Change8

v6.14.0

Breaking Changes
📦 invokeaiView on GitHub →
9 breaking17 features🐛 1 fixes🔧 14 symbols

Summary

This release introduces significant new features including video generation, support for multiple new models (Krea, Ernie, Ideogram, Flux.2), multi-GPU and Intel XPU support, and FP8 precision for reduced VRAM usage. It also includes numerous performance improvements and enhancements to Canvas and workflow management.

⚠️ Breaking Changes

  • Ideogram 4 and Ernie-Image-Turbo models do not support image-to-image, inpainting, or outpainting. These modes will be rejected at graph build time for Ernie-Image-Turbo. Ideogram 4 will block generation if inpaint masks are used.
  • Reference images are unsupported across Krea-2-Turbo, Krea-2-Raw, Ernie-Image-Turbo, and Ideogram-4. Anima explicitly rejects reference images in canvas validators.
  • Krea-2-Turbo negative prompting is only honored when CFG > 1. While Turbo can run with CFG > 1, this is off-spec and may degrade output quality.
  • Krea-2 regional guidance is positive text only. Regional negative prompts, auto-negative, and regional reference images will raise warnings on canvas. The denoise node supports masked negative conditioning in workflows, but the canvas graph does not wire it.
  • Ernie-Image-Turbo does not support masked modes or image-to-image generation as the denoise node lacks a `denoise_mask` input.
  • Ideogram-4 does not support negative prompts at any CFG. The sampler uses asymmetric CFG with a zeroed unconditional branch, meaning the denoise node has no negative conditioning input.
  • Ideogram-4 regional guidance only supports positive prompts and bounding boxes. Regional negatives, auto-negative, and reference images will be dropped with warnings.
  • Anima ControlNet adapters (depth, scribble, lineart, pose) trained on Preview3 are weaker on Anima Base 1.0. The mixed-conditioning sketch adapter is recommended for general use.
  • Each LLLite ControlNet model may only be applied once per generation.

Migration Steps

  1. To enable low-VRAM mode for Wan generation, add `wan_memory_optimization: true` to your `invokeai.yaml` configuration file.

✨ New Features

  • Video generation support via Wan 2.2, including text-to-video, image-to-video, and image-to-image-video.
  • Support for Krea.2-Turbo, Krea.2-Raw, Ernie Turbo, and Ideogram 4 models.
  • Flux.2 Dev support.
  • Anima controlnets and inpainting capabilities.
  • Flux.2 PiD support for super resolution up to 4K.
  • Multi-GPU support for parallel generation across multiple GPUs.
  • Native Intel XPU support.
  • FP8 support for Anima, Flux.2 Krea Turbo & Raw, and Z-Image Turbo models to reduce VRAM usage.
  • VAEs can now be run on the CPU to reduce VRAM contention.
  • Improved support for LoRA sliders, including UNet-only sliders and the ability to edit min, max, and default values.
  • Pressure-sensitive brush opacity on tablets and touch screens in Canvas.
  • Fine-grid hint added to the move tool in Canvas.
  • New transparency lock for Gradients and Shapes in Canvas.
  • Clip Strokes extended to Bbox and Rect/Oval shapes in Canvas.
  • Optimized graph execution for faster complex workflows.
  • A LoRA collection picker node allowing multiple LoRAs to be stacked without additional nodes.
  • Ability to call one workflow from another for easier workflow management.

🐛 Bug Fixes

  • Synchronous API routes will no longer stall the entire server.

Affected Symbols