InvokeAI v6.11.0

This is a feature release of InvokeAI which provides support for the new FLUX.2 Klein image generation and edit models as well as bug fixes and smaller feature releases. Before we get to the details, consider taking our 2026 User Engagement Survey. We want to know who you are, how you use InvokeAI, and what new features we can add to make the software even better.

About FLUX.2 Klein

The FLUX2 Klein family of models (F2K) are fast, high quality image generation and editing models. Invoke provides support for multiple versions, including both the fast-but-less-precise 4 billion (4B) and the slower-but-more-accurate 9 billion (9B) models, as well as quantized versions of these models suited for systems with limited VRAM. These models are small and fast; the fastest can render images in seconds with just four steps.

In addition to the usual features (txt2img, img2img, inpainting, outpainting) F2K offers a unique image editing feature which allows you to make targeted modifications to an image or set of images using prompts like "Change the goblet in the king's right hand from silver to gold," or "Transfer the style from image 1 to image 2".

Suggested hardware requirements are:

FLUX.2 Klein 4B - 1024×1024

• GPU: Nvidia 30xx series or later, 12GB+ VRAM (e.g. RTX 3090, RTX 4070). FP8 version works with 8GB+ VRAM.*
• Memory: At least 16GB RAM.
• Disk: 10GB for base installation plus 20GB for models (Diffusers format with encoder).

FLUX.2 Klein 9B - 1024×1024

• GPU: Nvidia 40xx series, 24GB+ VRAM (e.g. RTX 4090). FP8 version works with 12GB+ VRAM.
• Memory: At least 32GB RAM.
• Disk: 10GB for base installation plus 40GB for models (Diffusers format with encoder).

Getting Started with F2K

After updating InvokeAI, you will find a new FLUX.2 Klein starter pack for in the Starter Models section of the Model Manager. This will download three files: the Q4 quantized version of F2K 4B, which is suitable to run on low-end hardware, and two supporting files: the FLUX.2 VAE, and a quantized version of the FLUX.2 Qwen3 text encoder.

After installing the bundle, select the "FLUX.2 Klein 4B (GGUF Q4)" model in theGeneration section of Invoke's left panel. Also go to the Advanced section at the bottom of the panel and select the F2K VAE and text encoder models that were installed with the starter bundle. (If you don't select these, you will get an warning message on the first generation that tells you to do this.) Recommended generation settings are:

• Sampler: Euler
• Steps: 4-6
• CFG: 1-2

Modestly increasing the number of steps may increase accuracy somewhat. If you work with the Base versions of F2K (available from HuggingFace), increase the steps to >20 and the CFG to 3.5-5.0.

Text2img, img2img, inpainting and outpainting will all work as usual. InvokeAI does not currently support F2K LoRAs or ControlNets (there have not been many published so far). This functionality will be added in a future release.

Prompting with FLUX.2

Like ZiT, F2K's text encoder works best when you provide it with long prose prompts that follow the framework Subject + Setting + Details + Lighting + Atmosphere. For example: "An elderly king is standing on a low dais in front of a crowded and chaotic banquet hall bursting with courtiers and noblemen. He is shown in profile, facing his noblemen, holding high a jeweled chalice of wine to toast the unification of his fifedoms. This is a cinematic shot set that conveys historical grandeur and a medieval vibe."

F2K does not perform any form of prompt enhancement, so what you write is what the model sees. See FLUX.2 Prompting Guide for more guidance.

Image Editing

F2K provides an image editing mode that works like a souped-up version of Image Prompt (IP) Adapters. Drag-and-drop or upload an image to the Reference Image section of the Prompt panel. Then instruct the model on modifications you wish to make using active verbs. You may issue multiple instructions in the same prompt.

• Change the king's chalice from silver to gold. Give him a crown, and grow him a salt-and-pepper beard.
• Change the image style to a scifi/fantasy vibe.
• Use an anime style and give the noblemen and courtiers brightly-colored robes.

F2K editing supports multiple reference images, letting you transfer visual elements (subjects, style and background) from one to another. When prompting over multiple images, refer to them in order as "image 1," "image 2," and so forth.

• Give the king in image 1 the crown that appears in image 2.
• Transfer the style of image 1 to image 2.

Dealing with multiple reference images is tricky. There is no way to adjust the weightings of each image, and so you will have to be explicit in the prompt about which visual elements you are combining. If you cannot get the effect you are looking for by modifying the prompt, you may find success by changing the order of images.

Also be aware that each image significantly increases the model's VRAM usage. If you run into memory errors, use a smaller (quantized) model, or reduce the number and size of the reference images.

Other Versions of F2K Available in the Model Manager

To find additional supported versions of F2K, type "FLUX.2" into the Starter Models search box. This will show you the following types of files:

• FLUX.2 Klein 4B/9B (Diffusers) These are the full-size all-in-one diffusers versions of F2K which come bundled with the VAE and text encoder.
• FLUX.2 Klein 4B/9B These are standalone versions of the full-size F2K which require installation of separate VAE and text encoders. Note that the 4B and 9B models require different text encoders, "FLUX.2 Klein Qwen3 4B Encoder" and "FLUX.2 Klein Qwen3 8B Encoder" respectively. (Not a misprint: use the 9B F2K model with the 8B text encoder!)
• FLUX.2 Klein 4B/9B (FP8) These are the standalone versions quantized to 8 bits. The 4B model will run comfortably on macines with 8GB VRAM, while the 9B model will run on machines with 12GB or higher. As with all quantized versions, there is minor loss of generation accuracy.
• FLUX.2 Klein 4B/9B (Q4) These are standalone versions that have been quantized to 4 bits, resulting in very small and fast models that can run on cards with 6-8 GB VRAM.

There is only one F2K VAE, and it happens to be same as the one used by FLUX.1 and Z-Image Turbo. However, there are several text encoder options:

• FLUX.2 Klein Qwen3 4B Encoder Use this encoder with the F2K 4B versions. It also works with Z-Image Turbo.
• Z-Image Qwen3 Text Encoder (quantized) This is a Q6-quantized version of the text encoder, that works with both F2K and ZiT. You may use this on smaller memory systems to reduce swapping of models in and out of VRAM.
• FLUX.2 Klein Qwen3 8B Encoder Use this encoder with the F2K 9B versions. It is not compatible with ZiT.

You will find additional F2K models on HuggingFace and other model repositories, including the base models intended for fine-tuning and LoRA training. We have not exhaustively tested InvokeAI compatibility with all the available variants. Please report any incompatible models to InvokeAI Issues.

Many credits to @Pfannkuchensack for contributing F2K support.

Other Features in this Release

The other features in this release include:

Z-Image Turbo Variance Enhancer

ZiT tends to produce very similar images for a given prompt. To increase image diversity, @Pfannkuchensack contributed a Seed Variance Enhancer node which adds calibrated amounts of noise to the prompt conditioning prior to generation. You will find this feature in the Generation panel under Advanced Options. When activated, you will see two sliders, one for Variance Strength and the other for Randomize Percent. The first slider controls how much noise will be added to the conditioned prompt, and the second controls what proportion of the conditioning's weights will be altered. Using the default randomization of 50% of the values, a variance strength of 0.1 will produce subtle variations, while a strength of 0.5 will produce very marked deviation from the prompt. Increasing the percentage of weights modified will also increase the level of variation.

Improved Support for High-Resolution FLUX.1 Images

A new denoising tuning algorithm, introduced by @Pfannkuchensack, increases the accuracy of FLUX.1 generations at high resolutions. When a FLUX.1 model is selected, a new DyPE option will appear in the Generation panel. Its settings are Off (the default) to disable the algorithm, Auto to automatically activate DyPE when rendering images greater than 1536 pixels in either dimension, and 4K Optimized to activate the algorithm with parameters that are tuned for 4K images. Note that if you do not have sufficient VRAM to generate 4K images, this feature will not help you generate them. Instead, generate a smaller image and use Invoke's Upscaling feature.

Canvas high level transform smoothing

Another improvement contributed by @DustyShoe: The Canvas raster layer transform operation now supports multiple types of smoothing, thereby reducing the number of artifacts when an area is upscaled.

Text Search and Highlighting in the Image Metadata Tab

The Image Viewer's info (🛈) tab now has a search field that allows you to rapidly search and highlight text in image metadata, details, workflow and generation graph. In addition, the left margin of the metadata display has been widened to make the display more readable.

Thanks to @DustyShoe for this improvement.

Bugfixes

Multiple bugs were caught and fixed in this release and are listed in the detailed changelog below.

Installing and Updating

The Invoke Launcher is the recommended way to install, update and run Invoke. It takes care of a lot of details for you - like installing the right version of python - and runs Invoke as a desktop application.

Note: With recent updates to torch, users on older GPUs (20xx and 10xx series) will likely run into issues with installing/updating. We are still evaluating how we can support older GPUs, but in the meantime users have found success manually downgrading torch. Head over to discord if you need help.

Follow the Quick Start guide to get started with the launcher.

If you don't want to use the launcher, or need a headless install, you can follow the manual install guide.

Translation Credits

Many thanks to Riccardo Giovanetti for contributions to the Italian translation of InvokeAI. Also many thanks to Weblate for granting InvokeAI a free Open Source subscription to use its translation services.

What's Changed

Full Changelog: v6.10.0...v6.11.0.rc1.

• Add configurable model cache timeout for automatic memory management by @Copilot in #8693
• chore: Remove extraneous log debug statements from model loader by @lstein in #8738
• Feature: z-image + metadata node by @Pfannkuchensack in #8733
• feat(z-image): add add_noise option to Z-Image Denoise by @Pfannkuchensack in #8739
• (chore) Bump to version 6.10.0 by @lstein in #8742
• chore(release): bump development version to 6.10.0.post1 by @lstein in #8745
• Fix(model manager): Improve calculation of Z-Image VAE working memory needs by @lstein in #8740
• Chore: Fix weblate merge conflicts by @lstein in #8744
• ui: translations update from weblate by @weblate in #8747
• ui: translations update from weblate by @weblate in #8748
• Chore: Fix weblate rebase errors by @lstein in #8750
• fix(invocation stats): Report delta VRAM for each invocation; fix RAM cache reporting by @lstein in #8746
• Fix(UI): Error message for extract region by @DustyShoe in #8759
• Fix(UI): Canvas numeric brush size by @DustyShoe in #8761
• Feat(UI): Canvas high level transform smoothing by @DustyShoe in #8756
• chore(CI/CD): Remove codeowners from /docs directory by @lstein in #8737
• feat(z-image): add Seed Variance Enhancer node and Linear UI integration by @Pfannkuchensack in #8753
• fix(model_manager): prevent Z-Image LoRAs from being misclassified as main models by @Pfannkuchensack in #8754
• Add user survey section to README by @lstein in #8766
• ui: translations update from weblate by @weblate in #8767
• Limit automated issue closure to bug issues only by @Copilot in #8776
• Feat(UI): Search bar in image info code tabs and add vertical margins for improved UX in Recall Parameters tab. by @DustyShoe in #8786
• feat(flux2): add FLUX.2 klein model support by @Pfannkuchensack in #8768
• Feature: Add DyPE (Dynamic Position Extrapolation) support to FLUX models for improved high-resolution image generation by @Pfannkuchensack in #8763
• fix(model_manager): detect Flux1/2 VAE by latent space dimensions instead of filename by @Pfannkuchensack in #8790
• Prep for 6.11.0.rc1 by @lstein in #8771
• Fix ref_images metadata format for FLUX Kontext recall by @Pfannkuchensack in #8791