What's Changed

New IC-LoRAs

• AlphaGen: cut a subject out of a video with no green screen and no prompt. You get a transparent cutout (WebM) and a matte (MP4). Both files line up with your clip.
• Motion Track Control: draw up to three paths on a start frame to move objects or the camera. Set the pace and see a live preview.
• Video Guide: direct a generation with a guide video. Choose Pose, Depth, or Canny edges. A start frame is optional.
• Ingredients: combine up to 10 reference images into one video. It is in a new Marketing section.
• IC-LoRA tools keep every frame of your clip. The Source audio mode copies the original audio.

1440p and 4K

• Text to Video, Image to Video, Audio to Video, and the prompt and start-frame LoRAs offer 1440p and 4K.
• 1440p clips go up to 20 seconds. 4K clips go up to 10 seconds. 4K does not offer 21:9.
• Long clips run in overlapping time windows. The pickers tell you when this happens.
• A 20-second 1080p clip now runs on 32 GB graphics cards.

Assistant

• The new Assistant plans and queues jobs from plain words. It works on Desktop and on your phone.
• It can make videos, apply styles and effects, run Retake and Extend, find assets, and extract a frame.
• Use Ask assistant on an asset tile to attach it to a chat.
• Choose Gemini (API key) or a local model (download from the Models tab).

Prompt enhancement

• Enhance now looks at your source video for IC-LoRAs, Extend, and Retake.
• Extend and Retake write a prompt from the video when the prompt is empty and auto-enhance is on.
• Enhance works with an empty prompt when a video or start frame is present.
• Enhance with Gemini can run during a generation.

Queue and Support

• Every Queue row opens its generation on the feature page.
• The diagnostic logs viewer has a refreshed design that matches the rest of the app.

Requirements

• Windows: graphics cards with 9 GB or more of VRAM and fp8 support can run local generation with smaller settings.
• macOS: the app needs macOS 14 or later. Local generation needs macOS 26 or later.

Privacy

• Anonymous analytics now includes the OS version (and the distribution on Linux).

Bug Fixes

• Windows: a short memory refusal during decode no longer stops the job.
• Mac: long 1080p text-to-video clips no longer fail.
• Linux: large models no longer fail with "Cannot allocate memory" on default memory settings.
• Retake and Extend on LTX 2.3 no longer fail when LTX 2.5 is the active model.
• Mac: with a right-to-left system language, the window buttons no longer cover Contact Sales.

Full Changelog: v1.3.0...v1.3.1

What's Changed

A new look for the entire app

• LTX Desktop has a new look across the whole app, with surfaces that match LTX.io.
• The new Home page has a cinematic hero, a workflows gallery, a LoRAs row, and example media for each tool.
• The new sidebar has Home, Assets, and Remote, with Featured tools, Recent workflows, and Recent projects.
• Quick search (the / shortcut, the sidebar, or the Home search bar) finds any tool, with filters and previews.
• Home, the Video Editor, Settings, Keyboard Shortcuts, and all dialogs now share one design system and one light or dark theme.
• First-run is now one Setup screen for Hugging Face, API keys, and model downloads. It has an "Opt out of basic analytics" checkbox.
• A new splash screen plays the LTX-2 end-page animation while the Python environment installs.
• New back and forward navigation arrows in the header move you through the screens you visited.

New tools

• Text to Video, Image to Video, and Audio to Video run locally, with a live results feed.
• Audio to Video can record from the microphone on Windows and Mac, and trim long audio. It generates at 540p, 720p, and 1080p.
• Retake and Extend accept a source of any length, have an optional Trim button, and blend the edited section into the original clip.
• Pick a source video from the asset library when you use Retake, Extend, or an IC-LoRA.
• Open a tool from a link with ltxdesktop://fetcher/{tool}.

New LoRAs and IC-LoRAs

• New IC-LoRAs: Day to Night, Layout to Render, Restore, Deblur, Colorization, Clean Plate, Decompression, and Water Simulation.
• New LoRA recipes, including Cozy Felt and Dolly In. Download a LoRA from the Generate button, or from Settings → Models.
• Each LoRA now runs on the checkpoint it was trained for.

Queue and Assets

• A global generation queue shows pending, running, Done, and Failed jobs. Drag a job to change its order.
• A new Assets library shows your uploaded media and generated videos.
• Your generations and assets now persist across restarts.

Remote

• Text to Video, Image to Video, Audio to Video, Retake, Extend, the LoRAs, and the Assets library now work from the Remote app in any browser on the same Wi-Fi or local network (LAN), such as your phone.
• Pairing is secure. The QR code gives a one-time grant, and the Remote page lists paired devices that you can revoke.

More control over generation

• Local video now offers 270p and 360p, clips of 2 to 5 seconds, 25, 48, and 50 fps, and new aspect ratios: 21:9, 3:2, 4:3, 1:1, and 4:5. The model trained on 16:9 and 9:16, so other ratios can give lower quality.
• Prompts are now enhanced automatically before local video generation. Turn this off with the auto-enhance switch, or use the Enhance button to enhance a prompt yourself.
• Unlocked seeds are random. Every generate form has a seed control, and the seed is the same everywhere in the app.
• The progress bar fills from the real denoise steps.
• Local generation chooses full or streaming model loading for each job, and hides options that your machine cannot run.
• A warning shows when 720p or higher is likely to be slow on a Mac.
• A retake selection is limited to 10 seconds.

Settings and updates

• Settings has a clearer layout. The app now checks LTX, Fal, and Gemini API keys before it saves them, and you can remove a saved key.
• The model upgrade dialog and the app update dialog now match the design system.

Engine

• The app now uses ltx-core and ltx-pipelines 1.4.1, and the LTX API integration uses API v2.

Bug Fixes

• Mac: LoRA runs no longer decode to green or gray frames.
• Windows: IC-LoRA and Retake at 720p no longer use a very large amount of GPU memory on 32 GB cards.
• Linux: a 32 GB machine no longer runs out of memory when it loads CUDA weights.
• Mac: machines with 32 GB of total RAM are now correctly allowed to generate locally, and streamed model weights load faster.
• Photos from phones, especially Android, no longer appear rotated by 90 degrees.

Full Changelog: v1.2.7...v1.3.0

Bug Fixes

• Mac: local LTX 2.5 Fast no longer spikes memory into the tens of gigabytes and freezes during denoise on Apple Silicon.
• AMD (ROCm): local generation no longer treats the GPU as NVIDIA CUDA, so it skips FP8 and no longer runs out of VRAM when loading the full transformer on 32GB-class cards.

Full Changelog: v1.2.6...v1.2.7

What's Changed

• GenSpace video: optional last-frame image when a first frame is set (not available with Auto duration). Works for local and API generation.
• GenSpace prompt bar is resizable; the height is remembered.
• Enhance (API): choosing it without a valid Gemini key opens Settings → API Keys. The default Gemini model is Gemini 3.5 Flash-Lite.

Bug Fixes

• Mac: local Z-Image no longer freezes at the start of denoise.
• Windows 16GB GPUs: first local Fast generate no longer fails with a paging-file / HTTP 500 error when loading the checkpoint.
• First-run Choose Location shows real free disk space.

Full Changelog: v1.2.5...v1.2.6

What's Changed

• Local LTX 2.5: Retake and Extend work on-device. Cloud 2.5 Fast / Pro still do not offer retake or extend.
• GenSpace: 2160p is labeled 4K. API offerings include 720p. LTX-2.5 Pro (API) is 720p and 1080p only (no 48 fps, no 1440p/4K). Audio-to-video resolution and duration caps match the API.
• Video Editor gap fill can generate 2–5s clips for holes shorter than 6s. The GenSpace duration picker in API mode still starts at 6s.
• Finished GenSpace cards keep the model and resolution from when you clicked Generate, even if you change the picker while the job is running.

Bug Fixes

• Windows streaming: local generation on 16GB GPUs no longer fails with a fake CUDA out-of-memory error when the OS cannot pin host RAM.

Full Changelog: v1.2.4...v1.2.5

What's Changed

• Updates are opt-in. When a new version is found, a prompt offers Update now, Later (this session), or Skip this version. Nothing downloads until you choose Update now, and quitting never installs. Settings → About shows the same state, plus a manual check and an auto-check toggle. Restart-to-update is blocked while a generation is running.
• Stop local generations. Stop appears for on-device GPU jobs (video, image, IC-LoRA, retake, extend) and takes effect at the next generation step without unloading models. Cloud jobs (LTX / fal API) still run to completion — there is no cancel API for those.
• GenSpace gallery: filter All / Videos / Images, and sort by Date, Type, Duration, Resolution, or Ratio. Favorites still combines with the type filter.
• Enhance (Gemini API): choose the Gemini model next to your API key in Settings → API Keys. Default is now Gemini 2.5 Flash-Lite — the previous 2.5 Flash id 404s on new API keys. Local Enhance is unchanged (on-device Gemma).

Bug Fixes

• Local Fast: text + image + audio generation no longer crashes
• Downloading LTX 2.3 models no longer 404s on the spatial upscaler (Hugging Face removed the old file)

Full Changelog: v1.2.0...v1.2.4

What's Changed

• Local LTX 2.5 Fast generation
• LTX 2.5 Fast and Pro on the API generation path
• Hugging Face sign-in for gated model downloads

Full Changelog: v1.1.0...v1.2.0

What's Changed

• Apple Silicon Macs can now generate locally (tested on M4 Pro / 48GB — treat lower configs as a bonus, not a guarantee). Intel Macs stay API-only (no MPS backend)
• Built-in LoRA / IC-LoRA library — browse and download from inside the app, per-adapter strength slider, or drop your own .safetensors into loras/ (local generation only)
• Enhance: rewrites your prompt before generation (video, image, IC-LoRA). Picks up the selected LoRA/IC-LoRA's metadata, example prompts, and trigger words automatically. Runs on the local Gemma text
  encoder or via Gemini API
• Video extend: grow a generated clip forward or backward, locally or via API
• Image editing (Z-Image-Turbo): image-to-image editing alongside existing text-to-image, locally or via fal API
• Base model manager in Settings: download, switch, and delete LTX model versions; option to keep the previous checkpoint after an upgrade
• First-run setup now shows exact download sizes per file and a total
• Experimental "Diffusion Stage Cache" (Settings, NVIDIA/CUDA only, off by default) — can speed up fast generations by skipping redundant reload work

Bug Fixes

• Leaving a project mid-generation and returning no longer loses the result
• Fixed visible artifacts/glitches on upscaled videos
• Fixed a file-path bug that stopped some local video/image files from loading

Mac note

The app checks for ~15GB free RAM (not total) at launch. If a capable Apple Silicon Mac shows API-only mode, close memory-heavy apps (browsers are the usual culprit), then quit and relaunch — freeing RAM while the app stays open won't retrigger the check.

Full Changelog: v1.0.5...v1.1.0

Bug Fixes

• macOS: App didn't function for users without ffmpeg pre-installed
• Performance: The <16GB VRAM optimization introduced in a previous version caused a performance regression for users with high-VRAM GPUs — the optimization is now only active when VRAM is below 32GB
• Video Editor: Regression in gap fill generations causing generations to error
• Video Editor: Regression from previous version — pure audio tracks could not be dragged and dropped
• Video Editor: Video assets could be dropped onto audio tracks
• Video Editor: Panel resizing was unresponsive on first attempt when entering the editor
• Video Editor: Gap fill generation properties (resolution, fps, duration) were incorrect and out of sync with Gen Space properties
• Video Editor: Text clip properties panel incorrectly showed video clip properties
• Video Editor: Source monitor controls were missing a loop button
• Video Editor: Asset panel lasso was unresponsive to scroll when dragging out of panel bounds
• Video Editor: Could not start a lasso selection from a gap fill area in the timeline
• Video Editor: Custom asset bin creation didn't work
• Settings: Text encoder download showed a misleading UI — replaced with an accurate progress bar
• Settings: Removed configurations that were not connected to any functionality
• First Launch: LTX API key set during first launch was not reflected in the UI without a restart
• LTX API: Insufficient funds error now shows a dedicated message with a button linking to the LTX API console to purchase credits
• A2V: Local generations were unnecessarily restricted to landscape aspect ratio and specific resolutions — restrictions removed
• Backend: In specific failure cases, a launch error could appear with no message and a non-functional retry button

Improvements

• App version is now logged at startup in reported log files, aiding bug investigation
• Added volume control for video assets directly from the Gen Space asset thumbnails

Full Changelog: v1.0.4...v1.0.5

Full Changelog: v1.0.3...v1.0.4