github invoke-ai/InvokeAI v7.0.0-alpha.4
InvokeAI 7.0.0 (alpha 4)
on GitHub

pre-release3 hours ago

Caution

⚠️ Please read this before you install

If you point v7 at your existing InvokeAI root directory, it will upgrade your database, and that database will no longer work with InvokeAI v6. The migrations only go one way, and v6 refuses to open a database that contains migrations it doesn't recognise.

To try v7, create a brand-new InvokeAI root directory and experiment there. In the Invoke Launcher , choose a new, empty install location. When prompted for the version, select Manual and enter version v7.0.0-alpha.4. For a manual install, pass --root /path/to/new/folder or set INVOKEAI_ROOT. Keep your v6 install and its root directory exactly where they are.

Upgraded your v6 database by accident? Don't panic. Invoke makes a backup before it migrates anything, and from this release on the startup log names that backup before the upgrade begins. See Restoring your v6 database at the bottom of these notes, or the new Upgrading to v7 guide.

Warning

This is alpha software. Things will break, some features are half-finished, and a few corners still have cobwebs. Please don't trust it with work you can't afford to lose. When you hit a bug, and you will, please open an issue on GitHub and put v7 in the title, or tell us about it on our Discord. Screenshots, logs and the output of the Diagnostics panel all help a lot.

Note

Alpha 4 adds a database migration, so once alpha 4 has opened your root directory, alpha 3 and earlier will refuse to start on it. Canvases that use the new Noise, Clarity, Texture or Dehaze adjustments, or the new type-first control layers, are saved in a newer canvas format that alpha 3 can't open. Please stay on alpha 4 once you upgrade.

What's New

Alpha 4 brings three new image model families, character and edit references for Anima, Krea-2 and Qwen-Image-2.1, new ControlNets, much better live previews, faster int8 models, Windows on ARM64, and the return of more features from the legacy interface. A lot of this release is also the result of your feedback on the earlier alphas, so thank you to everyone who reported bugs.

Major Features

Major new features include:

New model families: FLUX.3 Klein and MOSS-Image

  • FLUX.3 Image Klein from Black Forest Labs, in both the step-distilled Klein (16 steps, no CFG) and the undistilled Klein Base. It generates from text and edits or combines up to ten reference images, and Canvas regional guidance is placed as FLUX.3 layout boxes. Both checkpoints are under the FLUX Kommunity License. (@Pfannkuchensack)
  • MOSS-Image from OpenMOSS, in Base and SFT versions (Apache 2.0). It works in Generate, on the Canvas (img2img, inpaint, outpaint and regional guidance as layout elements) and in workflows, can decode with PiD, and comes with a prompt enhancer that expands a short prompt into a full caption or a layout draft. (@Pfannkuchensack)

📖 FLUX.3 · MOSS-Image

Reference images and edits

  • Anima character references. Anima now supports IP-Adapter character reference images. (@lstein)
  • Krea-2 edit mode. Krea-2 can take up to three reference images, in Generate, on the Canvas and in workflows. This is what makes ai-toolkit edit LoRAs (the pose and depth "ControlNets" people have been asking for) actually work. (@Pfannkuchensack)
  • Qwen-Image-2.1 edits from reference images, plus LoRA support for Qwen-Image-2.1. (@Pfannkuchensack)

📖 Anima · Krea-2 · Qwen-Image-2.1

New ControlNets and type-first control layers

  • FLUX.2 [dev] Fun-ControlNet-Union and the InstantX Qwen-Image ControlNet Union are supported, with Canvas control layers. (@Pfannkuchensack)
  • Canvas control layers are now type-first: you say what the image should control (Pose, Depth, Edges, Line art, Sketch, Color, Tile and so on) and Invoke picks the detector and the model, runs detection in the background, and keeps the original beside the detected map so you can re-run it. Gallery images and raster layers get Use as Control. (@joshistoast)

Better live previews with Tiny AutoEncoders

Set tiny_autoencoder_previews: true in invokeai.yaml and install the matching Tiny AutoEncoder from the starter models, and the in-progress preview becomes a close likeness of the final image instead of a blurry approximation. A preview takes about 25 ms on an RTX 4090. The finished image is unchanged. (@Pfannkuchensack)

📖 Tiny AutoEncoder Previews

Faster int8 models, and more single-file and GGUF formats

  • int8_compute. A new opt-in setting runs ComfyUI int8 convrot builds on the GPU's int8 tensor cores. On an RTX 4090 the denoising step of Z-Image Turbo, FLUX.1 dev and Krea-2 Turbo runs 1.3–2× faster, and up to almost 4× when the model only partly fits in VRAM. It needs an RTX 30-series or newer NVIDIA GPU and Triton, which is now installed automatically on Windows NVIDIA installs. Images differ slightly from the default at the same seed. (@Pfannkuchensack)
  • int8 convrot weights now stay int8 while they stream from system RAM, so partially loaded models move half as much data. (@Pfannkuchensack)
  • New fp8 and int8 single-file starter models. (@Pfannkuchensack)
  • New formats: Krea-2 ComfyUI int4 convrot transformers, Anima transformer GGUFs, LTX-2.5 transformer GGUFs and the Gemma-4 text encoder as GGUF. GGUF models are also identified much faster, from the file header alone. (@Pfannkuchensack)
  • attention_backend can now pin a PyTorch attention kernel (cudnn, flash, efficient or math), which is mostly useful for debugging. (@Pfannkuchensack)

📖 int8 Compute · Model Format Support

Upscaling

  • The Upscale widget now works with FLUX.2 and Qwen-Image too. (@Pfannkuchensack)
  • A new Model only mode runs an upscaler such as 4x UltraSharp by itself, with no denoise pass, and each image's menu has an Upscale submenu listing your installed upscalers at their native scale. (@joshistoast)

📖 Upscale

Canvas adjustments

  • Layers and groups get Noise (film grain, with mono, colour and chroma grain and a midtone response), Clarity, Texture and Dehaze adjustments. (@joshistoast)
  • The same Clarity, Texture and Contrast Dehaze are available as workflow nodes. (@JPPhoto)
  • Layers can be saved to a board of your choice. (@joshistoast)

📖 Image Detail and Contrast Nodes

Shared workflows

In multi-user mode, you can now make a Library workflow Public so that other users of the same Invoke server can open, duplicate or download it from the Shared tab, and call it from compatible workflow nodes. Only the owner can rename, delete or change it. (@JPPhoto)

Windows on ARM64

Invoke now installs and runs natively on Windows on ARM64, including NVIDIA's RTX Spark developer preview, using NVIDIA's ARM64 build of PyTorch. As part of this work, password hashing in multi-user mode now uses bcrypt directly instead of passlib. (@lstein)

Working memory that fits the model

Invoke now reserves working memory for FLUX.2, FLUX.3, Z-Image and Krea-2 denoising, and for VAE decoding, from measurements of each model rather than from rough estimates. That means fewer out-of-memory errors without holding back more VRAM than needed. Administrators can also calibrate a model's VAE working memory from the Model Manager. (@Pfannkuchensack)

More old favourites, and lots of polish from your feedback

  • Hosted-provider options, Canvas modes and model record editing are back from the legacy interface, as are PBR map generation and image storage administration in the Gallery. (@joshistoast)
  • Undo instead of confirm. Moves, removals, archives, stars, preset changes and layout resets happen straight away and offer Undo. Only permanent or bulk actions still ask first. The Invoke button opens a card that shows the source, the destination and anything blocking the run, along with its fix. The Queue gets a Cache popover for the model and invocation caches. (@joshistoast)
  • Gallery and viewer: an Auto-switch menu (Newest / In progress / Off), previous and next arrows, search across All boards, and bulk delete for wildcards. Shift+click Invoke queues at the front. (@joshistoast)
  • Workflow editor: fixes for paste, board pickers, library sorting and templates. (@joshistoast)
  • External clients that enqueue with destination: "gallery:reveal" can now reveal videos as well as images. (@lstein)
  • Many more Canvas fixes for frames, layer actions and the canvas engine. (@joshistoast)

Major Bug Fixes

Notable bug fixes and performance improvements include:

  • Model downloads that are retried or restarted now fetch every missing file, so a failed download no longer leaves a model without its config.json or chat template, and partial downloads only resume when the server file hasn't changed (@joshistoast, @lorenzozanee)
  • Projects opened over the LAN or plain HTTP now show their images, and copy and paste explain when the browser needs HTTPS (@joshistoast)
  • Models now install and start on drives without hard links (@Pfannkuchensack)
  • SD 1.x, SD 2.x and SDXL GGUF files, which Invoke can't run, are now refused at install time (@Pfannkuchensack)
  • Mistral GGUF text encoders now load correctly (@Pfannkuchensack)
  • Z-Image only accepts its Qwen3 4B text encoder, Z-Image GGUFs with 2-D pad tokens load, and Z-Image Fun ControlNets run as trained (@joshistoast, @Pfannkuchensack)
  • Krea-2 Raw, Qwen-Image and Qwen-Image-2.1 Base use the correct schedule shift at large sizes, and FLUX.2 and Z-Image schedulers follow the node's sigma schedule (@Pfannkuchensack)
  • PiD now gets the normalized latent it was trained on (@Pfannkuchensack)
  • Many LoRA fixes: DoRA output scaling, conv-shaped LoHa, factorized LyCORIS, FLUX PEFT alpha values, FLUX guidance-embedder and final-layer modulation layers, diffusers FLUX LoRAs with text-encoder layers, and Anima kohya and diffusers LoRAs (@Saurav-Gupta-9741, @guishiw, @kevin9327, @chiliec, @Pfannkuchensack)
  • Wan video decoding only tiles when tiling actually saves memory (@Pfannkuchensack)
  • ERNIE-Image honours a partial denoise window with the Heun scheduler (@horizonbymuneeb)
  • MiniMax H3 now runs on Apple MPS (@joshistoast)
  • An enqueue that keeps failing with a 500 error is no longer retried forever, and the queue shows the full reason when an install is refused (@kevin9327, @tunglambk)
  • The Get Image Mask Bounding Box node returns an exclusive x_max/y_max, and String Split splits on whitespace when the delimiter is blank (@MohammadHijjawi97)
  • Translation updates now reach your browser without a hard refresh (@Pfannkuchensack)
  • Changing your password in a legacy-interface tab no longer logs out your other tabs (@wunianze666-netizen)
  • Image search vocabulary requests no longer block, and old vocabulary caches are cleaned up (@wunianze666-netizen)
  • Security updates for dependencies flagged by Dependabot, and diffusers 0.42.1 (@lstein, @Pfannkuchensack)

The full list of updates follows.

What's Changed

New Contributors

Full Changelog: v7.0.0-alpha.3...v7.0.0-alpha.4

🛟 Addendum: Restoring your v6 database

If you ran v7 against your existing root directory and now want to go back to v6, you can restore the backup that Invoke made just before it upgraded your database.

What you'll lose: anything you did while running v7. Images and videos you generated in v7 stay on disk in your outputs folder, but they won't appear in the v6 gallery. Models you installed in v7 stay in your models folder but won't be listed in the Model Manager until you install them again (the Scan Folder option makes this quick).

  1. Quit InvokeAI completely. Make sure no Invoke process is still running.

  2. Open the databases folder inside your InvokeAI root directory. (If you set db_dir in invokeai.yaml, look there instead.)

  3. Find the backup. You'll see files named like this: 

    invokeai.db
invokeai_backup_20261005-143012.db

    The timestamp is YYYYMMDD-HHMMSS. Choose the backup made when you first launched v7: it holds your database exactly as v6 left it. Earlier upgrades may have left older backups in the folder too, so pick by date. If you aren't sure which one it is, it's usually the oldest backup created on or after the day you installed v7.

  4. Move the upgraded database out of the way. Don't delete it yet. Create a folder such as v7-db and move these files into it (the last two may not exist, which is fine): 

    invokeai.db
invokeai.db-wal
invokeai.db-shm

    [!IMPORTANT]
    Don't skip the -wal and -shm files. If they're left behind, SQLite may try to apply v7's leftover changes to the restored database.

  5. Copy the backup into place and rename the copy to invokeai.db. Copy rather than move, so the original backup stays as a safety net.

    • Windows (PowerShell): Copy-Item invokeai_backup_20261005-143012.db invokeai.db
    • macOS / Linux: cp invokeai_backup_20261005-143012.db invokeai.db

  6. Check invokeai.yaml. v7 adds a few new settings (db_url, db_synchronous, fp8_compute, reserve_vram_gb, int8_compute, attention_backend, tiny_autoencoder_previews, prompt_files_dir, image_index_*, fonts_*). v6 won't start if any of these are in the file, so remove any you added.

  7. Start InvokeAI v6 (6.14.2). Your boards, images and models should be back where you left them.

The Upgrading to v7 guide covers the same steps in more detail, including how to go back to v7 afterwards.

If v6 still refuses to start, or you can't find a backup, ask in the Discord or open an issue before you try anything else, and keep every .db file you have.

Check out latest releases or
releases around invoke-ai/InvokeAI v7.0.0-alpha.4

Don't miss a new InvokeAI release

NewReleases is sending notifications on new releases.

Get notifications