UPDATES

Reframe

Change your shot's aspect ratio by generating beyond the original frame: portrait from landscape, widescreen from square, without a reshoot.

• Any aspect ratio. Seven presets from 9:21 tall to 21:9 ultrawide, to fit the screen or composition you have in mind
• 20 seconds. 4K. 10-bit color. Sub-second clips are supported, too
• Preserve shot. Your source stays completely intact, with the original performance, motion, and frame count preserved as the surrounding scene expands
• Expand freely. Choose where to position the source, how to scale it, and what to crop. Use a reference image to visually direct the expanded scene

To celebrate the launch, every Reframe generation is half price for two weeks. Reframe runs on SwitchFrame, our own video outpainting model: read the research for how it was built.

Open Reframe →

Effect Generator

Describe an effect and get a real node with controls built to match. It works on both images and video, and it's live in Canvas today.

• One prompt, any effect. Describe what you want and get a real effect with controls built to match: up to 16 sliders, colors, and handles you drag right on the preview
• It reads your depth and normal passes. Light, fog, and focus follow the actual scene, not a flat filter laid on top
• Regenerate without losing your work. Tweak the prompt and your control values carry over, and every effect is saved to your library for reuse on any canvas

Start from the starter canvas, with 20+ effects to duplicate and tweak, or make your own.

Open Canvas →

SwitchX 2.0

A major upgrade to our most advanced video-to-video model. Change anything in your video while preserving what matters — built from what you kept asking for: longer takes, 4K, color you can grade, camera motion that holds.

• 20s. Real 4K. 10-bit color. Takes up to 20 seconds, so you're no longer cutting shots to fit the tool. True 4K output, not an upscale, and 10-bit color that goes straight into your grade
• Camera tracking, built in. Pans, zooms, handheld — 2.0 reads the camera motion in your source and the new background follows it. No markers, no matchmove
• Fast motion holds. Whatever your mask covers stays untouched, now through fast motion — lip sync stays locked, quick movements stay clean
• Iterate in 720p, finish in 4K. Explore takes in 720p, then finish the one you like — the final pass is anchored to your original footage, so it stays consistent with the take you picked

To celebrate the launch, every SwitchX 2.0 generation is half price through September 28. For how the model was built, read the research.

Try SwitchX 2.0 →

MiniMax H3 Max

A new video model joins the Video Generator — Text and Image-to-Video at 480p or 768p, 24 fps, with audio included.

Generate with MiniMax H3 Max →

Improvements

• Canvas — SwitchX node now accepts image sources as well as video, matching the standalone SwitchX page
• Canvas — SwitchX node gains Auto Pilot: automatically regenerates its prompt when the reference image changes, plus a manual Suggest button
• Canvas — fixed Continue in Canvas occasionally triggering SwitchX's Auto Pilot unintentionally
• Canvas — Cmd/Ctrl+F finds nodes by name, with match cycling and highlighting
• Canvas — model picker gains a search field
• Canvas — SwitchX node hover overlay shows the Seed value
• Canvas — media badges consolidated into a single center-dot tag
• Canvas — fixed page zoom (Ctrl/Cmd + scroll, pinch) leaking into the browser's own zoom
• Canvas — fixed the zoomed-out view occasionally hiding a node's footer
• Canvas — fixed the SwitchX Infinity node's Unpack control not responding
• Canvas — fixed Continue in Canvas gallery Source/Alpha input tiles not rendering
• SwitchX — Seed moved out of Advanced into its own row
• SwitchX — job cards show the source's original resolution on thumbnail hover

Wan 3.0

Alibaba's newest video model joins the Video Generator — Text / Image / Reference-to-Video, up to 30-second single-shot clips with native audio, 480p to 1080p output.

Generate with Wan 3.0 →

Gemini Omni Flash 1.1

Google's multimodal video model gets a major upgrade — Text / Image / Reference-to-Video plus prompt-based editing, now with a real resolution picker from 360p up to 4K, generated audio throughout, and image-to-video end-frame support.

Try Gemini Omni Flash 1.1 →

New nodes: Rotoscope & Mixer

Rotoscope — layer-based manual rotoscoping. Draw shapes or paint strokes over your source and keyframe them over time, outputting a grayscale matte and the source with the matte applied.

Mixer — mask-driven key mix of two inputs. White in the mask shows the second input, black keeps the first; without a mask, Amount crossfades between them.

Try it in the Canvas →

Improvements

• Canvas — @ mention now auto-converts as you type or paste a matching reference name, not just when picked from the dropdown
• Canvas — Erode / Dilate node gains a Feather control and finer Size adjustment
• Canvas — Compare node gains Columns and Rows export
• Canvas — output gallery lets you compare different outputs from the same node
• Canvas — Image Generator and Video Generator nodes get inline prompt input
• Canvas — large canvases render faster when zoomed out
• Canvas — fixed trashed/purged canvases still being openable and silently autosaved by their owner
• Canvas — fixed applying a template not clearing undo history, so Undo could restore the previous graph

Canvas Despill

A new Canvas node removes green- and blue-screen spill in real time — no backend round trip, no credits. Feed it your keyed source, background, and mask, and it scrubs the color cast so composites read as believable footage.

Try it in the Canvas →

3D Scene Generator: GLB uploads

GLB uploads, and splat files dropped on the canvas now route straight to its upload panel.

Try the 3D Scene Generator →

Improvements

• Canvas — Video Generator's expand-dialog gallery gained Compare mode, matching Image Generator
• Canvas — new Audio Iterator node for batch-processing multiple audio files
• Canvas — Sequencer node gets a hover download button
• Canvas — node history now displays as an inline grid
• Canvas — repeatable input handles now add from one "+" at the bottom of the stack
• Canvas — the empty-canvas upload card shows format icons and a supported-formats legend
• Canvas — sticky notes gained a transparent-background option
• Canvas — fixed dropped newlines when editing text with @ mentions
• Canvas — fixed the zoom shortcut leaking into the browser's own zoom when an input was focused
• Canvas — text file drops are capped at 10MB, with a clear message instead of a silent failure
• SwitchX — history feed auto-loading restored, with tighter load-more skeleton spacing
• SwitchX — narrow history cards prioritize the preview across all products
• App — fixed a delay in credits reflecting after a completed purchase
• App — fixed the mobile sidebar tooltip popping out of place when opening

Audio references in the Video Generator

Seedance 2.0 and 2.5 Reference-to-Video now take audio references — hand the model an mp3 or wav and reference it in your prompt, the same way you already do with images and videos. Upload audio directly in the Video Generator, or add the new Audio Input node on the Canvas and wire it into a Video Generator node.

Generate with an audio reference →

Improvements

• Compare — reworked into four layouts, Wipe / Mix / Columns / Rows (Mix is new), shared across SwitchX, SDR to HDR, Background Remover, and Upscaler
• App — removed the top bar in favor of unified page headers, and reworked the mobile sidebar entry point
• Viewer playback — SwitchX, Canvas, Video Generator, and Upscaler players now start with sound and remember your last mute choice
• Image Generator — the expanded gallery gained a Compare mode with a source selector
• Video Generator — Generate Audio is now on by default for every Seedance model
• Canvas — drop a .md or .txt file onto the canvas to create a Text node from its contents
• Canvas — background notifications when an AI job finishes in another tab
• Canvas — "Continue in Canvas" now opens in a new tab, with less wait after clicking
• SwitchX — Create Reference: download the selected (or masked) frame at full quality

3D Scene Generator

3D Scene Generator is now live on Beeble. It generates full 3D scenes you can walk through, powered by World Labs.

• Marble 1.1 and 1.1 Plus, built in — generate a Gaussian splat scene in one click from a text prompt, an image, a video, or a panorama
• Intuitive 3D editor for shots — move through the scene freely, place up to six cameras, and capture at up to 4K
• Export anywhere — download the scene as a Gaussian splat and collider for Unreal, Blender, and beyond

It's also a dedicated node in the Canvas — no exporting, no bouncing between apps: connect it straight to SwitchX, or send a capture into the Image & Video Generator and start a Seedance 2.5 shot from it.

Try the 3D Scene Generator →

Seedance price drop

Every Seedance resolution now costs fewer credits per second — up to 50% less.

• Seedance 2.5: 480p 4 · 720p 10 · 1080p 20 credits / sec (was 7 / 15 / 34)
• Seedance 2.0 Standard: 480p 3 · 720p 6 · 1080p 13 · 4K 42 credits / sec (was 5 / 10 / 21 / 50)
• Seedance 2.0 Fast / Mini: now from 2 credits / sec

Full rates in the credit policy.

Generate with Seedance →

Seedance 2.5 — 1080p

Seedance 2.5 now renders at 1080p in the Video Generator, alongside the existing 480p and 720p options.

Improvements

• SwitchX — ProRes video sources are now ingested natively, with the upload limit for ProRes sources raised to 4.5GB
• SwitchX — the Showcase tab now leads with one large looping-video tile per category; hover to see the before still

MiniMax H3

MiniMax H3 joins the Video Generator — 2K / 24fps output with native stereo audio on every generation.

• Text / Image-to-Video — text-to-video picks from 6 aspect ratios; image-to-video follows your input image
• 5–15 second clips, 9 credits / sec — see the credit policy

Try MiniMax H3 →

Seedance 2.5 — Create / Edit / Extend

Seedance 2.5 now has Create / Edit / Extend task tabs in the Video Generator and on the Canvas — pick the tab that matches your prompt.

• Fixes edit-style prompts that used to fail on Seedance 2.5
• If a generation still fails, your credits are refunded automatically

Improvements

• Canvas — dragging a wire onto a node now shows the next input slot on nodes with expandable inputs; drop on it to create the slot and connect in one step
• Canvas — Connect to…: select several nodes, right-click, and connect them all to one target node in a single click
• Canvas — a selection box now picks up the wires it crosses, so you can delete them with one sweep
• Canvas — nine processing nodes gained the Expand button (Resize, Crop, Invert, Rotate / Flip, Curves, Blur, Erode / Dilate, Channels, and Sequencer), opening the full-quality result instead of the low-res proxy preview
• Mask editor — cursor-anchored zoom: the pixel under your cursor stays put while you zoom, instead of zooming from the image center (SwitchX and 3D Relight)
• Mask editor — brush strokes that leave the frame now follow the nearest edge pixel instead of being dropped, so masks reach all the way to the border
• SwitchX — a 30-second quick-guide video pops up the first time you open the reference creator, and a Guide button in the Reference Image section lets you rewatch it anytime