InvokeAI v6.14.2 Released with Custom Fonts and Model Fixes

InvokeAI v6.14.2 Released with Custom Fonts and Model Fixes

At a Glance

Item Value
Repository invoke-ai/InvokeAI
Version v6.14.2
Published 2026-09-28
License Apache-2.0
Source type Primary source (the publisher itself)

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

Overview

InvokeAI version 6.14.2 has been released. InvokeAI (invoke-ai/InvokeAI) is an open-source software developed in Python under the Apache-2.0 license that provides an advanced WebUI and creative engine for running and operating image generation models such as Stable Diffusion. In addition to image generation, it features versatile image editing capabilities including img2img, inpainting, and outpainting.

This version 6.14.2 introduces new features such as custom font importing on the Canvas and style transfer using reference images in krea-2. It also removes the non-functional FP4 starter model, improves VRAM management, and includes various bug fixes. Because it avoids errors in specific configurations and improves usability, users running target models or utilizing the Canvas features may want to consider updating.

Key Changes

Custom Font Import Support in Canvas Text Tool

Users can now import and use their own custom fonts in the text tool within the Canvas. This affects creators who add text to generated images on the Canvas. After updating, loading font files other than the standard built-in fonts expands design possibilities.

Style Transfer with Reference Images in krea-2

A feature has been added to transfer image styles to krea-2 by loading a reference image (Reference Image). This greatly benefits users who want to generate images using krea-2 while maintaining a specific art style or graphic style. Style transfer can be applied by specifying a reference image on the GUI.

Improved VRAM Allocation and Release for Quantized Qwen2.5-VL Encoder

A mechanism was introduced to pre-allocate the required VRAM space when running the quantized Qwen2.5-VL encoder and reliably release the VRAM after processing finishes (PR #9572). This affects users running the Qwen2.5-VL encoder in local environments with limited VRAM capacity. After the update, VRAM is properly released after generation processing, preventing errors due to VRAM shortages and slowdowns in subsequent processing.

Removal of FLUX.2 [dev] FP4 Mistral Starter Model

Because the Mistral loader did not support FP4, the FLUX.2 [dev] FP4 Mistral encoder starter model was removed (PR #9587). In previous environments, downloading 12.3 GB of model data still resulted in a shape mismatch error during initial generation. This affected users who encountered this issue and those utilizing the FLUX.2 [dev] model. After updating, FP4 will be excluded from the options. Users should select a supported variant such as FP8, BF16, or the GGUF format.

Limiting Supported Encoders for Z-Image Models to Qwen3 4B

For Z-Image models, the available Qwen3 encoder is now restricted to “Qwen3 4B" (PR #9587). Z-Image requires 2560-width embedding data, but previously the 8B encoder for Klein 9B (4096-width) could be selected, which caused errors during the denoising stage. This affects users generating images with Z-Image. After the update, only the Qwen3 4B model will be applied in the UI model selection and automatic assignment, preventing issues caused by incorrect settings.

Fixes to Validation and Save Processing When Renaming Models

When renaming a model, the validation conditions (character length check) have been corrected to allow setting short names of 3 characters. Additionally, a bug where the save button could not be clicked after form error states were cleared has been resolved (PR #9587). This affects users renaming or organizing models in the model management screen. Since the behavior during setting changes is now correct, intuitive renaming can be done without special workarounds.

Expanded Scope of “I" Keyboard Shortcut

Keyboard shortcut operations using the “I" key have been enabled across the Canvas, Gallery, and Viewer screens (PR #9569). This affects users who prefer efficient work centered around keyboard operations. No special preparation is required, and operations via the “I" key are available immediately after updating.

Support for ComfyUI Metadata (comfy.gguf.orig_shape)

Adjustments were made to correctly read and process comfy.gguf.orig_shape, which is the GGUF model metadata created and output by ComfyUI (PR #9564). This affects users who integrate model files and GGUFs in conjunction with a ComfyUI environment.

Supported Models and Hardware

In InvokeAI 6.14.2, restrictions and adjustments have been made to improve compatibility and prevent bugs regarding various supported image generation models and vision-language encoder models. The models and formats affected in this version are as follows:

krea-2

Image style transfer specifying a reference image (Reference Image) as input is now supported in krea-2. This makes it possible to reflect the style or texture of an existing image into the image generation process by krea-2.

FLUX.2 [dev] and Mistral Encoder

To address the issue where FP4 format support did not exist in the Mistral loader, the FLUX.2 [dev] FP4 Mistral encoder starter model was removed. The issue where generation stopped with a tensor shape mismatch error after downloading 12.3 GB of model data has been resolved. Note that while FP4 has been excluded, GGUF format, FP8, and BF16 variations of FLUX.2 [dev] remain unaffected and continue to be available.

Z-Image and Qwen3 Encoder

Encoders available for Z-Image models are now limited to “Qwen3 4B" (variant: qwen3_4b). While Z-Image is designed to process 2560-dimensional width embedding data internally, the 8B encoder for Klein 9B (4096-dimensional width) was previously selectable, causing errors due to inconsistencies during the initial denoising steps. With this fix, guards have been applied to the encoder selection picker, automatic selection when switching Z-Image, and metadata reading processes, ensuring only the appropriate Qwen3 4B encoder is provided.

Qwen2.5-VL (Quantized Model)

When running quantized Qwen2.5-VL encoder models, improvements were made to properly allocate VRAM space and release the space after processing completes. This prevents resource exhaustion and errors caused by tight VRAM capacity, allowing for more stable inference execution.

Support for ComfyUI GGUF Metadata

Support has been added to correctly interpret and process metadata comfy.gguf.orig_shape contained in GGUF models output and used in ComfyUI environments. Compatibility when handling GGUF format models quantized and created in ComfyUI has been enhanced.

Other Encoder-Related Fixes

An issue where SDNQ guard processing interrupted folder-type encoder identification midway has been resolved, improving encoder detection processing to run normally.

How to Get It

To install and update InvokeAI, using the officially provided Launcher is recommended. Follow the installation guide in the official repository and download the Launcher from the link below to perform the update.

Download the Launcher

If you are running an existing environment, you can apply the update to the latest version 6.14.2 through the launcher. After updating, please ensure that you avoid the removed FLUX.2 [dev] FP4 Mistral starter model and confirm that a BF16, FP8, or GGUF format encoder is specified. Also, if you are using the Z-Image model, make sure the encoder is set to Qwen3 4B.

Related Articles

What to Read Next

Sources

Engines and ToolsComfyUI,FLUX.2,InvokeAI,Qwen2.5-VL,Stable Diffusion,Verified,Z-Image