1.9.0

[AUTOMATIC1111]

1.9.0

Features:

• Make refiner switchover based on model timesteps instead of sampling steps (#14978)
• add an option to have old-style directory view instead of tree view; stylistic changes for extra network sorting/search controls
• add UI for reordering callbacks, support for specifying callback order in extension metadata (#15205)
• Sgm uniform scheduler for SDXL-Lightning models (#15325)
• Scheduler selection in main UI (#15333, #15361, #15394)

Minor:

• "open images directory" button now opens the actual dir (#14947)
• Support inference with LyCORIS BOFT networks (#14871, #14973)
• make extra network card description plaintext by default, with an option to re-enable HTML as it was
• resize handle for extra networks (#15041)
• cmd args: --unix-filenames-sanitization and --filenames-max-length (#15031)
• show extra networks parameters in HTML table rather than raw JSON (#15131)
• Add DoRA (weight-decompose) support for LoRA/LoHa/LoKr (#15160, #15283)
• Add '--no-prompt-history' cmd args for disable last generation prompt history (#15189)
• update preview on Replace Preview (#15201)
• only fetch updates for extensions' active git branches (#15233)
• put upscale postprocessing UI into an accordion (#15223)
• Support dragdrop for URLs to read infotext (#15262)
• use diskcache library for caching (#15287, #15299)
• Allow PNG-RGBA for Extras Tab (#15334)
• Support cover images embedded in safetensors metadata (#15319)
• faster interrupt when using NN upscale (#15380)
• Extras upscaler: an input field to limit maximul side length for the output image (#15293, #15415, #15417, #15425)
• add an option to hide postprocessing options in Extras tab

Extensions and API:

• ResizeHandleRow - allow overriden column scale parametr (#15004)
• call script_callbacks.ui_settings_callback earlier; fix extra-options-section built-in extension killing the ui if using a setting that doesn't exist
• make it possible to use zoom.js outside webui context (#15286, #15288)
• allow variants for extension name in metadata.ini (#15290)
• make reloading UI scripts optional when doing Reload UI, and off by default
• put request: gr.Request at start of img2img function similar to txt2img
• open_folder as util (#15442)
• make it possible to import extensions' script files as import scripts.<filename> (#15423)

Performance:

• performance optimization for extra networks HTML pages
• optimization for extra networks filtering
• optimization for extra networks sorting

Bug Fixes:

• prevent escape button causing an interrupt when no generation has been made yet
• [bug] avoid doble upscaling in inpaint (#14966)
• possible fix for reload button not appearing in some cases for extra networks.
• fix: the split_threshold parameter does not work when running Split oversized images (#15006)
• Fix resize-handle visability for vertical layout (mobile) (#15010)
• register_tmp_file also for mtime (#15012)
• Protect alphas_cumprod during refiner switchover (#14979)
• Fix EXIF orientation in API image loading (#15062)
• Only override emphasis if actually used in prompt (#15141)
• Fix emphasis infotext missing from params.txt (#15142)
• fix extract_style_text_from_prompt [Bug]: Error when hitting ♻️ recycle button #15132 (#15135)
• Fix Soft Inpaint for AnimateDiff (#15148)
• edit-attention: deselect surrounding whitespace (#15178)
• chore: fix font not loaded (#15183)
• use natural sort in extra networks when ordering by path
• Fix built-in lora system bugs caused by torch.nn.MultiheadAttention (#15190)
• Avoid error from None in get_learned_conditioning (#15191)
• Add entry to MassFileLister after writing metadata (#15199)
• fix issue with Styles when Hires prompt is used (#15269, #15276)
• Strip comments from hires fix prompt (#15263)
• Make imageviewer event listeners browser consistent (#15261)
• Fix AttributeError in OFT when trying to get MultiheadAttention weight (#15260)
• Add missing .mean() back (#15239)
• fix "Restore progress" button (#15221)
• fix ui-config for InputAccordion [custom_script_source] (#15231)
• handle 0 wheel deltaY (#15268)
• prevent alt menu for firefox (#15267)
• fix: fix syntax errors (#15179)
• restore outputs path (#15307)
• Escape btn_copy_path filename (#15316)
• Fix extra networks buttons when filename contains an apostrophe (#15331)
• escape brackets in lora random prompt generator (#15343)
• fix: Python version check for PyTorch installation compatibility (#15390)
• fix typo in call_queue.py (#15386)
• fix: when find already_loaded model, remove loaded by array index (#15382)
• minor bug fix of sd model memory management (#15350)
• Fix CodeFormer weight (#15414)
• Fix: Remove script callbacks in ordered_callbacks_map (#15428)
• fix limited file write (thanks, Sylwia)
• Fix extra-single-image API not doing upscale failed (#15465)
• error handling paste_field callables (#15470)

Hardware:

• Add training support and change lspci for Ascend NPU (#14981)
• Update to ROCm5.7 and PyTorch (#14820)
• Better workaround for Navi1, removing --pre for Navi3 (#15224)
• Ascend NPU wiki page (#15228)

Other:

• Update comment for Pad prompt/negative prompt v0 to add a warning about truncation, make it override the v1 implementation
• support resizable columns for touch (tablets) (#15002)
• Fix [Bug]: The categories layout is different when localization is on #14591 using translated content to do categories mapping (#14995)
• Use absolute path for normalized filepath (#15035)
• resizeHandle handle double tap (#15065)
• --dat-models-path cmd flag (#15039)
• Add a direct link to the binary release (#15059)
• upscaler_utils: Reduce logging (#15084)
• Fix various typos with crate-ci/typos (#15116)
• fix_jpeg_live_preview (#15102)
• [alternative fix] can't load webui if selected wrong extra option in ui (#15121)
• Error handling for unsupported transparency (#14958)
• Add model description to searched terms (#15198)
• bump action version (#15272)
• PEP 604 annotations (#15259)
• Automatically Set the Scale by value when user selects an Upscale Model (#15244)
• move postprocessing-for-training into builtin extensions (#15222)
• type hinting in shared.py (#15211)
• update ruff to 0.3.3
• Update pytorch lightning utilities (#15310)
• Add Size as an XYZ Grid option (#15354)
• Use HF_ENDPOINT variable for HuggingFace domain with default (#15443)
• re-add update_file_entry (#15446)
• create_infotext allow index and callable, re-work Hires prompt infotext (#15460)
• update restricted_opts to include more options for --hide-ui-dir-config (#15492)

---

This discussion was created from the release 1.9.0.

[highnrgappalachian]

Haven't had a chance to check it out yet, but thanks for providing the option to revert extra networks to what they used to look like before the previous (atrocious) change.

[Soulreaver90]

When I load a img via PNG info and send to txt2img, it doesn't update the scheduler type, it simply remains at automatic. This is true for images generated before and with 1.9.0.

[bobba84]

I have this issue too!

[Soulreaver90]

Yup! I do not expect this to apply to generations before 1.9, but anything current should note what scheduler was used so it can load correctly in the dropdown. Probably an obvious oversight that can be fixed and yet remain compatible with older or forked builds.

[Appare-Ranman]

When I try to use img2img it fails with the message "TypeError: 'bool' object is not iterable". 20 hours ago everything was fine.

Part of the error in the console: Traceback (most recent call last):
File "E:\NEURAL NETWORK\Stable Diffusion\modules\call_queue.py", line 57, in f
res = list(func(*args, **kwargs))
File "E:\NEURAL NETWORK\Stable Diffusion\modules\call_queue.py", line 36, in f
res = func(*args, **kwargs)
File "E:\NEURAL NETWORK\Stable Diffusion\extensions\sd-webui-inpaint-background\lib_inpaint_background\webui_hijacks.py", line 28, in hijack_func
return original_img2img_processing(id_task, mode, prompt, negative_prompt, prompt_styles, init_img, sketch,
File "E:\NEURAL NETWORK\Stable Diffusion\modules\img2img.py", line 150, in img2img
override_settings = create_override_settings_dict(override_settings_texts)
File "E:\NEURAL NETWORK\Stable Diffusion\modules\infotext_utils.py", line 416, in create_override_settings_dict
for pair in text_pairs:
TypeError: 'bool' object is not iterable

[martinvra]

I have the same issue after last update of modules a few days ago. There was something wrong.

It helps if you didable extension "sd-webui-inpaint-background"

[StanleyPain]

Anyone else getting VERY slow generations after upgrading? Seeing things taking 2x longer to generate @ same settings compared to 1.8

[Soulreaver90]

I wonder if this is a Nvidia/Cuda only issue on 1.9. I am on AMD and I have 0 performance issues since upgrading.

[Kuzlata]

I can confirm that image generation takes at least 2 times longer. Back to 1.8.0 ( torch 2.1.2+cu121, xformers 0.0.23.post1) and everything is fast as it should. Settings on both Versions are the same. My graphics card GTX 1070.

[vagabund44]

How you did a clean rollback to 1.8?

[Kuzlata]

to start from scratch in a new folder

git clone -b "v1.8.0" https://github.com/AUTOMATIC1111/stable-diffusion-webui.git

[Enferlain]

same here, like 4 times slower

[yonpere]

Selected prompt text can not mouse-dragged moving. Are there any options that have changed?

[vagabund44]

After the update the generations are way down in quality. I used saved presets for generations, so im sure that changed parameters are not the reason.

Esspecially the faces are mostly coming out with artefacts or looking weird.

Generall quality of all generations seem to be lower than before V1.9

Am i the only one who has that issue? And what could be the reason?

[rumoredlegends]

Doing a fresh install of 1.9 seems to fix the problem with Loras from any 1.8 instances that were upgraded. The speed is definitely 5x faster at the cost of compatibility with older gens.

[StanleyPain]

Going to have to do a fresh install. Looks like too much stuff changed to try and repair my current install ;)

[Kuzlata]

A complete reinstall of 1.9.0 fixes all the speed problems

[vagabund44]

A fresh install would mean, that all add-ons, loras, etc, need be added new, right?

[ljleb]

You can copy them or specify the path in the command line for a lot of models including controlnet and loras.

[neoktion]

Yesterday automatic was updated to 1.9.0. img2img stopped working. Throws an error (TypeError: 'bool' object is not iterable).
Is there a way to fix it?

"sd-webui-inpaint-background" is installed.

[maxmarkeev]

With this update all generations looks very bad

[ljleb]

consider installing in a separate folder and leaving the current install untouched, much less prone to breaking something.

[vagabund44]

Would it work to reinstall new in a new folder and then copy the lora, model, embeddings folders into the new installation?

[StanleyPain]

Yes, I have done this a few times before.

As always, have backups of things you don't want to lose.

[ljleb]

You can copy things, that works. If you're not happy with the extra space this requires to store duplicated models, know there are command line options to specify the path for things like controlnet, lora, embeddings, upscalers, etc.

[Enferlain]

I did a separate clone of 1.9.0 and copied my config and extensions and venv and I got the slowdown. Probably the fact that I copied the venv if I have to guess so I'll just redo it

[vagabund44]

I made a comparison between V1.9 and V 1.8:

My feeling was right, The faces are weird in V1.9, while everyhting else is exact same quality:

(Only face restauration of A1111, no ADetailer used, no highrez)

V 1.8 Clean and detailed face and hair, well highlighted

V 1.9 Blurry face and hair, not good lighting

More clearly to see is the difference in "from side" images: (V 1.8 takes some seconds longer for every generations)

V 1.8, 22 seconds

V 1.9, 17,2 seconds (face and eyes totally weird

Anyone has a solution for that? As long it cannot be solved, i will keep using 1.8