Please add a json selector to the UI - and please remove the XL from the name

[OlivioSarikas]

Please add a json selector to the UI, so we can load our custom lists on the fly.

Please remove the XL from the name, as this also works well with SD 1.5 and a lot of people don't or can't use XL, so it is more difficult to make videos or posts about it and it is confusing the community of users

[barepixels]

Check out Olivio Sarikas video on Style Selector

Style Selector for A1111 + ChatGPT HACK!!!!

[3Diva]

@OlivioSarikas Did you intend to leave that comment under the Fooocus-MRE discussion? Or did you intend to leave it under the Style Selector XL github?

[barepixels]

[3Diva]

@barepixels You can already add different style json files to Fooocus-MRE.

[barepixels]

nooooo I mean in one style sheet, to be able to drag a specific style such as "Artstyle Pop Art" and move it to the 3rd position from the top. Basically, drag and drop your favorite ones to the top. Reordering then export

[MoonRide303]

@barepixels Styles are being listed in following order:

1. default filenames (if they're present): default_styles_files = ['sdxl_styles_sai.json', 'sdxl_styles_twri.json', 'sdxl_styles_diva.json', 'sdxl_styles_mre.json'],
2. any other .json file, alphabetically.

So if you want order change, you can simply use names like "styles_1.json", "styles_2.json", etc. You can do that for sdxl*.json styles, too - just the "Default (Slightly Cinematic)" in coming from the code. So for example if I do something like that:

then It will be ordered accordingly in the UI:

Re-ordering json content is something you can do in any text editor. And if you'd like to do it in a smarter way you can use tools like jq. For example if you wanted to have all your style sorted alphabetically by name, you can do something like that:
type styles-in.json | jq "sort_by(.name)" > styles-out.json (Windows) or
cat styles-in.json | jq "sort_by(.name)" > styles-out.json (*nix).

[barepixels]

@MoonRide303 Found a solution

For example, change ones I like to:
{
"name": "Pop Art",
"prompt": "Pop Art style {prompt} . Bright colors, bold outlines, popular culture themes, ironic or kitsch",
"negative_prompt": "ugly, deformed, noisy, blurry, low contrast, realism, photorealistic, minimalist"
},

to:
{
"name": "001_Pop Art",
"prompt": "Pop Art style {prompt} . Bright colors, bold outlines, popular culture themes, ironic or kitsch",
"negative_prompt": "ugly, deformed, noisy, blurry, low contrast, realism, photorealistic, minimalist"
},

then sorted alphabetically by "name"

No coding needed

[barepixels]

@OlivioSarikas I just tested, while 1.5 models seem to work, even with refiner. 1.5 loras does not. Here is what @MoonRide303 said about support for 1.5 or 2.1 #29

EDIT: Am thinking 1.5 model does not work... I think it doesn't show errors but defaults to SDXL base

For example, I tried to recreate this in Fooocus and results looks similar to SDXL base NOT Deliberate
https://civitai.com/images/183148?modelVersionId=15236&prioritizedUserIds=4055&period=AllTime&sort=Most+Reactions&limit=20

[MoonRide303]

@OlivioSarikas You can add files with any names you want on style files you add into sdxl_styles folder - any file ending with .json will be loaded (as long as it has compatible syntax). Default file names starting with "sdxl" are result of two things:

1. those were the original names used in the repository they originally came from: sdxl_prompt_styler,
2. while styles using commonly recognized phrases can work pretty well between different models (and generations), it's not so obvious they will work as expected - different models might be trained using different keywords, so it's best to not make too far-reaching assumptions about styles compatibilities. Styles that were created and tested with SDXL model should be marked as sdxl, IMO.

Non-SDXL models are not currently supported - trying to load anything non-SDXL it will just result in reverting to the default model. As I previously mentioned and (in the issue mentioned by @barepixels) it is theoretically possible (even hybrid SD base + SDXL refiner workflows using VAE decode/encode at switch step), but it would complicate the flow - normalisation to model-native resolutions where it's needed, inpainting model being delivered as patch for SDXL architecture, LoRAs compatibility etc.

I could replace protection from loading non-SDXL models and replace it with warning messege, though - so people who know what they're doing could play with previous generation (it should work fine after disabling refiner, and setting resolution to something around 512x512).

[barepixels]

protection from loading non-SDXL models meaning models and loras?

we can add a set of 1.5 and 2.1 in our resolutions.json :)

Ideally, if choose 1.5 or 2.1 refiner should be off and hidden

With that said, I rather see you get DW Pose working before supporting 1.5 and 2.1 haha

DW Pose for A1111 - Better than OpenPose ControlNET

on YouTube I saw someone wrote in the comment
For those interested, the DWPose preprocessor also works great for ComfyUI and ControlNet

[MoonRide303]

@barepixels You can check out v2.0.78.3, I've just added limited support for SD models. Functions that will not work with non-SDXL models: refiner, Control-LoRAs, Revision, inpainting, outpainting.

[barepixels]

Oh yes. Deliberate v3, SD1.5, Prompt: redhead

Total time: 8.39 seconds on 3070 8GB 512 x 768

[barepixels]

Switch to SDXLbase and Refiner > Upscale Fast 2x

Total time: 10.80 seconds 1024 x 1536, NO METADATA

[MoonRide303]

@barepixels Fast upscale fixed (another side effect of merging changes from vanilla - MRE flow wasn't compatible with "direct_return" from vanilla). Restart to pull update, should be fine now.

[barepixels]

update Total time: 10.80 seconds, NO METADATA

[barepixels]

Switch to SDXLbase and Refiner > img2img scale 2

Total time: 48.58 seconds 1024 x 1536, METADATA included

[barepixels]

Switch to SDXLbase and Refiner > Upscale 2x (NOT FAST) Judging by her eyes, this is best out of 3 from Fooocus

Total time: 38.59 seconds 1024 x 1536, METADATA included

[barepixels]

Topaz GigaPixels (standard-scale-2_00x)

Total Time: N/A but very fast 1024 x 1536, METADATA included, UPDATE: Foooocus can load prompt from Topaz GigaPixels output

[MoonRide303]

@barepixels Fast upscale doesn't involve diffusion model - it's just simple ESRGAN type upscaler (not using prompt or any other generation parameters), so saving metadata doesn't make much sense in this case.

[barepixels]

understood, base on test images above, I won't be using Fast 2x anyway.