Updates

H3 Max Director now offers higher quality 1080p output, 15-minute session lengths, and enhanced media inputs that support audio tracks and end frames. Live directions in the Director can now carry audio and one-shot target frames.

1080p Director Sessions

The H3 Max Director now supports 1080p resolution output, billed at twice the standard rate. Session lengths have been increased from 2 minutes to a maximum of 15 minutes. The interface displays your account’s session allowance, calculated from your available credit balance, giving you a clear view of how long your session can run before it starts.

Director Media Inputs

You can now provide media directly to the Director at session start. The node accepts a Start Frame, an End Frame for the first chunk, and a Target Audio soundtrack. Audio bitrate can be configured manually or left on Auto. The system validates that media URLs are accessible to the provider before starting a session to prevent failed attempts.

Live Directions with Media

Directions sent during a Director session can now include audio attachments and target end frames alongside text prompts. Use the “Apply at next chunk” switch to insert directions immediately or append them to the existing queue. The transcript and captions track audio applications and rejections separately from text prompts. An audio behavior setting controls whether a track replaces the current audio or queues to play next.

Quality of life

• Clips in the Compositor timeline now render at their true proportional width, even when very short.
• Narrow timeline clips display a single, bidirectional trim handle for easier manipulation.
• Dragging from a node’s input handle to empty canvas and selecting “Upload File” or “Import from Library” now correctly wires the new Import node.
• The Generate button in the Compositor is no longer blocked by unrelated active jobs.

New models headline this release—OpenAI’s GPT Image 2.5 is now available, and MiniMax H3 gets a draft-quality resolution to save on costs. Your realtime session recordings are now visible in the asset library for organization.

GPT Image 2.5

OpenAI’s GPT Image 2.5 is now available as a new node. The model includes two distinct variants you can select: Flare, optimized for speed and natural lighting, and Sunburst, which offers higher precision and detail for premium visuals. The node automatically switches between text-to-image and image editing modes based on whether you connect an input image, and you can specify image size, quality, and background transparency.

Lower-Cost Drafts for MiniMax H3

The MiniMax H3 node now includes a resolution selector, letting you choose between the standard 2K output or a new 768p draft tier. The draft resolution is approximately 30% cheaper and is ideal for testing concepts before committing to a full-quality render. The default remains 2K to preserve existing workflow behavior.

Realtime Recordings in Asset Library

Video recordings from realtime sessions, such as those generated by H3 Max Director or Decart Lucy nodes, are now accessible in the asset library. You can favorite, tag, rename, move to folders, or soft-delete these recordings just like other generated media. They appear under the Generations tab and display the originating node type.

Quality of life

• The new model banner on the canvas and jobs page now highlights GPT Image 2.5.

Fixes for video layer duration and balance calculations, alongside refinements to the compositor playback and timeline layout.

Video layer duration in the compositor

Video layers now default to their estimated duration when added to the compositor timeline. Previously, all generated layers stretched to the full timeline length by default. A 3-second video will now automatically span only 3 seconds on the timeline, giving you the correct duration immediately.

Quality of life

In the compositor:

• Playback now automatically rewinds to the start of the timeline upon reaching the end.
• The Invert Matte control uses a consistent Switch component.
• The timeline scroll container properly fits the number of visible rows.

Toast notifications now appear when generations complete or fail, and recordings save with the correct file extensions. The video compositor gains image-to-video and audio generation, while a new drag-snapping preference helps align nodes on the canvas.

Toast notifications for generation events

After a generation job finishes, a brief toast notification confirms the outcome. Successful runs show the model name followed by “completed,” while failures show the model name and the specific error. If you rename a layer, the toast uses your custom label instead of the model’s default name, keeping the feedback personalized to your workflow.

Correct file extensions for WebM recordings

Realtime video recordings now download with the correct .webm extension instead of the previous .mp4 mismatch. This resolves issues where players like QuickTime and VLC would refuse to open WebM-encoded files that were incorrectly labeled. All downloads, including videos, images, and audio previews, inherit this fix for accurate file typing.

AI generation in the compositor

The video compositor now supports generating images, videos, and audio directly within the editor. A new mode toggle filters available models by the type of media you want to create.

Generate videos using multi-shot editors for models that support sequential prompting, with individual prompts and durations per shot. Input slots for start frames, reference images, and audio layers integrate the LayerPicker for easy media selection. A new “Override Cut Format” toggle lets you choose between using the project’s dimensions or the model’s native resolution controls.

Snap nodes to alignment guides

A new “Snap to nearby nodes” preference in the Canvas settings (enabled by default) adds magnetic alignment when dragging nodes. As you move a node or a multi-node selection, light-blue guide lines appear when the selection aligns with a nearby node’s edges or center on either axis. The feature works with groups and respects grid snapping when the Shift key is held.

Quality of life

• Playhead on the video timeline turns red during playback.
• Time format display on the timeline is now a click-to-cycle button.
• Inspector controls use a unified height, with blend mode and matte channel selectors replacing native dropdowns.
• Generate modal features a redesigned model picker with provider icons and a scrollable list.
• Seed input in the Generate modal displays a dice icon and an “Auto” toggle pill.

Generate images directly in the Video Compositor without leaving the editor, export your full timeline to DaVinci Resolve, and enjoy a refreshed media lightbox with a new filmstrip.

Generate images in the Video Compositor

You can now generate images from inside the Video Compositor editor, keeping you in your creative flow. Open the editor, click “Generate with AI”, and choose from available image generation models. The resulting image appears as a new layer that you can composite, transform, and render alongside your other footage. Generated images save with your project and appear alongside your jobs.

Export timeline to DaVinci Resolve

Export your entire video composition to DaVinci Resolve for finishing work. The Export XML button now packs your timeline, source media, and import instructions into a downloadable ZIP file. Transforms, opacity, crop, and supported blend modes are preserved in the export. After downloading, extract the archive and import timeline.xml into Resolve—the app guides you through the relinking process to recreate your edit.

Refreshed media lightbox

The media lightbox now features a filmstrip navigation panel and updated controls. A thumbnail rail shows all items in your gallery, with keyboard navigation using arrow keys. The interface adopts Studio’s design language with a cleaner caption bar and intuitive control clusters. Video items autoplay muted with native controls, and zoom controls are more accessible.

Quality of life

• Multi-select frames on the canvas now display a connection pin with a count badge for batch operations.
• The “Selected N nodes” toast no longer appears over the Run Selected button when using Select All.

This release introduces wireless Get/Set nodes for cleaner workflows, generative video editing in the compositor, and a refreshed visual design with consistent corner radii and elevations across the app.

Get/Set Nodes

Introducing Get and Set nodes, which act as named wireless connections to help organize complex workflows. A Set node binds a name to any input value, and Get nodes placed anywhere on the canvas behave as if they are wired directly to that source. This lets you route a shared resource like a hero image to multiple consumers without creating visual spaghetti. Rename a Set to retarget all its connected Gets instantly. The system maintains referential integrity, preventing cycles and handling dangling references gracefully.

Generative Video Edits

The video compositor now supports prompt-driven generative edits. Right-click any video layer and select “Edit video” to open a popup for selecting a clip segment, choosing an edit model like Seedance 2.5 or Kling O3, and writing a prompt. The generated clip replaces the original segment in previews, exports, and node thumbnails. A new frame annotation tool lets you capture a reference frame from the clip and draw on it to guide the edit.

Output Settings Modal

Output settings in the video compositor have moved from the sidebar to a dedicated modal, accessible from the editor toolbar or the node parameter panel. The modal consolidates aspect ratio presets, which now lock the width and height to maintain the selected ratio until you switch to a custom setting.

Matte Layer Picker

The matte source selector in the compositor has been upgraded from a simple dropdown to a visual picker showing thumbnails, media types, and resolutions for each layer. This makes it easy to distinguish between layers, even when they share the same name.

Visual Refresh

The entire interface has been updated with a consistent design system for corner radii and elevations. Buttons, cards, panels, and menus now use role-based tokens, giving the app a more cohesive and polished appearance. Selected nodes on the canvas now feature a subtle accent border, ring, and glow that eases in and out smoothly.

Quality of life

• Hovering a Set or Get node highlights its bound pairs and adds a “Go to Set” button for quick navigation.
• The active icon in the left navigation rail now uses an inverted color scheme for better visibility.
• Routine actions like opening a project from the asset library or duplicating a workflow now navigate in the current tab instead of spawning new ones, reducing tab clutter.

This release adds pre-submit file size validation for the MiniMax H3 model and makes node error messages copyable.

Copyable node errors

A new “Copy error” button now appears under the error message when a canvas node fails. This makes it easy to share detailed error information when troubleshooting or reporting issues.

Quality of life

• MiniMax H3 now validates reference files before submission, showing a clear error if they exceed the 50 MB per-file limit
• The Reference Videos input tooltip now states the 50 MB per-file limit upfront
• Provider errors for file size limits now show actionable guidance to trim, downscale, or re-encode files

Collapsible side panels in the video compositor give you more space for your timeline, and a node lets you generate 3D splats for free.

Meshy 7 3D Generation

A new suite of Meshy 7 models is now available for 3D generation. You can create 3D assets from a single image, multiple images, or directly from text. The node includes parameters for different model types, including standard, low-poly, and smart-topology, giving you greater control over the final output’s style and complexity.

Free Image to 3D Splat

A new node allows you to generate Gaussian splats from images for free. The generated splat is automatically framed in the 3D viewer for immediate preview and can be used in scenes just like any other 3D assets.

This update improves refund handling for failed jobs and adds clearer error messages for content policy violations.

Error classification

Content policy rejections for some models now show as invalid input errors instead of generic failures. This makes it easier to distinguish between policy violations and technical issues. Timeouts during generation are also now properly categorized separately from other failures.

This release adds machine-to-machine authentication support.