Releases

Fixed a crash when training LoRA for SD v1.5 models.

Fixed a crash when using LoRA with SD 3.5 Medium base model. These LoRAs need to be re-imported.

A lot of functional and usability improvements on LoRA training for FLUX.1 models: a). Defaults to cosine annealing with warm restarts scheduler for learning rate; b). Introducing "Weights Memory Management" allowing you to loads weights only needs it layer-by-layer; c). Added TensorBoard support; d). Easier syntax for layer indices selection.

Added support for SD3 Medium 3.5.

Added support for training FLUX.1 [dev] LoRA.

Fixed a bug that Intel Mac talks to Apple Silicon Mac through gRPC server can cause a crash.

Added support for SD3 Large 3.5 and SD3 Large Turbo 3.5.

Fix instant crash on macOS Ventura.

Enhance "Server Offload" feature to default TLS enabled, allows you to turn off response compression, and adding server manually.

For larger models such as FLUX.1, allows to cancel the generation immediately.

Fix some animation issues.

Add "Server Offload" feature which allows you to enable gRPC server on one instance of the app on a different device and connect to that device to use its compute.

Fix import issues with Diffusers format FLUX.1 model.

Fix excessive RAM usage when merge some SDXL models.

Fix some more UI bugs.

Fixed a crash related to PuLID.

Fixed issues of importing FLUX.1 ControlNet that contains only double blocks. Alimama ControlNet Inpaint, Jasper AI ControlNet should now all work;

Fixed issues with guidance embed usage on FLUX.1 ControlNet. Now existing ControlNet should work better;

Fixed a long-standing outpainting edge cases that will incorrectly erase some of your content;

Better settings section separation for easier readability.

This is a hotfix version fixed a crash when resize then window down while image is generating;

Also included in this release: ControlNet ProMax import support.

1. Added support for PuLID 0.9 for FLUX.1 models;

2. Fixed issues related to ControlNet Union Pro release: a). crash when selecting scribble / softedge, b). Canny Edge Map broken for SDXL ControlNet;

3. Fixed issue that Community Scripts cannot be changed;

4. Added support for default LoRA weight / range, in custom_loras.json, you can edit it by adding: "weight": { "value": 0.5, "lower_bound": -10, "upper_bound": 10 }, such that the LoRA when selected will use the custom range rather than default -1.5 to 2.5 range.

1. Phase 1 integration with MFAv2, improves generation speed 5%-15% on M3 / M4 and ~2% on M1 / M2;

2. Added support for ControlNet Union Pro for FLUX.1 models;

3. Added support to import ControlNet models for FLUX.1;

4. Fixed issue FLUX.1 [dev] used with LoRA can blow-up RAM usage;

5. Some QoL UI improvements: copy settings now only contains relevant fields, new Script API to query which models are already available locally, remove unsupported models from LoRA training or exporting.

1. Fixed issue with OOTDiffusion's full body model. You need to redownload the Garment UNet model for it if you downloaded it before 8/30;

2. Added support for Kwai Kolor's IP Adapter FaceID Plus model. This enables better identity keeping than IP Adapter Face Full we previously support;

3. DDIM samplers now do fractional steps, so you can do img2img with DDIM samplers at 4 steps at any given percentage;

4. Some scripts improvements: fixed a timing issue when loading masks;

5. Some QoL UI improvements: added option to copy / paste canvas locations so you can working off the same canvas settings when switching between pictures; images added to moodboard now won't be cropped when switching aspect ratios.