This release includes support for FLUX ControlNets and improves the UX for common img2img flows in Canvas.
We've also added a fix for Apple Silicon users for the mushy noise issue and updated the installer to skip xformers for GPUs where it causes a performance hit.
Be sure to review the v5 release notes if you haven't already upgraded to v5.
FLUX ControlNets
We now support both XLabs and InstantX ControlNets for FLUX. We've found the Union Pro model substantially outperforms the other models and added it to the starter models. Other models work, but outputs are not as good.
You can use FLUX ControlNets in both Workflows and the Linear UI. We will be adding a union mode control to the Linear UI in a a future release. You can select the union mode in Workflows today.
Canvas img2img Flow
We've made a number of changes to support the common img2img flow in the Canvas:
- Transform now supports 3 modes:
fill(old behaviour): The layer is stretched to fit the generation bbox exactly. Its aspect ratio is not maintained.contain(new default): The layer is stretched to fit the generation bbox, retaining its aspect ratio.cover: The layer is scaled up so that its smallest dimension fits the bbox exactly, retaining its aspect ratio.
- Add a layer context menu item to fit the layer to the bbox using the
containmode. - Update the
New Canvas from Imageimage context menu item to streamline theimg2imgflow. It now resizes the bbox to match the image's aspect ratio, respecting the currently-selected models' optimal size. The image will fit exactly in the box. You can click this and then immediately Invoke to doimg2img.
Installer Updates
The performance of torch-sdp attention is substantially faster than xformers on 30xx and 40xx series GPUs. We've made two changes to ensure you generate with the best settings:
- Add an installer option for 30xx & 40xx series GPUs, which does not install
xformers. - When the attention type is set to
auto(the default), and you do havexformersinstalled, we choose the best option oftorch-sdpandxformers, based on your GPU.
Apple Silicon Fix
The mushy noise issue on Apple Silicon is related to sliced attention. We've temporarily forced all MPS devices to use torch-sdp. Memory usage is a bit higher, but you won't get mushy noise (unless, of course, you prompt for it 😅).
This appears to be a torch bug, and we'll revert this change once it is resolved.
Other Changes
Enhancements
- Updated workflow list menu UI, restoring
New Workflowconfirmation dialog. - Handful of rendering optimizations for Canvas.
Fixes
- Fixed misc UI jank in workflow list menu UI.
- Fixed longstanding issue where workflows were marked as unsaved immediately after loading.
- Fixed canvas layer preview not updating if layer is disabled.
Installation and Updating
To install or update to v5.2.0rc1, download the latest installer and follow the installation instructions
To update, select the same installation location. Your user data (images, models, etc) will be retained.
What's Changed
- feat(ui): img2img UX by @psychedelicious in #7081
- feat(ui): add option to copy share link for workflows by @maryhipp in #7082
- Allow explicitly selecting xFormers at install time by @ebr in #7065
- refactor(ui): modals state, workflow list modal misc by @psychedelicious in #7087
- perf(ui): rendering optimizations by @psychedelicious in #7061
- feat(ui): img2img UX 2 by @psychedelicious in #7090
- Add support for FLUX ControlNet models (XLabs and InstantX) by @RyanJDick in #7070
- feat(ui): consolidate img2img canvas flow by @psychedelicious in #7093
- fix(ui): use non-icon version of delete menu item on canvas context menu by @psychedelicious in #7094
- ui: translations update from weblate by @weblate in #7086
- chore: bump version to v5.2.0rc1 by @psychedelicious in #7096
- Add a workaround for broken sliced attention on MPS with torch 2.4.1 by @RyanJDick in #7066
- fix(ui): edge case where controladapters added counts could be off by @psychedelicious in #7095
Full Changelog: v5.1.1...v5.2.0rc1