing it.

Project assistant, MCP, and API

• Run whole workflows through the MCP project agent. `prompt_project_agent` can now create and connect nodes, run or schedule them, wait for execution, inspect results, and keep working in the same call. Calls stay independent, without retained conversation history, and waits are bounded.

• Discover and retrieve published outputs through MCP. New tools — `list_workspace_assets`, `get_workspace_asset`, and `list_workspace_asset_creators` — expose asset discovery, metadata, creator filters, and authenticated download URLs. An agent-generated asset can be retrieved after publication without exporting and executing the project as an API.

• Ask the assistant to reuse existing assets. The project assistant can find workspace assets and place or assign them to compatible input nodes, including scalar and array inputs, and can reuse files already in the current project — in both browser and MCP workflows.

• Ask for a node's credit estimate before running it. The assistant can return a server-calculated estimate for planned node inputs, distinguishing an exact charge from a maximum hold and flagging assumed inputs.

• A visible waiting state in agent chat. When the browser assistant is waiting for a subscribed node to finish, the composer explains the wait, keeps the Stop control, and locks model and run settings until the wait ends or is stopped.

3D generation, texturing, and animation

• Hi3D v3.0 Quality and Master. New single-image and multiview backends replace the older Hi3D choices, with configurable face counts up to 5 million and texture de-shading controls.

• Hi3D v3.0 texturing. Retexture meshes with the updated backend, including multiple reference views and adjustable de-shading strength.

• Hi3D Split. Split a mesh in character or general mode, with mode-specific controls for templates, joints, connectors, or granularity.

• More relief-generation controls. Set relief height, background removal and its strength, and the retained base's shape and thickness in the Hi3D Relief/Depth workflow.

• Meshy v7 generation and texturing. Choose v7 triangle or quad generation and multiview generation, use Detailed or Ultra geometry where supported, and select 2K/4K/8K textures on supported Meshy v6/v7 paths. Supported generation paths also expose the pre-remeshed GLB.

• Meshy adaptive geometry controls. Choose adaptive Ultra/High/Medium/Low face density or an explicit face target in supported generation and remeshing workflows; Meshy-T2 adds explicit Auto or custom face targeting.

• Meshy Prototype and Build Meshy Prototype. Turn an image into a figure prototype with a preview and reusable prototype ID, then build that prototype into a GLB mesh.

• Meshy Text to Motion. Generate a standalone, Mixamo-compatible character animation from a text prompt in Prime or Swift mode, with a duration from 2 to 10 seconds. Requires animation-node access.

• More Rodin 2.5 controls. Use HighPack, HD/UHD texture controls, texture de-lighting, T/A pose, and bounding-box dimensions on supported paths. Fast Image to 3D gains a Rodin 2.5 Fast option with PBR control.

• Tripo P2 new 3D model. Tripo P2 Preview. Tripo's next-generation model for turning images and text into higher-fidelity 3D meshes also with an 8k textures option.

• Tripo animated-mesh output. Receive a complete animated GLB from Animate Tripo Mesh, control animation baking, and inspect individual clips in the viewer. Rigging and animation nodes expose Tripo task IDs; existing individual-animation outputs are marked as legacy.

  

Image and vector generation

• GPT 2.5 Sunburst and Flare.

• Reusable Recraft styles. Create a Style ID from 1–10 reference images with Recraft V4 or V4 Pro, then reuse it in supported raster and vector nodes. Choose Precise or Flexible style matching, and a vector-oriented base style when needed.

• Recraft V4 Text to Vector. Choose the base V4 vector backend alongside the existing Pro variants.

  
• Explicit GPT-Image-2 controls. Use one GPT-Image-2 backend with separate aspect/resolution and Low/Medium/High quality inputs in Quality Text-to-Image and the Multimodal nodes. Editing supports input-image matching, including a resolution choice for the matching mode.

  
• Qwen Image 3 and Image 3 Pro. Select the new models in image generation and multimodal editing, with negative-prompt and prompt-expansion controls.

Video generation and editing

• Gemini Omni 1.1 Flash. Generation and editing move to 1.1, with an explicit output resolution up to 4K. Generation supports 3–10-second clips and up to 10 reference images; editing applies to the first 10 seconds of the input.

• Wan 3.0 and Wan 3.0 Prime. New choices replace Wan 2.7 in text-, image-, and reference-to-video workflows, with 2–30-second duration, optional audio, and optional thinking; image-to-video also accepts an end frame.

  
• A broader Reference to Video node. The former Seedance Reference to Video node now also offers Wan 3.0 and Prime, with image, video, and audio references and resolutions from 480p to 1080p. The separate text- and image-reference video node supports a Wan array of up to 10 reference images.

Speech, voices, and music

• Gemini 3.5 Transcribe. A new Speech to Text node offers language detection and hints, custom vocabulary, speaker diarization, word timestamps, and smart transcription cleanup, returning text with any requested speaker or timestamp annotations. Smart mode can't be combined with diarization or word timestamps.

• Qwen Audio 3 TTS. Generate speech with preset voices, clone a reference voice, or design a voice from a description, with controls for language, a reference transcript, and speaking style.

• MiniMax Speech 2.8. Choose HD or Turbo speech generation and select a preset or reusable custom voice.

• Create MiniMax Voice. Clone a recording or design a voice from text, returning a reusable Voice ID and preview for later speech generation.

• MiniMax Music 3. Generate music with an explicit 1–300-second duration and controls for seed, inference steps, and guidance scale. This replaces the previous MiniMax Music v2.6 option.

• ElevenLabs Music v2. New music generation uses the updated Music v2 backend; legacy composition plans keep their compatibility path.

  

Text models, sign-in, and credits

• Updated text-model choices. The catalog adds GPT-6 Astra and a GPT-5.6 Sol option with reasoning disabled, and updates Gemini Flash to 3.8. Applicable text, vision, and agent nodes use the revised catalog and defaults; older selected models are remapped (see Changes and removals below).

• Continue with Microsoft. Microsoft joins the sign-in and account-creation options, with account selection and the usual email-verification flow where required. Availability depends on the environment's Microsoft authentication configuration.

• Credit estimates tied to your inputs. Node costs are calculated from the selected configuration. The editor shows the credit amount beside the run area, distinguishes maximum holds, and surfaces last-run cost and cached status, Free states, missing pricing inputs, and insufficient-credit states. The Node Pricing page shows the corresponding default-input estimates or bounds.