Extensions for Forge Neo

[hirorohi03]

There have been many reports that certain well-known, long-standing extension does not work on Forge Neo or support new models such as Anima.
There are forks or alternative extensions developed to work on Forge Neo, but they are not necessarily widely known.

I think it would be helpful to have all the information about extensions that work on Forge Neo in one place.

The table below summarizes the tools I know.
If you could reply about other extensions for Forge Neo I will add or update this table.

I won't add them to the list, but if you know of any new and useful extensions—even if they aren't traditional ones—please share them in this thread.

Note:

• This list includes only the following extensions. (I will ignore extensions that do not meet these criteria.)
  • Forks and alternatives to extensions that no longer work properly in Forge Neo.
  • Forks and alternatives to extensions that are incompatible with new models.
  • Forks and alternatives featuring a UI that is more optimized for Gradio 4 than existing extensions.
• If the original repository that was forked updated to work on Forge Neo, I may remove that fork from the list.
• I haven't tested all the extensions on the list, so I might be wrong.

🧩 Extension	🌐 GitHub Repository	📝 Note
AADetailer Forge Neo	https://github.com/abzaloff/aadetailer-neoforge	(Original)
ADetailer Neo	https://github.com/Haoming02/ADetailer-Neo	"Official" fork (Original)
Agent Scheduler Neo	https://github.com/cataclisma/sd-webui-agent-scheduler-neo	(Original)
Civitai	There are many forks of Civitai Helper and Civitai Browser, and most of them run on Forge Neo.	Find a fork you like on your own.
Dynamic Prompts	https://github.com/abzaloff/sd-dynamic-prompts	(Original)
Forge NegPip	https://github.com/flyfront/sd-forge-negpip (for Krea 2, MOSS-Image, Anima, SD1, SDXL)	Currently, Original supports Anima, Krea2 and Forge Neo.
Lama cleaner	https://github.com/abzaloff/forge-neo-lama-cleaner	(Original)
LoRA Block Weight ForgeNeo Anima	https://github.com/kiygskr/sd-webui-lora-block-weight-forgeneo-anima	Currently, Original supports Anima, Krea2 and Forge Neo.
Prompt All-in-One NEO	https://github.com/eduardoabreu81/sd-webui-prompt-all-in-one-neo	Original also works on Neo
Repeat Generate for NEO	https://github.com/hirorohi03/repeat-generate-4NEO	Alternative of “Generate forever”
State Manager Neo	https://github.com/eduardoabreu81/sd-webui-state-manager-neo	(Original)
SuperMerger-ForgeNeo-Anima	https://github.com/kiygskr/sd-webui-supermerger-forgeneo-anima (For Anima only)	Currently, Original supports Anima, Krea2 and Forge Neo.
TagComplete Neo	https://github.com/eduardoabreu81/sd-webui-tagcomplete-neo	Original also works on Neo
WD14 Tagger	https://github.com/hirorohi03/stable-diffusion-webui-wd14-tagger	(Original)
WD14 Tagger	https://github.com/szxex/stable-diffusion-webui-wd14-tagger	Supports Pixai Tagger, Oppai Oracle, and others.

[Git-Maniac]

Very good list!

Actually you could put on that list all NEO releases from abzaloff and eduardoabreu81 like:

composer_forge_neo: https://github.com/abzaloff/composer_forge_neo
sd-civitai-browser-neo: https://github.com/eduardoabreu81/sd-civitai-browser-neo
sd-webui-state-manager-neo: https://github.com/eduardoabreu81/sd-webui-state-manager-neo

Greetings

[hirorohi03]

The extensions forked by eduardoabreu81 are great, but most of the original extensions work fine on Forge Neo.
I'm torn on whether I should even add it to the list.

[abzaloff]

I forked this extension yesterday and now it works on Forge Neo.
https://github.com/abzaloff/forge-neo-lama-cleaner

[SiliconeShojo]

ScribeNEO: https://github.com/SiliconeShojo/ScribeNEO
Civitai Helper: https://github.com/SiliconeShojo/Civitai-Helper
LoRA Block Weight Neo: https://github.com/SiliconeShojo/lora-block-weight-neo
Codex: https://github.com/SiliconeShojo/Codex

[flux-error-man]

I use Zixaphir.
It didn't work with forge neo as is, so I found a way to modify model.py somewhere and executed it.
The Civitai Helper likely experienced bugs every time the web UI was updated or a new web UI was released,
which is probably why so many forks were created.
However, it seems that the various improvers haven't actually made any improvements or additions to the functionality,
so the shortcomings remain unchanged.
I use Web Browser when downloading from Civitai.
What I think is necessary for Civitai Helper is LoRA and Checkpoint tabs, and LoRA duplicate checks.
But for some reason, this tab doesn't appear in Civitai Helper.
I have to switch to another tab each time to bring up these tabs.
It would be convenient if LoRA's duplicate check could select and delete the duplicate data it finds.
While there are ways to correct incorrect links in LoRA and Checkpoint, the procedure is not very clear.
I can search for LoRA data using the LoRA tab, but I find it inconvenient that I can only search by the beginning of the string.
LoRA's name sorting method is also different from Windows.
If I add a link to data downloaded from other sites like SeaArt, will it be possible to at least check for updates?
I can't code, so I wonder if someone could make a good one for me?

[hirorohi03]

@flux-error-man
You're right.
I don't use the [Civitai Helper Browser] tab in Civitai Helper. This is because, in addition to changes to the WebUI, there are also changes to the Civitai API and the impact of various issues, so accessing Civitai directly through a web browser is much more convenient.

I use ComfyUI-Lora-Manager to manage Lora. It also comes in a Portable Standalone Edition, so you can use it without ComfyUI.

[flux-error-man]

I used both Comfy and Forge in the past. I uninstalled both after switching to Forge Neo.
I was looking for an alternative to Civitai Helper for Comfy. I think it might have been ComfyUI-Lora-Manager.
It's probably not the same as Civitai Helper; perhaps it requires a separate thumbnail image, or the specifications
of other files are different?
Or can ComfyUI-Lora-Manager share files with the Civitai helper?
I felt it was a waste of resources, so I discarded it when I switched to Forge Neo.
Also, while Forge Neo now supports AVIF,
The Civitai Helper I use can handle AVIF thumbnails, which is convenient.

[flux-error-man]

This is specifically designed for Forge Neo, so it might be more stable.
https://github.com/abzaloff/sd-webui-lora-quickpick
It's a shame that it doesn't seem to support AVIF.
I also think the Civitai browser is unnecessary, so I've set it to not display, but I would be even happier if it could be completely removed.

[flux-error-man]

About LoRA Block Weight Neo.
Sometimes, presets I've created myself do not appear in the list.
It does not appear in the list even after restarting only the web browser in Forge neo.
It appears in the list after a full restart of Forge neo.
But there are also presets that show up in the list normally, right?
Also, it seems that the presets provided for Flux are not necessarily correct.
Double-bits do not seem to be useful for the Flux LoRA I created.

[PunkRat9]

Thanks man, you're a life saver. I was wondering why my extensions weren't working on neo.

[PixelArmony]

Here are my contributions and findings for you all. They all work in Forge Neo:

Img2img Hires Fix: https://github.com/Amadeus-AI/img2img-hires-fix

• Hires fix upscale right after the img2img process, using its latent.
• Inspired by wcde/custom-hires-fix-for-automatic1111, which has been archived and isn't functioning since 2023.
• Simple 150 lines extension with minimized features and dependency.
• It also works in text2img, you can check the consistency between the Original "Hires. fix" and this extension.
• (Originally, if you upscale using/before img2img, the anatomy will be broken at higher resolution.)

ReActor NSFW Freedom: https://github.com/kainatquaderee/sd-webui-reactor-Nsfw_freedom

• This extension removes the built-in NSFW (Not Safe For Work) censorship filter from ReActor (the popular face-swap extension).
• By default, ReActor scans the image you upload and the image you want to swap onto. If it detects nudity or explicit content, it blocks the process and refuses to swap the face.
• Installing this modified file overrides that restriction, giving you total freedom to use the face-swapper on any image without triggering content blocks or refusal errors.
• There is no difference for working in Forge and now Forge-Neo.

Face Editor: https://github.com/ototadana/sd-face-editor

• Fixing broken faces, including faces in the background!
• Changing facial expressions
• Apply blurring or other processing

Forge Attention Couple: https://github.com/Haoming02/sd-forge-couple

• This extension allows you to generate couples target different conditionings at specific regions.
• No more color bleeds or mixed features!

Forge IC-Light: https://github.com/Haoming02/sd-forge-ic-light

• This is an extension allows you to manipulate the illumination of images.

Detail Daemon: https://github.com/muerrilla/sd-webui-detail-daemon

• This extension allows users to adjust the amount of detail/smoothness in an image, during the sampling steps.
• It uses no LORAs, ControlNets, etc. and as a result its performance is not biased towards any certain style and it introduces no new stylistic or semantic features of its own into the generation.
• This also means that it can work with any model and on any type of image.
• Works with SD 1.5, SDXL, Flux, Qwen Image etc.

Vectorscope CC: https://github.com/Haoming02/sd-webui-vectorscope-cc

• This extension performs a kind of Offset Noise natively during inference, allowing you to adjust the brightness, contrast, and color of the generations.

Yapping: https://github.com/Haoming02/sd-webui-yapping

• This extension adds fully-customizable preset buttons, that set specified parameters to specified values when clicked.

Test my prompt!: https://github.com/Extraltodeus/test_my_prompt

• Have you ever used a very long prompt full of words that you are not sure have any actual impact on your image? Did you lose the courage to try to remove them one by one to test if their effects are worthy of your pwescious GPU? Well now you don't need any courage as this script has been MADE FOR YOU!
• It generate as many images as there are words in your prompt (you can select the separator of course).

PNGINFO Beautify: https://github.com/bluelovers/sd-webui-pnginfo-beautify

• Raw text metadata inside your generated images into an organized, readable dashboard.
• When you drag an image into the standard PNG Info tab, Stable Diffusion normally spits out a giant block of plain text containing your prompts, seeds, sampler, and hash codes mixed together. This extension intercepts that raw text and organizes it visually.
• Color-Coded Prompts: It adds syntax highlighting to your positive and negative prompts, separating individual tags cleanly so you can read them effortlessly.
• Clickable Model/LoRA Links: It automatically detects the model hashes, VAEs, and LoRA names embedded in the image metadata and formats them as links. If the file exists on Civitai, it shows the resource data directly.
• Recovers Dynamic Templates: If you used the Dynamic Prompts script to randomize a generation, this extension recovers the original wildcard template formula (e.g., apparel), rather than just the final flattened text, making it much easier to resume your exact workflow if your PC crashed.

Webui Prompt Format: https://github.com/Haoming02/sd-webui-prompt-format

• Sometimes, when you type too fast or copy prompts from all over the places, you end up with duplicated spaces and commas. This simple Extension helps removing them whenever you click Generate!
• Works in both txt2img and img2img.
• Works in both Positive and Negative, as well as Hires. fix prompts.
• Remove extra spaces and commas.
• Fix misplaced brackets, commas, and periods.
• Enable Remove Duplicates to remove identical tags found in the prompts.
• Only checks within the same line.
• Only works for tag-based prompt, not sentence-based prompt.
• Enable Remove Underscores to replace _ with space.
• Append a comma every line break.
• Toggle between auto formatting and manual formatting.
• In Auto mode: The process is ran whenever you click Generate.
• In Manual mode: The process is only ran when you click on the Format button.
• Press Alt + Shift + F to manually trigger formatting.
• Format the text pasted from clipboard.
• Clean up unwanted texts when pasting from Booru sites.
• supports danbooru and gelbooru.

[keenerb]

Is it approprite to put in a request here?

https://github.com/thygate/stable-diffusion-webui-depthmap-script

I'd quite like this one to work.

[wktra]

@kiygskr

Just adding a note for anyone testing SuperMerger / Plus LoRA with Forge Neo.

I was trying to merge an AI Toolkit-trained SDXL LoKR into an Illustrious-based checkpoint.
The adapter was trained as LoKR, not plain LoRA, and text encoder training was enabled.

In sd-webui-supermerger-forgeneo-anima, Plus LoRA started correctly and did try to merge the adapter into the checkpoint. It loaded the base model, read the LoKR, synchronized metadata, and processed the modules. So it was not simply user error or the wrong tab.

First issue I hit was:

AttributeError: 'Options' object has no attribute 'lora_in_memory_limit'

That came from the extension expecting an A1111-style option that Forge Neo did not expose. I locally patched it by replacing the direct option access with a fallback value.

After that, the merge got farther, but then failed during checkpoint validation with:

RuntimeError: Non-finite values detected while merging LoRA: conditioner.embedders.0.transformer.text_model.encoder.layers.11.mlp.fc1.weight

That tensor is on the SDXL text encoder / conditioner side, not the UNet. My guess is that the Forge Neo fork is catching non-finite values created during the text-encoder-side merge of this LoKR.

I also saw this warning in both Forge Neo and old Forge:

lora_unet_up_blocks_0_upsamplers_0_conv.alpha
lora_unet_up_blocks_0_upsamplers_0_conv.lokr_w1
lora_unet_up_blocks_0_upsamplers_0_conv.lokr_w2

Those unmatched upsampler conv keys did not appear to be fatal.

What worked:

I launched my older non-Neo Forge install with the original SuperMerger extension and ran the same Plus LoRA merge there. It processed all 1051 modules, saved the merged checkpoint, and then crashed only after saving while trying to reload/revert the model:

ValueError: You do not have any model! Please download at least one model in [models/Stable-diffusion].

Despite that post-save reload error, the checkpoint file was created successfully. I copied the merged checkpoint into Forge Neo, loaded it normally as a checkpoint, and it generated correctly.

So the practical workaround for me was:

1. Use old Forge + original SuperMerger to run the Plus LoRA merge.
2. Ignore the post-save reload crash if the file was already saved.
3. Move the merged checkpoint back into Forge Neo.
4. Test it there.

This seems specifically relevant to SDXL LoKR / LyCORIS-style adapters trained with text encoder enabled. The Forge Neo SuperMerger fork may need better compatibility around lora_in_memory_limit and around text-encoder-side LoKR merge validation.

TLDR: The Supermerger Neo Extension DOESN'T WORK! @hirorohi03

[hirorohi03]

Thank you for the information.
I was wondering why you posted such detailed information here, but I guess it's because @kiygskr isn't accepting issues.

[hirorohi03]

It appears that the following is a fork of kiygskr/sd-webui-supermerger-forgeneo-anima, modified to work with anima.
https://github.com/aiai-r/sd-webui-supermerger-forgeneo-anima

I'm not really interested, so I think everyone would appreciate if you tried it and reported back the results.

[wktra]

Thank you for the information. I was wondering why you posted such detailed information here, but I guess it's because @kiygskr isn't accepting issues.

I apologize for taking up this space.
I was mistaken. I thought this extensions list were for extensions that worked properly with all models.
I was not aware that these extensions only had to function for anima processes.

[zeydsama]

does anyone have port this extension? https://github.com/[ClownsharkBatwing/RES4LYF](https://github.com/ClownsharkBatwing/RES4LYF/tree/main)/tree/main

[zeydsama]

have a try:
https://github.com/zeydsama/sd-forge-beta57.git
https://github.com/zeydsama/sd-forge-res4lyf.git

[SiliconeShojo]

[flux-error-man]

I'd like to try it.
Please tell me what it can and cannot do.
Can I install and use it while keeping Civitai Helper installed?
Can thumbnails and metadata be shared with Civitai Helper?
If it is not affiliated with Civitai, or if it has access to other sites, it might want to change it's name.
Is it something like a LoRA management helper?

[Squibeel]

@SiliconeShojo

Actually looks good. How does it filter per model? metadata or do we have to manually set the model?
Does it still retain the card buttons? (rename, delete, etc)?

I'll give it a try later, please consider adding this website in the future btw: https://arcenciel.io/models

EDIT: After trying answered the above questions myself.

[zeydsama]

@SiliconeShojo hii nice works btw! id like to try your extension, but i cant found it on your repository.

[SiliconeShojo]

@SiliconeShojo hii nice works btw! id like to try your extension, but i cant found it on your repository.

This was originally just a proof of concept and was never heavily optimized. As a browser extension, it relied on a significant amount of JavaScript, which resulted in noticeable performance issues. I've since decided to discontinue the project and will no longer be maintaining it. If you're interested, you can still browse the public archive here:

https://github.com/SiliconeShojo/civitai-neo

[Haoming02]

"Official" fork of ADetailer: https://github.com/Haoming02/ADetailer-Neo

[SiliconeShojo]

@Haoming02

I love the cleaned-up UI, especially the inpaint section with the toggleable settings.

Just so you know, the extension will fail if the built-in ControlNet is disabled (as it is in my setup), it throws a ModuleNotFoundError. I temporarily hotfixed it by wrapping the imports in a try/except block, though I know that’s not the ideal solution.

Also, class filtering would be a great feature. This fork claims to support it: adetailer-ultimate, although I haven’t tested it yet.

[Squibeel]

@Haoming02 Would you consider adding some of the stuff introduced in this fork? https://github.com/Anzhc/aadetailer-reforge

• Automatically include loras If loras are present in prompt, they will be automatically added. If their name include main trigger, it can be included too. Schema for trigger in name: <lora:lora name (trigger) blah blah:1> - basically what is inside () is considered trigger.
• Autotag before inpaint (Autotags crop area, so inpaint is stable, and doesn't require re-prompting each gen)
• Reworked resolution, now based on scaling. (Define a scale(multiplier) for resolution over base, so it's always bigger than original)

The first one at the very least.

[Haoming02]

Automatically include loras If loras are present in prompt, they will be automatically added

So this is why dozens of people complained in an Issue that LoRA fried their images when using ADetailer. It was the Extension after all...

Reworked resolution, now based on scaling. (Define a scale(multiplier) for resolution over base, so it's always bigger than original)

I... don't see the point of this. You should always inpaint at 1 MP no?

[Haoming02]

class filtering

It is included. You just need to install clip from Ultralystics

[Squibeel]

So this is why dozens of people complained in an Issue that LoRA fried their images when using ADetailer. It was the Extension after all...

Never happened to me but maybe you could test it but i feel its unrelated to that issue.

[TohnoSeika]

Hires for Neo in t2i is urgent needed.

[hirorohi03]

It is because Anima has a low maximum resolution at which it does not generate noise; it is not due to Forge Neo.
You'll get the same result in ComfyUI, too.

[TohnoSeika]

It is because Anima has a low maximum resolution at which it does not generate noise; it is not due to Forge Neo. You'll get the same result in ComfyUI, too.

Yes sadly that’s the case. 😭

[Haoming02]

In my experience, Anima Base 1.0 can generate up to ~3 MP just fine.

Higher than that, you will need to use MultiDiffusion in img2img

[hirorohi03]

@MinagiCN
If you want to learn more about using Anima for upscaling (Hires.fix, Multidiffusion and SD Upscale), in Forge Neo, check out my article
https://note.com/hirorohi03/n/n3eaa9da386f8

Articles on note.com are currently available only in the language in which they were written (mostly Japanese), but note.com plan to introduce multilingual support starting tomorrow.

[Haoming02]

#1179 (comment)

[SiliconeShojo]

Models Info: https://github.com/SiliconeShojo/models-info.git

This is not a Civitai extension fork, this was built entirely from scratch with a strict focus on performance, objectivity, and zero bloat.
It leverages the Civitai and Arcenciel APIs to handle checkpoint and LoRA metadata seamlessly without any extra hassle.

What it does

• Smart Downloads & Compression: Queries online databases to retrieve model metadata and preview images. It automatically compresses heavy preview files into optimized local JPEGs to save your disk space.
• Instant UI Filtering: Injects a base model selection dropdown directly into the native Extra Networks header toolbar. You can instantly filter your model card grids by architecture (e.g., SD 1.5, SDXL, Flux, Pony, SD 3.5) with zero page reloads.
• Automated LoRA Organization: Scans model tags on download and automatically routes new LoRAs into organized subfolders matching their actual usage categories (Style, Character, Concept, etc.).

What it doesn't do

To keep the extension lightning-fast and hyper-focused, features like checking for updates, scanning for duplicates, or downloading actual model files are not and will not be implemented.

If you strictly want speed, clean organization, and a lightweight footprint, give it a try! If you need a heavy-duty library manager with update checking, you're better off using one of the existing forks.

Feedback and contributions are always welcome!

[Squibeel]

I would live to say I restored this project from a local copy but the copy I have has a lot of stuff added on top (by Claude).
https://github.com/Squibeel/models-info

Basically added the ability to scan for updates, ignore updates, hide buttons, card size no longer ignore size set in settings and probably some other stuff I can't recall. Unless arc and ciel or civitai break something I plan to leave it as is.

[zeydsama]

Prompt save: sd-forge-prompt-save

What it does: Simply save your prompt, just like you would when using a lora.

[seti9585]

Forge Neo does not have any high-order ODE samplers.
So I ported two projects to work on Forge Neo: "ComfyUI-ODE", which is likely the basis for reForge's built-in ODE sampler, and "ComfyUI-RK-Sampler", a separate Runge-Kutta sampler.
Verification documents confirming that both work correctly from a mathematical standpoint have been uploaded to each repository.
I am also working on Forge Neo-compatible versions of MaHiRo, TCFG, FreSca, and SkimmedCFG.

https://github.com/seti9585/sd-webui-tde-sampler
https://github.com/seti9585/sd-webui-rk-sampler

[BulbulLeung]

Re-layout WebUI Forge NEO for Ultra-Wide Monitors.
I know zero programming, but I used Cursor AI to fix the layout pain points in Forge NEO for wide-screen monitors. No more endless vertical scrolling or constantly switching between the Generation and Extra network tabs!
https://civitai.com/articles/31164/ultra-wide-monitor-friendly-layout-extension-for-webui-forge-neo

https://github.com/BulbulLeung/forge-split-extra-networks
Split Extra Networks Tab:
Splits Checkpoints/Loras to the right panel.
Features: Built-in Output Browser (with batch delete), Wildcard tab, Prompt manager with Local AI Translation, and direct Lora weight editing controls.

https://github.com/BulbulLeung/sd-generation-layout
Custom Generation Layout:
Redesigns the core generation layout to fit Ultra-Wide screens perfectly, placing parameters and previews side-by-side.

[eduardoabreu81]

Iris-3B for Forge Neo

An extension that runs Iris-3B inside the normal txt2img tab - same checkpoint list, VAE / Text Encoder selector and Generate button. Iris paints every pixel directly, so there is no VAE to pick. It changes no Forge Neo file, and it stays disabled if your Forge Neo is missing something it needs.

• Runs on 8 GB GPUs - tested on an RTX 2070, with the bf16 and the lighter INT8 checkpoints
• Uses the same Qwen3-VL 4B text encoder as Krea 2
• Turbo LoRA support - a 1024x1024 image in 4 steps, about 4 seconds on an RTX 2070, with an Iris Turbo schedule for it
• Restore x4 in img2img - restores and enlarges a picture four times, with the upscaler LoRA or the official upscaler
• Depth in img2img - a depth map of any picture, ready for depth ControlNets
• An iris UI preset with the Iris DPM++ 2M sampler, the one the model was made with

Iris 3B for Forge Neo

Needs an up-to-date Forge Neo. Install from URL and restart. Examples with their prompts and settings are in the wiki.

[yuki-lang89]

hi @eduardoabreu81 , may i ask, since you has a capability to build a big and faster forge community extension, is it possible to you to rewrite forge architecture to rust? (*sorry its kinda crazy question, but i'm curious does it worth-it and possible) <3

[eduardoabreu81]

Hey @yuki-lang89 ! Technically possible, but I don't think a full Rust rewrite would be worth it.

Haoming has already put tremendous effort into modernizing Automatic1111 for Forge Neo. Rewriting everything would mean rebuilding much of that work and potentially breaking compatibility with existing extensions and the Python/PyTorch ecosystem.

Besides, most inference work already runs on the GPU, so Rust wouldn't necessarily improve generation speed.

A hybrid approach makes more sense: preserve what works and introduce Rust only where it brings measurable benefits. :)

[zeydsama]

the memory management from rust is hella usefull, especially for mobile usage

[eduardoabreu81]

Forge Neo Extensions Index

An index for the Available tab that brings the Forge Neo ecosystem in - the default one (AUTOMATIC1111) almost never receives new entries, so most Neo extensions never show up there. It carries the classic list, revalidated with everything that no longer exists removed, plus 54 Forge Neo extensions from this thread.

• Every Neo extension is marked with a 🟢 in the name and with a forge-neo tag, so the ecosystem is easy to spot and easy to filter
• To list only the Neo extensions, tick the forge-neo tag and switch Showing type to show
• Go to Extensions, Available and Load from: https://raw.githubusercontent.com/eduardoabreu81/forge-neo-extensions/main/index.json to see the new index
• The classic list, revalidated: every repository checked, dead and archived ones removed
• Extensions you already have keep showing as Installed, as usual
• Updated monthly

Forge Neo Extensions Index

[eighty-griff]

How do you change that default URL or is it even possible?

[AMGarkin]

https://github.com/[eduardoabreu81/forge-neo-extensions](https://github.com/eduardoabreu81/forge-neo-extensions#make-it-the-default)#make-it-the-default

Forge Neo reads the index URL from the WEBUI_EXTENSIONS_INDEX environment variable. Add this line to webui-user.bat, before call webui.bat, and the Available tab loads this list by default, even after Forge updates:
set WEBUI_EXTENSIONS_INDEX=https://raw.githubusercontent.com/eduardoabreu81/forge-neo-extensions/main/index.json
On Linux or macOS, use export WEBUI_EXTENSIONS_INDEX=... in webui-user.sh.

[eduardoabreu81]

FLUX 3 Image Klein for Forge Neo

An extension that runs FLUX 3 Image Klein inside the normal txt2img and img2img tabs - same checkpoint list, VAE / Text Encoder selector and Generate button. It changes no Forge Neo file, and it stays disabled if your Forge Neo is missing something it needs.

• Runs on 8 GB GPUs - tested on an RTX 2070 with the INT8 files: a 1024x1024 image in about 25 seconds with the 16-step distilled model
• The distilled and the base model, in BF16, INT8 ConvRot and NVFP4 (RTX 50) files from Black Forest Labs' ComfyUI repository
• Uses the same Qwen3-VL 4B text encoder as Krea 2, and the FLUX.2 autoencoder
• Works on GPUs without bf16 (RTX 20xx) - the few values that outgrow float16 are kept in range, with no slowdown
• Reference images - edit a picture in img2img, or combine up to 10 with ImageStitch Integrated
• Layouts and edits by bounding boxes in the prompt - place, add, move or remove things by their boxes
• LoRA support, and a flux3 UI preset with the FLUX.3 schedule

FLUX 3 Image Klein for Forge Neo

Needs an up-to-date Forge Neo. Install from URL and restart, then place the files from black-forest-labs/flux-3-image-klein-comfy in the usual folders. Examples with their prompts and settings are in the wiki.

---

Regional Boxes for Forge Neo

A panel in txt2img and img2img to draw boxes with the mouse and tell the model where things go - no coordinates to type, no image editor. Each model reads boxes in its own way, so the panel writes them in the format it was trained on:

• FLUX 3 Image Klein - coordinates in the prompt, in Black Forest Labs' layout format: place things in a new picture, or add, move and remove them in an edit
• Qwen-Image 2.1 - marks on the picture: draw over the img2img input and say what to do inside each box; the panel writes a note per box and draws the boxes on a copy of the input when you generate (the "sketch marks" technique)
• Auto picks the format by the loaded model; the input picture shows behind the frame in img2img
• It only writes the prompt and, for Qwen-Image 2.1, draws on the input - nothing of Forge Neo is changed

Regional Boxes for Forge Neo

Works with the FLUX 3 extension above, and with Forge Neo's own Qwen-Image 2.1 support or the Qwen-Image 2.1 extension. For models without boxes of their own (SD 1.5, SDXL, Anima, Z-Image, Krea 2), Forge Couple and Regional Prompter are the tools to use.

[yuki-lang89]

woah.. so many great open weight model lately.. also, you might be interested a flux-level image model, OpenOSS model.
https://huggingface.co/OpenMOSS-Team/MOSS-Image-SFT
https://huggingface.co/OpenMOSS-Team/MOSS-Image-Base
https://huggingface.co/Comfy-Org/MOSS-Image

[eduardoabreu81]

woah.. so many great open weight model lately.. also, you might be interested a flux-level image model, OpenOSS model. https://huggingface.co/OpenMOSS-Team/MOSS-Image-SFT https://huggingface.co/OpenMOSS-Team/MOSS-Image-Base https://huggingface.co/Comfy-Org/MOSS-Image

Maybe... in 15 minutes? =D

[yuki-lang89]

holly moly..

[eduardoabreu81]

Regional Boxes 0.2.2 - now with MOSS-Image layouts

MOSS-Image can take its whole prompt as a JSON layout (OpenMOSS's user guide): a caption for the scene, then one item per thing or piece of text, each with its own description, the exact words for text, and a box in pixels counted from the center of the picture. It is how you put a title at the top and a subject in the middle - but writing it by hand means working out every box's pixels yourself.

Regional Boxes now writes it for you: draw the boxes, fill in what each one holds, and the layout lands in the prompt.

• Describe the scene in the prompt, then draw a box for each thing and each piece of text
• For each box, pick object, character or text, say how it looks, and type the exact words for text boxes (Enter for a new line)
• Write to prompt turns it all into MOSS-Image's layout, the way OpenMOSS's own tools write it - your description becomes the layout's caption
• Move a box by its name tag, resize it by its corners, or center it with ⇔ / ⇕ - dragging anywhere else draws a new box, even inside another one
• Each box shows its position in pixels, and you can type a number to set an edge exactly
• Read from prompt loads a layout back - yours or one you pasted - to adjust it
• The layout follows the picture size - change Width and Height, or switch img2img to Resize by, and the boxes move to fit on their own
• Works in img2img - draw over the input picture and describe the new version: things keep their place, while captions and words can change
• LoRA tags stay outside the layout, and Auto picks the format by the loaded model: FLUX 3, Qwen-Image 2.1 or MOSS-Image

What the panel writes for the poster above

{"summary":{"meta":["poster","flat illustration"],"canvas":"single","caption":"A minimalist winter travel poster with a pale blue background, a white arctic fox sitting in fresh snow, a bold dark-navy title at the top and a small subtitle at the bottom. Flat vector illustration."},"items":[{"label":"text","caption":"Large bold dark-navy uppercase title, centered","text":"ARCTIC FOX","bbox":[-538,-358,-376,358]},{"label":"object","caption":"A white arctic fox sitting in fresh snow, flat vector style","bbox":[-289,-234,424,316]},{"label":"text","caption":"Small thin navy subtitle, centered","text":"FORGE NEO","bbox":[474,-250,560,250]}],"bbox_coord_space":"abs_centralized","bbox_format":"yxyx"}

Each bbox is [top, left, bottom, right] in pixels from the center of an 832x1248 picture.

_{txt2img, then the same boxes in img2img with a new title and an arctic fox}

Needs MOSS-Image for Forge Neo 0.1.2, which also fixes layouts with sd-dynamic-prompts installed - it was stripping the { } out of the JSON before the model read it.

[hirorohi03]

A version supporting MOSS-Image, forked from @Haoming02's sd-forge-negpip, has been released.
https://github.com/flyfront/sd-forge-negpip

[eduardoabreu81]

MOSS-Image for Forge Neo

An extension that runs MOSS-Image inside the normal txt2img and img2img tabs - same checkpoint list, VAE / Text Encoder selector and Generate button. It changes no Forge Neo file, and it stays disabled if your Forge Neo is missing something it needs.

• Sharp text in the picture, in English and Chinese - put the exact words in quotes in the prompt
• The SFT and the Base model, in the INT8 ConvRot files from Comfy-Org's repository - tested on an A40 and an RTX A6000: a 1024x1024 image in about a minute at 50 steps
• Brings its own port of the Qwen3.5 9B text encoder, which Forge Neo does not have yet, and uses the FLUX.2 autoencoder
• Layouts by boxes - write the prompt as MOSS-Image's JSON layout and each thing goes in its box
• Image-to-image - change a picture's details while keeping the scene
• Upstream's guidance, which skips block 22 in the negative pass, with a setting to change it
• A moss UI preset with the MOSS schedule - and the values you save for the preset in Settings are kept

MOSS-Image for Forge Neo

Needs an up-to-date Forge Neo. Install from URL and restart, then place the files from Comfy-Org/MOSS-Image in the usual folders and restart once more, so the text encoder shows up. Examples are in the README.

[tusharbhutt]

@eduardoabreu81, you are amazing! ever thought of partnering up with @Haoming02 officially?

My dream is that someone updates Reactor to the version ComfyUI users enjoy: https://codeberg.org/Gourieff/comfyui-reactor-node

I have zero skills myself :)

[tusharbhutt]

Thanks so much! I will have a look at it and let you know how it goes on this side!

[Blueblizard]

@eduardoabreu81 Great work :)

[tusharbhutt]

@eduardoabreu81 it is an amazing extension! No issues so far on a bog standard 4060TI/16.

Question though: to create a model, can we also use any swapper, or is it insight128 only?

[eduardoabreu81]

Thanks, great to hear it runs well on the 4060 Ti!

Any swapper. A Visage face model doesn’t belong to a particular swapper: it only stores the averaged ArcFace identity of your source images. When you swap, each model converts that identity into what it expects, so the same face model works with inswapper, ReSwapper, HyperSwap, AlphaFace, GHOST, SimSwap and HiFiFace. Build it once and switch swappers freely. ReActor face models (.safetensors) load too, since they store the same kind of embedding.

One tip: likeness varies between swappers with the same face model, so if one doesn’t look right, try another (HyperSwap or inswapper are good starting points) and check the identity score.

[tusharbhutt]

Thanks, it is as I suspected. I was just curious if the higher resolution swappers are "better" than good ol inswapper128, or if that swapper has some other tricks to get better results dispute having 1/4 the resolution.

[sca-285]

Hi! I'd like to share some extensions I've forked, adapted or built for the Stable Diffusion WebUI Forge family ecosystem.

Some are adaptations of existing projects for better compatibility, while others are my own tools for prompting, image processing, and workflow customization.

Forks and adaptations

• [Photopea Embed for ReForge / Neo](https://github.com/sca-285/sd-webui-photopea-embed-ReForge-Neo)
• [Lobe Theme Redux](https://github.com/sca-285/sd-webui-lobe-theme-redux)
• [Impact ADetailer Neo](https://github.com/sca-285/sd-Impact-ADetailer-Neo)

Other extensions

• [Prompt Vault](https://github.com/sca-285/sd-webui-prompt-vault)
• [Kitchen Samplers](https://github.com/sca-285/sd-kitchen-samplers)
• [Auto Color Corrector](https://github.com/sca-285/sd-forge-auto-color-corrector)
• [Optical Realism](https://github.com/sca-285/sd-forge-optical-realism)
• [Digital Mastering](https://github.com/sca-285/sd-webui-digital-mastering)
• [Forge Notebook](https://github.com/sca-285/sd-forge-notebook)
• [Steer CFG Attention Guidance](https://github.com/sca-285/sd-steer-cfg-attention-guidance)
• [LORA Mixer](https://github.com/sca-285/sd-forge-neo-LORA-mixer)

I'd like to share these projects with the community in case they're useful to other users.

Thanks for maintaining the project!

[Haoming02]

Don't post/advertise, especially not in README, about 18+ features

I don't want my repo banned

[eduardoabreu81]

FIBO 2 for Forge Neo - Extension

An extension that runs Bria's FIBO 2.0 (Dev and Edit Turbo) inside the normal txt2img and img2img tabs, with the checkpoint, preset and Generate button you already use.

• Text-to-image from a structured JSON caption (Bria's VGL) or plain text
• Placement: each object's location and relative_size in the caption puts it where you say
• Text rendering with text_render, accented letters included
• Editing with Edit Turbo in img2img: write the change and set Denoising strength to 1
• More pictures through ImageStitch: <image_1>, <image_2>...
• Masked edits with Inpaint
• Turbo LoRA from Comfy-Org: Dev in 4 steps
• INT8 and GGUF checkpoints, with the FLUX.2 VAE and the Qwen3-VL 4B text encoder
• fibo2 UI preset with Dev's settings
• Changes no Forge Neo file

On an RTX 2070 (8 GB) with INT8 at 1024x1024: Edit Turbo takes about 40 s (4 steps), Dev about 5 min (50 steps).

For edits, use about 1 megapixel in the picture's proportions with Crop and resize. The console tells you when the size is off.

Models: Comfy-Org/Fibo-2.0. The weights are CC BY-NC 4.0, non-commercial only.

https://github.com/eduardoabreu81/fibo2-forge-neo

Needs an up-to-date Forge Neo. Install from URL and restart.

[SillySilk]

https://claude.ai/artifact/UHi4RarA5tS1TqK8xV8t4K A Claude artifact with all viable extensions listed so far, or at least I hope. <https://www.avast.com/sig-email?utm_medium=email&utm_source=link&utm_campaign=sig-email&utm_content=webmail> Virus-free.www.avast.com <https://www.avast.com/sig-email?utm_medium=email&utm_source=link&utm_campaign=sig-email&utm_content=webmail> <#DAB4FAD8-2DD7-40BB-A1B8-4E2AA1F9FDF2>

…

On Thu, Oct 22, 2026 at 5:41 PM digitalhunters ***@***.***> wrote: Thanks, great to hear it runs well on the 4060 Ti! Any swapper. A Visage face model doesn’t belong to a particular swapper: it only stores the averaged ArcFace identity of your source images. When you swap, each model converts that identity into what it expects, so the same face model works with inswapper, ReSwapper, HyperSwap, AlphaFace, GHOST, SimSwap and HiFiFace. Build it once and switch swappers freely. ReActor face models (.safetensors) load too, since they store the same kind of embedding. One tip: likeness varies between swappers with the same face model, so if one doesn’t look right, try another (HyperSwap or inswapper are good starting points) and check the identity score. — Reply to this email directly, view it on GitHub <#1128?email_source=notifications&email_token=AHN7U77KP6HHO7OXXJM2AQ35VKSMNA5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCOJRGM2TGMZQUZZGKYLTN5XKU43VMJZWG4TJMJSWJJLFOZSW45FMMZXW65DFOJPWG3DJMNVQ#discussioncomment-19135330>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AHN7U77BTED6VULGZE7VKRD5VKSMNAVCNFSNUABIKJSXA33TNF2G64TZHM4DIMJQGYZDAMJWHNCGS43DOVZXG2LPNY5TCMBQGU2DOMBZUF3AE> . Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS <https://github.com/notifications/mobile/ios/AHN7U74OVYB7JJMKQ7ZQD435VKSMNA5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCOJRGM2TGMZQUZZGKYLTN5XKU43VMJZWG4TJMJSWJJLFOZSW45FKMZXW65DFOJPWS33T> and Android <https://github.com/notifications/mobile/android/AHN7U77ETLM3W2644XJUH5T5VKSMNA5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCOJRGM2TGMZQUZZGKYLTN5XKU43VMJZWG4TJMJSWJJLFOZSW45FOMZXW65DFOJPWC3TEOJXWSZA>. Download it today! You are receiving this because you are subscribed to this thread.Message ID: <Haoming02/sd-webui-forge-classic/repo-discussions/1128/comments/19135330@ github.com>

[AMGarkin]

ilyq's sd-webui-florence-2 extension adapted for Forge Neo:

https://github.com/AMGarkin/sd-webui-florence-2-neo

It is an image captioning tool based on the Microsoft Florence-2 model.

[eduardoabreu81]

Krea 2 Edit V2 for Forge Neo

Forge Neo already runs Krea 2 edit LoRAs on its own (Settings → Stable Diffusion → [Krea2] Enable Reference), but its edit path predates the V2 release of the Krea 2 Identity Edit / General Edit LoRAs, so it has none of the controls they were trained with. This extension adds them on top of Forge's own path - an accordion in txt2img and img2img, ported from the ComfyUI-Krea2Edit 2.0 nodes:

• Subject / scene boost (ref_boost) - how hard the result follows each picture: 2 for Identity Edit V2, 4 for General Edit V2
• Placement - aligned, beside or bag of features, for the subject and the scene; scene aligned + subject beside for composites
• Fit - the V2 geometry: pictures of other proportions are fitted in pixels instead of resized as latents (a portrait photo in a landscape result also costs fewer tokens: 6 minutes instead of 18 in my composite test)
• Grounding size, and the pictures shown with the negative prompt too
• The img2img picture comes first, the ImageStitch ones after it; the last picture is the subject
• Changes no Forge Neo file; with the accordion closed, Forge's own path runs untouched

Tested on an RTX 2070 (8 GB) with the _r64 LoRAs: an img2img edit, a person put into a scene, and a txt2img restage from one photo - where Forge alone drew the person twice.

Krea 2 Edit V2 for Forge Neo

Needs an up-to-date Forge Neo. Install from URL and restart.

---

The Forge Neo Extensions Index is updated too, with the extensions shared here since October 16: Visage, MOSS-Image, FIBO 2, Regional Boxes, flyfront's NegPiP, AMGarkin's Florence-2, abzaloff's ReActor, sca-285's eleven extensions and Krea 2 Edit V2 - 86 Forge Neo extensions in all. Load it from Extensions → Available: https://raw.githubusercontent.com/eduardoabreu81/forge-neo-extensions/main/index.json

[eduardoabreu81]

INT4 W4A8 for Forge Neo

Many of the small INT4 checkpoints made for 8 GB cards - the INT4 ConvRot ones, for Krea 2, MOSS-Image and others - ask Forge Neo to run both the weights and the activations in 4 bits (W4A4). comfy-kitchen has a second path for the same files that keeps the INT4 weights and runs the activations in INT8 (W4A8). This extension switches the loaded model to it - same file, no conversion, practically the same VRAM:

• Works with any INT4 ConvRot checkpoint, whatever the model; INT8, FP8, NVFP4, GGUF and BF16 checkpoints are not touched
• Settings → Stable Diffusion → [INT4 W4A8] Activations of INT4 ConvRot checkpoints: INT8 (W4A8), the default, or As stored (W4A4) - applied at the next generation, with no restart and no checkpoint reload
• The switch happens in memory, on the model already loaded - the checkpoint file is never changed
• A little slower: 15 s instead of 13 s for an 896x1152 MOSS-Image picture at 8 steps on an RTX 2070
• Changes no Forge Neo file

Tested on an RTX 2070 (8 GB) with Muse by Stable Yogi v4.0 INT4 and KREA 2 Turbo v10 INT4 (Krea 2), and Realism by Stable Yogi v1.0 INT4 (MOSS-Image).

INT4 W4A8 for Forge Neo

Needs an up-to-date Forge Neo. Install from URL and restart. If you use MOSS-Image for Forge Neo, update it too: this setting used to live there and moved here.

[eduardoabreu81]

Anima Reference for Forge Neo

An extension that puts your character from one picture into any scene with Anima, in the normal txt2img and img2img tabs - no panel: add the method's LoRA tag to the prompt and the picture to ImageStitch Integrated, as with Anima Edit. Hair, eyes, outfit and accessories come from the picture; pose, place and framing come from the prompt - no character name, artist tag or appearance tags needed, so it works for original characters too. It changes no Forge Neo file, and it stays disabled if your Forge Neo is missing something it needs.

• Two community methods, each with its authors' own weights: ElectronGirl's Ref Adapter (with its own reference guidance and APG) and AnticPan's Role Ref KV
• The tag's weight is the strength; Ref Adapter's reference guidance lives in Settings → Stable Diffusion → Anima Reference
• Works with Anima base v1.0, turbo checkpoints at 8 steps and CFG 1, and the Anima 2.9B models
• Hires. fix and img2img work too
• Their weights run only inside the reference passes - they never load as ordinary LoRAs

Anima Reference for Forge Neo

Needs an up-to-date Forge Neo. Install from URL and restart, then put the weights of the method you want in models/Lora (Ref Adapter, Role Ref KV - the file in its comfy folder). Side-by-side comparisons of both methods are in the README. It is also in the Forge Neo Extensions Index.

[mirabarukaso]

Forge Turing FP16

Forge Turing FP16 is now an extension, not a core patch. #1599 was closed for that reason.

Install it from Extensions → Install from URL:
https://github.com/mirabarukaso/forge-turing-fp16

Leave the branch and the local directory name empty. Forge clones it to extensions/forge-turing-fp16. Then run the installer in that repository and keep --skip-install in the launch arguments, so startup does not replace the Turing SageAttention build.

It adds two Settings → Optimizations switches for Turing (sm75):

• forge_compute_fp16: run the loaded UNet in fp16. Stored weights stay as they are.
• forge_sage_fp16_cast: cast attention q/k/v to fp16 for SageAttention and cast the output back.

Both are available through API override_settings. Later GPUs may see no speedup.

Latest check on an NVIDIA TITAN RTX, 24 GB. Model moss_image_sft_bf16.safetensors, stored as fp32, with flux2-vae and qwen3.5_9b_int8_convrot. 8 steps, LoRA 8steps_lora_latest at 0.7. SageAttention 2 is active through its Triton kernels (SageAttention CUDA extension unavailable; using Triton kernels). Times are the sampling loop.

Switches	Sampling	UNet
Both on	2.88 s/it (23 s)	19387 MB, fully loaded
Both off	10.73 s/it (1 m 25 s)	15907 MB loaded, 3480 MB offloaded, 49 lowvram patches

That is about 3.7×. The off state is the same weight set split across GPU and CPU: 15907 + 3480 = 19387 MB. fp32 compute reserves more VRAM, so the loader offloads part of the UNet. It is not a second copy left behind by the fp16 run. POST /sdapi/v1/unload-checkpoint between the two requests does free the previous weights, and cuda.system.free rises. The next load still offloads, because Forge sets the weight budget from computation_dtype (here 19403 MB usable with fp16, 16221 MB with fp32). An int8 UNet that already fits does not get this split.

An earlier check on the same card, also 8 steps and the same LoRA, is in #1599:

Model	Switches off	fp16 + xformers	fp16 + SageAttention 2
moss_image_sft_int8_convrot (5784 MB)	5.81 s/it (46 s)	2.20 s/it (17 s), 2.6×	1.68 s/it (13 s), 3.5×
moss_image_sft_bf16 (stored as fp32)	13.85 s/it (1 m 50 s), 3480 MB offloaded	2.62 s/it (20 s), 5.3×	2.53 s/it (20 s), 5.5×

On the int8 model the UNet stays at 5784 MB, so that speedup is not from leaving lowvram. Against xformers with the same fp16 switches, SageAttention 2 was about 1.3× on int8 and about the same on the fp32 model.

forge_sage_fp16_cast does not change linear compute. It only runs inside SageAttention. If SageAttention is not the active backend, the switch logs a warning and does nothing.

Same idea, measured in ComfyUI

Same GPU: TITAN RTX, sm75, no hardware bf16. SageAttention on that card uses the Triton path.

Full write-up: https://github.com/mirabarukaso/ComfyUI_MiraSubPack#sagefp16switch-benchmark-moss-image-sft-bf16-vs-int8-titan-rtx-sm75

[eduardoabreu81]

Thanks for the numbers, and for the original finding in ComfyUI_MiraSubPack: my extension started from it.

I took a smaller route to the same speed-up: Turing FP16 for Forge Neo. Install from URL and restart; no installer, nothing changed in the venv, no --skip-install.

Instead of patching the loaded UNet, it hooks the two places where Forge picks the dtypes while loading (memory_management.unet_dtype and inference_cast) and changes only the float32 fallback that happens for lack of bfloat16:

• compute: float16 instead of float32
• storage: a BF16 or FP16 file keeps its own weights instead of being widened to float32 (the "stored as fp32" in your Titan run)

GPUs with bfloat16, real float32 files and --fp32-unet are left alone. Since the dtype is decided before the model is built, Forge sets up manual cast, LoRA and ControlNet on its own. It is on by default, with a switch under Settings → Stable Diffusion. SageAttention is not part of it; your extension is still the way to get Sage on Turing.

RTX 2070 (8 GB), Forge Neo's built-in MOSS-Image, 1024x1024, 8 steps, CFG 1, whole request with the model already loaded:

Model	Off	On
hypermosslush turbo (INT8 merge)	32 s	16.6 s
Refined Realism v2 (BF16 merge)	78 s	24 s

On 8 GB, Forge already stores a 5B BF16 model in bfloat16 because float32 does not fit, so the storage part only kicks in on larger cards, which I can't test here. If you have a moment, a run of v0.2.0 on your Titan RTX with moss_image_sft_bf16 would be very welcome: the console should show storage: torch.bfloat16, computation: torch.float16, with about half the UNet memory of your fp16 run.

[mirabarukaso]

Thanks for the numbers, and for the original finding in ComfyUI_MiraSubPack: my extension started from it.

I took a smaller route to the same speed-up: Turing FP16 for Forge Neo. Install from URL and restart; no installer, nothing changed in the venv, no --skip-install.

...

Post in your repo, in short, it's almost the same. Good job!
I was working on another project this week.

[masterofobzene]

Forge neo notification sound 🔔

You may know that if you use the standard notification configuration, if you minimize the browser the sound will not be played until you have focus on it again. This super small extension will play your 'notification.mp3' sound even when the browser is not in the forefront.

https://github.com/masterofobzene/sd-forge-neo-notification-sound

[eduardoabreu81]

Idol for Forge Neo

An extension that turns one picture of an object into a 3D Gaussian splat with Idol, inside the normal img2img tab: same checkpoint list, VAE / Text Encoder selector and Generate button. One picture goes in; the whole object comes out, including the sides the picture never showed. It changes no Forge Neo file, and it stays disabled if your Forge Neo is missing something it needs.

A flat anime drawing works too, with limits: its lines and painted shadows become part of the object's colour, and the sides it does not show come out less convincing.

• .ply, .spz and .glb - choose what to save under Save as; the .spz is about a fifth of the size of the .ply
• Idol Turbo - 5 steps instead of 50, softer details; its LoRAs download by themselves the first time you tick it
• Turn the finished object on X, Y and Z without generating it again - no GPU, and the original stays as it was
• Background removed for you, unless your picture already has transparency
• A 3D viewer in the panel, and an idol UI preset
• The same seed gives the same object, byte for byte

Rendered figures and generated images give the best results. The unseen sides are the model's guess, and objects can arrive turned sideways.

The .glb mesh is a draft: fine detail goes, the surface ripples, and anime faces drawn with lines do not survive it. Use the .ply or the .spz for that style. The Examples show this too.

This is a first preview, tested in Forge Neo 2.31 on one machine with a handful of objects. Real photographs and busy backgrounds are not tested yet.

Idol for Forge Neo - guides in the wiki

Needs an up-to-date Forge Neo. Install from URL and restart, then put the four Idol files in the usual folders (links in the README, no Hugging Face account needed).