Repository navigation
InvokeAI 7.0.0 (beta 2)
Pre-releaseCaution
⚠️ 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-beta.2. 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 the startup log names that backup before the upgrade begins. See Restoring your v6 database at the bottom of these notes, or the Upgrading to v7 guide.
Warning
This is beta software. Please keep backups of work you care about. When you hit a bug, 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
Coming from beta 1 or an alpha? You can install beta 2 over beta 1 (or alpha 4) and keep using the same root directory. Beta 2 adds one small database migration (it stores the FLUX.3 prompt upsampler's system prompt), so once beta 2 has opened your root directory, beta 1 and the alphas won't start on it.
InvokeAI 7.0.0 (beta 2)
Thank you to everyone who tried beta 1 and told us what broke. Beta 2 is mostly a bug-fix release: about 65 pull requests, most of them fixes for memory use, model loading, LoRAs, recall and the queue. There are also some new features, led by vector layers on the Canvas and a redesigned workflow library.
New to v7? The beta 1 release notes have the full tour of what's new in InvokeAI 7 (projects, the workbench, the new canvas, video, the gallery and more), and the documentation lives at v7.invoke.ai.
Contents: Bug fixes · New in beta 2 · Contributors · Restoring your v6 database · Full changelog
🐛 Bug fixes
🧠 Memory and VRAM
- A failed generation gives its VRAM back. When a generation ran out of memory, the GPU memory it had been using stayed reserved until you cancelled something. Now any session that ends early (cancelled, failed, or stopped by an error) releases it. Empty Cache tells you which models it dropped, and says so when the GPU driver refuses to release memory instead of failing silently. (@Pfannkuchensack, #10267)
- LoRAs no longer fragment VRAM. Patching a large LoRA could leave around 5 GB of VRAM that the model cache thought was free but wasn't. On Windows, the VAE decode after it could then spill into shared memory and take over a minute instead of a fraction of a second. (@Pfannkuchensack, #10272)
- Wan 2.2 uses much less system RAM. Wan was holding both experts in RAM during expert swaps, pinning whole Diffusers shards, and reading entire single-file checkpoints before deciding what to keep. All three are fixed, and Wan now stays within the model cache's RAM budget. (@Pfannkuchensack, #10192)
- Qwen-Image-2.1 VAE decodes on ROCm reserve the memory they actually need, measured on AMD and NVIDIA cards, so large decodes no longer run out of memory on a busy card. (@lstein, #10043)
- VAEs set to run on CPU now really run there, including the FLUX Fill and Control LoRA paths. (@wunianze666-netizen, #10145)
- Diffusers and GGUF T5 encoders run in float32 when on the CPU, avoiding slow or broken half-precision CPU paths. (@Pfannkuchensack, #10216)
- GGUF files are probed without reading the whole file. (@Pfannkuchensack, #10301)
📦 Model loading
- ComfyUI
fp8_scaledcheckpoints stay packed. Wan checkpoints (#10262) and Qwen3 and Qwen2.5-VL text encoders (#10291) were being expanded to bf16 when loaded, doubling their memory. (@Pfannkuchensack) - Wan GGUF transformers with BF16 and embedder tensors unpack correctly. (@Pfannkuchensack, #10256)
- One model file, more than one job. A whole Qwen3.5 9B vision-language file now installs as a Text LLM and serves as the MOSS-Image Qwen3.5 encoder, whichever way you install it. Likewise, a full Qwen3 folder (e.g.
Qwen/Qwen3-4B) also serves as the Qwen3 encoder of its size for Anima, Z-Image and FLUX.2 Klein. (@Pfannkuchensack, #10221) - Related models now matter. Models you link in the Model Manager are listed first in the VAE, encoder, LoRA and reference-image pickers for the active main model. (@Pfannkuchensack, #10246)
- PatchMatch works offline. Invoke now requires
pypatchmatch2.2, which ships its native library in the wheel instead of downloading it from GitHub on first import. (@Pfannkuchensack, #10298)
🧬 LoRAs
- DoRA magnitudes now apply on Z-Image, Qwen-Image and Anima, and on quantised transformers. (@Pfannkuchensack, #10245)
- FLUX kohya and OneTrainer LoRAs that train layers outside the transformer blocks are identified and converted. (@Pfannkuchensack, #10206)
- Qwen-Image LyCORIS LoRAs, and LoKr and LoHa files with dotted key names, are identified correctly. (@Pfannkuchensack, #10189)
- Diffusers LoRAs made for other DiT models are no longer mistaken for Anima LoRAs. (@Pfannkuchensack, #10223)
- LoRA sub-patches stay in place during sidecar forward passes. (@Pfannkuchensack, #10243)
🔁 Generation and recall
- Recall clears reference images when the recalled image had none, instead of leaving stale ones in place. (@lstein, #10284)
- SD 3 text encoders are recorded in metadata and recalled. (@renningg, #10228)
- FLUX Kontext reference images survive being renamed. (@renningg, #10238)
- The FLUX.2 VAE starter model has the right description. (@renningg, #10236)
🧩 Workflows
Ifnodes run their branches eagerly by default, which fixes graphs that combineIfwith iteration and loops (#9918). Saved execution state moves to version 3, and older snapshots still load. (@JPPhoto, #10032)- Legacy workflow child rows load correctly. (@JPPhoto, #10242)
⏱️ Queue, projects and saving
- Queue submissions recover from flaky connections. A submission whose acknowledgement got lost is now reconciled instead of being doubled or dropped, and missed completion events are caught up every 15 seconds. (@joshistoast, #10162)
- Edits made just before closing a tab are kept, including in Safari and other WebKit browsers. An older save can no longer overwrite a newer recovery journal. (@joshistoast, #10165)
- iOS Safari can list recovery projects again. (@lstein, #10229)
- The queue duration label no longer shows values like "1m 60s". (@kevin9327, #10086)
🖼️ Gallery
- Board downloads no longer fail with "Response content longer than Content-Length". The zip is only served once it has been fully written. (@Pfannkuchensack, #10270)
- Very long board names no longer break bulk download archive names. (@Pfannkuchensack, #10279)
🔒 Security
- Custom nodes can no longer tamper with the server. Core API routes and the web UI mount are captured before custom nodes load, so a node can't replace them or register its own HTTP endpoints. (@JPPhoto, #10211)
- The Image Map's search-by-URL fetches images through the SSRF-guarded session, which ignores proxy environment variables and blocks NAT64 addresses. (@Laihu08, #10264)
⌨️ Interface
✨ New in beta 2
✏️ Vector layers and spline editing
The Canvas gets vector layers. Draw rectangles, ellipses, stars, polygons and freehand shapes into a vector layer, then reshape them with the new Spline Edit tool (S): drag control points and Bézier handles, add points, switch between corner, smooth and symmetric points, split paths and join them back up. Each layer has its own fill, stroke, stroke width and taper settings. When you're ready, Trace or Fill a path or the whole layer into a new raster layer, and the vector paths stay editable. (@DustyShoe, #10281)
The Transform tool also gets resampling controls (Bilinear, Bicubic, Hamming, Lanczos) and Contain / Cover / Fill options for fitting a layer to the Frame. (@DustyShoe, #10152)
🗂️ A redesigned workflow library
Finding a workflow is much easier now. The library has a sidebar with In this project, Recent, My library, Templates and Shared with me, each with a count, plus tag filters. You can sort by last run, and switch between a grid of smaller cards and a list. Each workflow shows a readiness dot that tells you whether it's ready to run, needs models installed or is missing node packs. The previews and the nodes in the editor have been redesigned to match. (@joshistoast, #10222)
In the editor:
- Fit selection in the graph's context menu frames the nodes you've selected. (@renningg, #10225)
- Shift+Z moves keyboard focus between the graph and the workflow form. (@renningg, #10226)
🪄 FLUX.3 prompt upsampler in Generate
FLUX.3 generations can now run the prompt upsampler themselves. Turn on Prompt upsampler in the Render section and pick the upsampler model under Components. It rewrites your prompt into the kind of caption FLUX.3 was trained on, and for edits it sees your reference images too. (@Pfannkuchensack, #10219)
📦 More model formats
- torchao fp8 builds from unsloth for FLUX.3 (#10182), Qwen-Image-2.1 (#10179) and MOSS-Image (#10181). (@Pfannkuchensack)
- SD 3.5 transformers from GGUF builds. (@Pfannkuchensack, #9972)
- Krea-2 edit references with a GGUF encoder. The Krea-2 loader takes a Qwen3-VL
mmprojvision tower beside a GGUF Qwen3-VL 4B encoder, so pose and depth edit LoRAs work without the full encoder. (@Pfannkuchensack, #10244) - MOSS-Image LoRAs. (@Pfannkuchensack, #10271)
- MOSS-Image in the Upscale widget. (@FrankBro, #10257)
👀 Tiny AutoEncoder previews on the CPU
On small GPUs, Tiny AutoEncoder previews used to fall back to the plain linear preview whenever the decoder didn't fit beside the running model. Each preview decoder now has a Decode previews on CPU switch in the Model Manager. CPU previews are capped at 512 px and update less often, but they need no VRAM. (@Pfannkuchensack, #10231)
🎬 MiniMax H3 hybrid checkpoints
For H3 checkpoints that support both FL2VA and Ref2VA, you can now choose which task to use per generation from the Video panel, without editing the model. (@jaffer1979, #10141)
🎛️ Generate and the workbench
- Update a preset. A floppy-disk button beside Save preset overwrites the loaded preset with your current settings, with Undo. (@lstein, #10285)
- Apply model defaults. The Generate header's reset becomes a Reset settings menu. Apply model defaults changes only what the model's own defaults define, as in v6. (@joshistoast, #10136)
- Recall from the Queue. Queue entries offer Recall all, Remix, prompt, seed, size and CLIP skip, and A recalls everything from the focused row. (@joshistoast, #10136)
- Drop files from your desktop onto the Canvas drop zones or Preview. They upload to the current board first. (@joshistoast, #10136)
- Plus a broad polish pass over forms, menus, the top and status bars, Gallery, Layers, Upscale and Projects, and the command palette now runs on Chakra UI 3.38. (@joshistoast, #10136, #10247)
🔗 Metadata from every denoise node
All 19 image and video denoise nodes now take a metadata input and return a metadata output, so workflows record their generation parameters without the old model-specific *_meta wrapper nodes. IP-Adapters used in workflows now recall as v7 reference images. (@skunkworxdark, #10083)
💜 Contributors
Thank you to everyone who wrote code for this release:
Core development team
![]() @Pfannkuchensack |
![]() @joshistoast |
![]() @lstein |
![]() @JPPhoto |
![]() @DustyShoe |
Community contributors
![]() @renningg |
![]() @skunkworxdark |
![]() @jaffer1979 |
![]() @FrankBro |
|
![]() @wunianze666-netizen |
![]() @Laihu08 |
![]() @kevin9327 |
![]() @jadelike-wine |
![]() @deepratna-awale |
And an equally big thank-you to everyone who tested beta 1 and reported what was broken.
🛟 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).
-
Quit InvokeAI completely. Make sure no Invoke process is still running.
-
Open the
databasesfolder inside your InvokeAI root directory. (If you setdb_dirininvokeai.yaml, look there instead.) -
Find the backup. The startup log names the backup when v7 upgrades your database. You'll see files named like this:
invokeai.db invokeai_backup_20261005-143012.dbThe timestamp is
YYYYMMDD-HHMMSS. Choose the backup made when you first launched v7: it holds your database exactly as v6 left it. Later upgrades (including beta 2's) leave newer 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. -
Move the upgraded database out of the way. Don't delete it yet. Create a folder such as
v7-dband 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-waland-shmfiles. If they're left behind, SQLite may try to apply v7's leftover changes to the restored database. -
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
- Windows (PowerShell):
-
Check
invokeai.yaml. v7 adds new settings (for exampledb_synchronous,fp8_compute,int8_compute,attention_backend,tiny_autoencoder_previews,reserve_vram_gb,image_index_*andfonts_*). v6 won't start if any of these are in the file, so remove any you added. -
Start InvokeAI v6 (6.14.4). Your boards, images and models should be back where you left them.
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.
What's Changed Since Beta.1
- feat(shell): replace the alpha-build notice with a beta-build notice by @lstein in #10209
- chore(release): bump version by @lstein in #10210
- test(workflows): wait for the library card before selecting it by @lstein in #10212
- chore(performance): re-record build and browser baselines by @lstein in #10213
- fix(frontend): preserve edits and validate browser lifecycles by @joshistoast in #10165
- fix(queue): recover ambiguous admission and delivery acknowledgements by @joshistoast in #10162
- fix(queue): carry a rounded duration into the next unit by @kevin9327 (#10086)
- feat(workflows): fit selected nodes in the editor by @renningg in #10225
- fix(api): harden core ASGI routes against custom-node mutation by @JPPhoto in #10211
- fix(workflows): drop unused child policies and read legacy child rows by @JPPhoto in #10242
- fix(qwen-image): identify LyCORIS LoRA layouts and group dotted LoKr and LoHa keys by @Pfannkuchensack in #10189
- fix(queue): list recovery projects without nextunique cursors on iOS Safari by @lstein in #10229
- fix(wan): keep system RAM within the model cache budget by @Pfannkuchensack in #10192
- fix(model-manager): run Diffusers and GGUF T5 encoders in float32 on CPU by @Pfannkuchensack in #10216
- feat(flux3): run the prompt upsampler inside FLUX.3 generations by @Pfannkuchensack in #10219
- test(recall): fail when a recorded metadata key has no recall contract by @Pfannkuchensack in #10220
- fix(lora): keep wrapper sub-patches in place during sidecar forwards by @Pfannkuchensack in #10243
- fix(lora): apply DoRA magnitudes on Z-Image, Qwen-Image and Anima and on quantized transformers by @Pfannkuchensack in #10245
- feat(krea-2): read edit references with a GGUF encoder through the mmproj vision tower by @Pfannkuchensack in #10244
- feat(model-manager): load unsloth's torchao fp8 Qwen-Image-2.1 builds by @Pfannkuchensack in #10179
- fix(models): rank related models by the active main model in pickers by @Pfannkuchensack in #10246
- fix(wan): unpack BF16 and embedder tensors when loading GGUF transformers by @Pfannkuchensack in #10256
- feat(flux3): load unsloth's diffusers-named torchao fp8 build by @Pfannkuchensack in #10182
- feat(previews): decode Tiny AutoEncoder previews on the CPU by @Pfannkuchensack in #10231
- fix(lora): identify and convert FLUX kohya and OneTrainer layers outside the blocks by @Pfannkuchensack in #10206
- feat(moss-image): support LoRAs by @Pfannkuchensack in #10271
- fix(lora): stop direct patching from fragmenting VRAM by @Pfannkuchensack in #10272
- fix(gallery): serve bulk downloads only once the archive is complete by @Pfannkuchensack in #10270
- docs: rename AutoExpress to Emotica in community nodes by @deepratna-awale in #10277
- fix(models): correct the FLUX.2 VAE starter description by @renningg in #10236
- chore(agents): enhance agents environment by @joshistoast in #10278
- fix(gallery): cap the board name in bulk download archive names by @Pfannkuchensack in #10279
- fix(generation): keep FLUX Kontext references after renaming by @renningg in #10238
- fix(generation): record and recall SD 3 text encoders by @renningg in #10228
- feat(upscale): support MOSS-Image in the Upscale widget by @FrankBro in #10257
- feat(workflows): redesign the workflow library, previews and editor nodes by @joshistoast in #10222
- feat(ui): workbench polish: desktop file drops, compare waiting state, reference image rows, Upscale prompt tools, compact dropdowns by @joshistoast in #10136
- fix(wan): keep ComfyUI fp8_scaled checkpoints packed instead of folding them to bf16 by @Pfannkuchensack in #10262
- fix(graph): default If branches to eager execution by @JPPhoto in #10032
- fix(models): keep ComfyUI fp8_scaled Qwen3 and Qwen2.5-VL encoders packed by @Pfannkuchensack in #10291
- feat(workflows): switch between graph and form by keyboard by @renningg in #10226
- fix(deps): require pypatchmatch 2.2 so PatchMatch loads offline by @Pfannkuchensack in #10298
- feat(sd3): load SD 3.5 transformers from GGUF builds by @Pfannkuchensack in #9972
- fix(image-map): fetch search URLs through the SSRF-guarded session by @Laihu08 in #10264
- fix(qwen-image-2.1): reserve measured VAE working memory on ROCm by @lstein in #10043
- refactor(palette): upgrade Chakra UI to 3.38 and build the command palette on Command by @joshistoast in #10247
- perf(model-manager): probe GGUF files without a full collection; drop dead tests by @Pfannkuchensack in #10301
- fix(queue): release a failed generation's VRAM and keep Empty Cache working by @Pfannkuchensack in #10267
- feat(canvas): add transform resampling controls by @DustyShoe in #10152
- fix(lora): reject diffusers LoRAs of other DiTs by Anima's attention widths by @Pfannkuchensack in #10223
- refactor(layers): give layer duplicate, delete, group and ungroup one owner by @joshistoast in #10307
- feat(moss-image): load unsloth's torchao fp8 transformers by @Pfannkuchensack in #10181
- fix(ui): keep scrolling menu lists out of the tab order; fix flaky workflow library tests by @lstein in #10315
- fix(vae): honor the loaded model compute device by @wunianze666-netizen in #10145
- fix(docs): map exact-base sidebar href to the empty page id in makeHrefToId by @jadelike-wine in #10269
- test(canvas,generation): remove timing races from two flaky browser tests by @DustyShoe in #10314
- feat(generation): update a loaded preset from the current settings by @lstein in #10285
- feat(canvas): add vector layers and spline editing by @DustyShoe in #10281
- chore(frontend): parallelize test CI and cut the suite and comments to essentials by @joshistoast in #10305
- fix(recall): clear reference images when the recalled image had none by @lstein in #10284
- refactor(canvas): judge canvas generation inputs once for readiness and invoke by @joshistoast in #10308
- feat(video): support H3 hybrid task overrides by @jaffer1979 in #10141
- refactor(settings): require resolved widget values wherever layered keys are read by @joshistoast in #10312
- refactor(generation): add reference images through one Generation operation by @joshistoast in #10311
- fix(models): let one model file serve as more than one type by @Pfannkuchensack in #10221
- refactor(models): own the architecture table and link VAEs as the backend serves them by @joshistoast in #10309
- feat(nodes): add native metadata input and output across denoise invocations by @skunkworxdark in #10083
Full Changelog: v7.0.0-beta.1...v7.0.0-beta.2













