SD3 support, innovations and possible problems. Is it worth updating?

Hello everyone, today an update has been released popular interface for generating images on Stable Diffusion models – Automatic 1111 up to version 1.10.0. My name is IlyaI am the founder of an online neural network for creating images ArtGeneration.mea tech blogger and neuro-evangelist, and we'll take a look at whether the new update is worth your attention or whether you can skip it.

I want to point out right away that this update is mainly focused on fixing bugs and minor performance improvements. The most notable new feature is support for Stable Diffusion 3, but unfortunately, many users, including me, have had problems running it. Despite the correct installation of the model, the system returns an error.

I also encountered compatibility issues with some extensions. For example, I had to remove the “replacer” extension because it was causing errors. Given these complications, I wouldn't recommend rushing to update just yet. It's better to wait a week or two until fixes are released. Especially since SD3 isn't working yet, and you probably won't notice any other significant changes.

However, if you do decide to upgrade, here is the full list of changes in version 1.10.0:

Version 1.10.0

Peculiarities:

• Many performance improvements (more details in the Performance section)

• Stable Diffusion 3 support

  • It is recommended to use the Euler sampler; DDIM and other time-stamped samplers are not currently supported.

  • The T5 text model is disabled by default, it can be enabled in the settings

• New planners:

  • Align Your Steps

  • KL Optimal

  • Normal

  • DDIM

  • Simple

  • Beta

• New sampler: DDIM CFG++

Minor changes:

• Ability to skip CFG at early stages

• Added –models-dir option

• Allows mobile users to open the context menu with a two-finger tap

• Infotext: Add Lora name as TI hashes for built-in text inversion

• Checking the model hash after downloading it to prevent corrupted downloads

• Additional extension tag filtering options

• When saving in AVIF format, the JPEG quality setting is used.

• Added file name template: [basename]

• Added option to enable skip clip for clip L on SDXL

• Option to prevent screen from going to sleep during generation

• ToggleLivePreview button in image viewer

• Eliminate interface flickering when rebooting and scrolling quickly

• Option to disable recording in log.csv of the save button

Extensions and API:

• Added hook process_before_every_sampling

• Return HTTP 400 instead of 404 on invalid sampler error

Performance:

• [Производительность 1/6] use_checkpoint = False

• [Производительность 2/6] Replacing einops.rearrange with native torch operations

• [Производительность 4/6] Preliminary calculation of the is_sdxl_inpaint flag

• [Производительность 5/6] Prevent unnecessary backup of additional network offsets

• [Производительность 6/6] Added –precision half option to avoid type casting during output

• [Производительность] LDM Optimization Patches

• [Производительность] Storing sigmas on CPU

• Check for nan in unet only once, after all steps are completed

• Added the ability to run the torch profiler to generate images

Bug fixes:

• Fix for grids without full infotext

• Partial lora update precedes full update

• Fixed a bug where the file extension had an extra dot under some circumstances

• Fix for loop when loading a corrupted model

• Allowing old sampler names to be used in API

• Improved compatibility with older sampler schedulers

• Fix Hypertile xyz

• XYZ CSV skipinitialspace

• Fix soft inpainting on mps and xpu, torch_utils.float64

• Fix for extensions updating when not on master branch

• Update safe file names for pickle

• Using relative path for css webui-assets

• When creating a virtual environment, update pip in webui.bat/webui.sh

• Fixing AttributeError

• Using script_path for webui root in launch_utils

• Fix for P Transparency Extra Party Mode

• Using gradio theme colors in css

• Fix for dragging text inside hint input field

• Adding the correct MIME type for .mjs files

• Improved handling of metadata issues for SD, Loras and embedding models

• Replacing wsl-open with wslpath and explorer.exe

• SDXL Inpaint Fix

• Multidimensional grid

• Fix for preview replacement

• Possible fix for incorrect scale in weight decomposition

• Making Python Usable from venv on Mac and Linux

• Prefer python3.10 over python3 if both are available on Linux and Mac (with fallback)

• Stopping generation of add-ons

• SD2 boot fix

• Fix Lora hashes in infotext for hires fix with different lora

• Fix sampler scheduler autocorrection warning

• Fix for interface flickering when rebooting and fast scrolling

• Fixing upscale logic

• Correction of the error when recording images in P mode as JPEG

• Other changes (including documentation updates, typo fixes, dependency updates, and minor improvements)

If you decide to upgrade despite the possible problems, the process is the same as I described for previous versions. Add the “git pull” startup argument to your webui-user.bat file by opening it in Notepad. If you don't want to upgrade every time you run it, you can remove this argument after upgrading.

For portable version users: check for update.bat file and run it to update.

If you have problems, I recommend either doing a clean install or at least deleting the “venv” folder in the Automatic1111 directory. This will reinstall all dependencies and may resolve any issues you may have.

Remember that stability is more important than new features, especially if those features don't work correctly yet. Be careful with updates and always have a backup of your working configuration.

I talk more about neural networks on my site YouTubeV telegram and on Busty. I will be glad for your subscription and support. Hugged everyone. Happy generations!