InvokeAI v6.14.1 Released with Krea-2 LoRA and Memory Fixes

September 7, 2026September 20, 2026

At a Glance

Values determined by this site’s code at collection time. Dates are JST.

Overview

InvokeAI v6.14.1 has been released. InvokeAI is a creative engine featuring an industry-leading WebUI for professionals and artists to generate visual media using Stable Diffusion models.

The biggest changes in this update are the optimization of memory management and support for Krea-2 LoRAs. It includes multiple improvements regarding VRAM and RAM reservations, making it a valuable update especially for users seeking stable operation while reducing resource consumption.

Key Changes

Expanded Support for Krea-2 LoRA

Krea-2 LoRAs built using ai-toolkit’s DoRA magnitudes are now supported. Users who build LoRAs using ai-toolkit or utilize Krea-2 format LoRs will be able to handle these new LoRA formats without issues with this update. Additionally, Kohya key processing has been improved with more robust validation.

UI and Workflow Feature Additions

New features have been added to enhance user workflow efficiency. First, current workflows can now be saved as PNG screenshots. This makes it easy to visually record and share complex node configurations.

Furthermore, a feature has been added to open media in the gallery directly in a new browser tab with a middle click. To use this feature, “Use Middle Click to Open Images/Videos in New Tab" must be enabled in the settings screen. This helps users who manage a large number of generations in the gallery to quickly review them.

Improvements to Memory Management and VRAM/RAM Reservation

Several important fixes related to resource management have been made. Specifically, these include improvements to VRAM memory reservations for VAEs, improvements to internal memory management during shutdown, and a fix for the issue where RAM reservations overshoot when using FP8.

In addition, working memory estimation accuracy during the denoising process and bidirectional VAE operations has been improved for Flux2 models. These improvements are extremely effective for engineers running large models using Flux2 and FP8, especially in hardware environments with limited VRAM or RAM capacity, to avoid out-of-memory errors and enhance operational stability.

Hardware Requirements

This update brings improvements to the following model architectures, quantization formats, and memory management.

Krea-2 LoRA

Krea-2 LoRAs built using ai-toolkit’s DoRA magnitudes are supported. This allows users training LoRAs with ai-toolkit to use their created LoRs directly in the InvokeAI environment. Moreover, the Kohya key validation process has also been improved to validate via un-flattening rather than prefix-based determination, realizing more robust LoRA loading.

Flux2

Working memory estimation accuracy during the denoising process and both VAE directions has been improved when using Flux2 models. This allows for more accurate memory allocation when using Flux2, aiming to optimize resource management.

FP8

Memory management when using the FP8 format has been improved. Specifically, it has been fixed to respect all declared skip lists, resolving the issue where the RAM reservation amount overshoots. This reduces the risk of out-of-memory errors during inference using FP8.

VAE

Processing related to VRAM memory reservations for VAEs has been improved. This makes memory usage during VAE execution more stable.

How to Get It

To update InvokeAI, please download and apply the latest version from the official launcher below.

Download the Launcher

Releases Since Our Last Article

Compiled by Local Model Watch from the project’s GitHub releases: the versions between this release and the last one we covered, which did not get separate articles. Full history: release tracker.

Sources

• InvokeAI Version 6.14.1 (GitHub Release)
• fix(lora): support ai-toolkit DoRA magnitudes for Krea-2 LoRAs (PR #9517)
• fix(krea2-lora): validate kohya keys by un-flattening, not by prefix (PR #9518)
• fix(fp8): honor every declared skip list, and stop overshooting the RAM reservation (PR #9538)
• fix(flux2): estimate working memory for denoise and both VAE directions (PR #9519)
• chore(deps): remove dependency on Jupyter Notebook (and update uv) (PR #9409)

Update History

• 2026-09-19: Rewrote the article from re-collected sources and restored it from draft to published.