Forge Announcement (June 8)

[lllyasviel]

Hi forge users,

Today the dev branch of upstream sd-webui has updated many progress about performance. Many previous bottlenecks should be resolved. As discussed here, we recommend a majority of users to change back to upstream webui (directly use webui dev branch or wait for the dev branch to be merged to main).

At the same time, many features of forge (like unet-patcher and modern memory management) are considered to be too costly to be implemented in the current webui’s ecosystem.

Forge will then be turned into an experimental repo to mainly test features that are costly to integrate. We will experiment with Gradio 4 and add our implementation of a local GPU version of huggingface space’ zero GPU memory management based on LRU process scheduling and pickle-based process communication in the next version of forge. This will lead to a new Tab in forge called “Forge Space” (based on Gradio 4 SDK @spaces.GPU namespace) and another Tab titled “LLM”.

These updates are likely to break almost all extensions, and we recommend all users in production environments to change back to upstream webui for daily use.

We invite a small group of users to stay here to test Gradio 4, since feedback and extensions for Gradio 4 are also necessary for upstream’s considerations or adaptations, with regard to gradio’s recent advancement in LLM interface and streaming system, image editors and displays, and Gradio sdk’s seamless integration of zero-gpu computation management system.

Finally, we recommend forge users to backup your files right now (or just change back to upstream webui if possible). If you mistakenly updated forge without being aware of this announcement, the last commit before this announcement is 29be1da

---

Update June 27:

Hi Forge Users,

Here are some updates regarding the recent announcement:

1. There are no code changes between June 8 and today (June 27). If something is broken, it is likely due to other reasons, not this announcement.

2. We will provide a “download previous version” page similar to most other desktop software. If you are not an advanced Git user, please do not add, modify, or delete program files.

3. The “break extensions” refers to the extensions’ Gradio version. This mainly affects compatibility with A1111 extensions. For newer Forge extensions, upgrading will only require a few lines of modifications related to Gradio calls. The logical patching API will not change.

4. We understand your concerns and hear your feedback, but please do not misinterpret our announcements as an implication to eliminate the repository (which is false).

We recommend users to back up their files because the code base may undergo major changes in a few weeks. Right now, users do not need to do anything. If you are a professional user in a production environment, after the update happens (which will be in several weeks later), we recommend using version 29be1da (which will soon be available on a “download previous version” page) or using the upstream webui if necessary. There is also a possibility that our updates will be seamless, with no noticeable errors, but this chance is relatively small.

Finally, please note that the repository is not being "re-oriented." The original purpose of Forge is to "make development easier, optimize resource management, speed up inference, and study experimental features." This announcement will soon bring Gradio4 and a newer memory management system, which aligns with the original purpose of Forge.

Forge

---

Update July 21:

Hi Forge Users,

Here are some updates regarding the recent announcement:

1. The "download previous version" is provided now.

Forge

[LumiWasTaken]

Yikes

[ghost]

Oof, F in the chat boyos

[PanduZen]

I just created a github account just so I can comment here:

This announcement has broken my heart, I have been using forge for months and I was hoping that someday you would come and say that you would continue to update it.

For me it is impossible to use A1111 on XL models, I have tried a thousand ways and the VRAM consumption and time is very high compared to forge, I don't know what kind of black magic you have used to create forge but it is a real pity that you abandon the development.

I have installed the DEV version of a1111 you are talking about with those “improvements” you are talking about.

But I still have the same problems as always, for a simple image in XL, with a lighting model of 8 steps, a1111 takes more than 2 minutes to create it, consuming all the VRAM, instead in forge, the same image with the same model, does not take more than 10-12 seconds, for things like this it is impossible for me to switch to A1111 in its current state, I do not know what problem I have that forge works much better.

I hope you change your mind and do not abandon the development, it is the best program without doubt for this kind of things, and hopefully someday you can continue updating it and helping those people who have a humble computer.

I don't know how difficult it will be to implement the VRAM and u-net improvements in a1111, but I sincerely believe that no matter how difficult it is, it should be the number 1 priority of the A1111 developer, it completely changes absolutely everything and I don't understand how at this point, after demonstrating that forge handles VRAM in a wonderful way, the dev of A1111 is not trying to implement this. Because it would be the biggest revolution for the program.

Sorry for my english and if you didn't understand what I said, I'm using a translator while I'm crying.

[YiFoneChen]

is shit, inflation is up, and lots of uncertain things are happening in the world.

Forge is open source...if you love it so much CONTENUE DEVOLOPING IT ON A FORK and quit crying like a baby.

What you said is like saying, "When you're sick and need surgery, you can just do it yourself because you can easily find a knife at home." Haha.

[Koitenshin]

I'm so glad Github saw reason and now allows us to block Edgelords like the above user.

[toyssamurai]

What you said is like saying, "When you're sick and need surgery, you can just do it yourself because you can easily find a knife at home." Haha.

Or, when you need a mortgage and no bank/no one is willing to lend you the money, you can just borrow it from yourself.

[tpcdaz]

is shit, inflation is up, and lots of uncertain things are happening in the world.

Forge is open source...if you love it so much CONTENUE DEVOLOPING IT ON A FORK and quit crying like a baby.

Try not being a cunt more people might like you.

[abzaloff]

This is very sad news to me. I believed and hoped until the last that you would continue this wonderful project. I like Forge very much and it will be very hard for me to give it up. My heart is broken...

[b-7777777]

R I P

[TetyLike3]

I just got a friend into using Forge, and to see this announcement made me feel a little annoyed, since his PC already can barely handle the size of SD, but atleast it would run unlike on A1111. I could be wrong, perhaps it runs on par with Forge now (the last time I used A1111 was probably in March), and besides I was already thinking of just letting him remote access my copy. ¯_(ツ)_/¯
Either way, I'll miss this project, it was great to play around with, and I'll probably never feel the joy of watching the sheer drop in VRAM usage and gen times with the optimisations in Forge.

[rterreereeeee]

no point of using automatic1111 if they don't implement VRAM optimization like in forge, in seems like the dev branch did nothing in that regard

[Bullseye110]

no point of using automatic1111 if they don't implement VRAM optimization like in forge, in seems like the dev branch did nothing in that regard

Exactly...

[Panchovix]

I compared forge vs A1111 on the dev branch, and A1111 seems to be generally faster on SDXL on a RTX 4090, BUT without using loras.

The moment you add a LoRA, A1111 takes a good while to start to do inference. Then it's the same or faster. Not sure what is happening on that side.

Anyways, really thanks for all your work on Forge.

[ShadowAimai]

Would you please reconsider?

[BNP1111]

Guys, don't make things difficult for him or put pressure on him. His return is more important than forge.

[Valentinbejan]

Ok, so if I understand correctly the speed optimizations from forge will be implemented in automatic 1111 2.0. What about VRAM memory optimizations? you said that unet-patcher and modern memory management will not be implemented currently. Will they ever be implemented in automatic 1111? I hope so, it's good that automatic 1111 is getting faster, but my main problem with automatic 1111 is that I don't have much VRAM ( I have a 1060 6gb). Could they be added in automatic1111 2.1 or a future version?🤔

[xhoxye]

If you want to continue using the old ver forge program, remember to delete the update script “update.bat”, avoid accidentally clicking on it

[ExissNA]

"update.bat" is the only one I'm aware of. Is there another? I use the standalone zip install, not the Git stuff.

[hoodady]

remove git pull

[metapea]

for later on, are their any instruction on rollback to 29be1da?
i got:
"HEAD is now at 29be1da7 pass options to cross attention class error: RPC failed; curl 28 Recu failure: Connection was reset fatal: expected 'acknowledgments' Press any key to continue"
and git reseted my COMMANDLINE_ARGS=, are there logs for with COMMANDLINE_ARGS= i can find?
Thankfully it was nothing right now, but my webui-user.bat was reset and constantly does now when clicking on update.bat unlike the few times i did update before.

[sgtpepper67]

remove git pull

from where?

[PladsElsker]

I think this is great news. I have paid some attention to the developpments in A1111 dev branch, the devs there have pulled some phenomenal improvements to inference speeds in the last few weeks. I'd like to think forge's incredible inference pipeline was the push in the back that they needed to actually go and address the issue.

I also think it's great that this repo will live on to test some other features that are "costly" to implement right now in A1111. This sounds like a good compromise.

Cheers to the team! Keep going guys, your work is useful to thousands of people :)

[gltchcom]

Is it possible to switch our install to the dev branch of A1111? Or do we need a brand new install? If we can switch, can someone give directions?

[LeF0URBE]

fresh install then cut and paste your model folder+ embeddings folder and reinstall your extensions one by one.(you can have problem by cutting them out).
this is what i would do anyway. i first have to check on VRAM consumption.

[BetaDoggo]

The reactions in this thread are disappointing. Lots of whining about a free tool which will continue to function just as it always has. There's no need to switch to anything else if the current version already works. Forge is a great project and I'm excited to see where it goes in the future.

[wywywywy]

Yea I'm a bit puzzled. This should be good news really.

A1111 has vastly improved, and lllyasviel is back to carry on experimenting (which may help upstream in the future). Plus we finally got the closure so many people have asked for.

We should be celebrating, no?

[ExissNA]

I think people are mostly upset about a lack of proper fixes to certain things in Forge. While Forge is amazing, it still needs some tweaking to be truly functionally competitive for certain things (broken controlnet primarily for me). I was thoroughly ready to continue on an old and dated version of Forge than to switch to other things because it functioned so much faster and more stable than the alternatives. If A1111 can function even remotely closely to the performance Forge has given me, then I would consider returning. Given the track record of my experiences though, I continue to expect failure and frustration. That's my experience, and I think that in many people's minds Forge is the only option, and it's now basically "dead." I believe that's the most probable cause of being upset.

[Kmr196]

Yea I'm a bit puzzled. This should be good news really.

A1111 has vastly improved

As long as it takes 3-4 minutes for one gen in A1111 and in forge it takes 30 sec, I wouldn't call it vastly improved

[drdancm]

If you work with the Extras Tab, things are much faster, than otherwise, and also much faster than prior to the 1.93 and 1.94 updates. Unfortunately not all extensions are available in the Extras Tab.

[9wow]

It's not going anywhere in the future because It's a crash dummy now. Read the post.

[BNP1111]

We don't care about a1111 or forge; we are just very happy that you are back, our old friend. We missed you

[Kmr196]

But he never left, he just ghosted forge for 3 months and worked on other stuff. For the majority of people this "return" will be the same as it just being still abandoned

[Vendaciousness]

Showing up after over 4 months of inactivity (zero bug fixes), just to break our installs, so we have to use 'git checkout 29be1da' & 'git reset --hard' to make them work again? You're a jerk. A highly gifted dev, initially helpful, but ultimately the most frustrating developer I've ever dealt with on Github.

You hate production artists, don't you? You love the MidJourney and FoOocus type users, who can only create with text. Your work is aimed at replacing us, not empowering us. I think you want art for the masses (no artists/everyone an artist), and you've never intended to cater to existing artists in any way. I feel like ControlNet became used by artists incidentally and you broke it and stopped developing it, possibly as a result.

I'm not gonna kiss your ass like everyone else here, since I'm giving up SD and Deforum because of you. Thanks for when your tools actually worked.

I wrote up bugs in your CN repo over a year ago that you've never implemented, like allowing Inpainting with any other model. You ought to give this and your other repos to people who will ACTUALLY WORK on them. You'd make a great contributor, with occasional pushes, but you can't manage all the work as the sole dev. Not when you show up once every few months.

Advice for anyone on Stable Forge: Switch back to A1111. They've probably implemented most or all of the good code from this midden heap.

Goodbye Illyasviel and Generative Art.

[LadyFlames]

sadly A1111 does not have any of the Forge Features i went back to test A1111 if it has changed but it hasnt its as slow as it used to be yea you might want to try Sd.Next or Invoke or Krita

[brknsoul]

@LadyFlames Did you try the dev branch? (git checkout dev && git pull)

[LadyFlames]

yea i did but since im using Stability matrix as a hub its kind of pain to Delete and reinstall the webui everytime if i want to switch branches im using both Forge in Dev2 and A1111 in Dev Branch

[brknsoul]

You don't need to do that, just install a new package, name it something else, and select the correct branch. Now you have a master branch and a separate dev branch

[Susanoo1337]

Sorry to break it to you, you're not a production artist.
You're not using a production level tool.
And if you run anything with auto-updates, you're really just asking for it, Illya doesn't deserve this vitriol.
Yes, the premise of Forge was a little over-hyped, but all things considered it was never really sold to us as anything but an experiment, if you'd just taken a second to look at how busy Illya is, this would not have come as a surprise to you.
Typing 'git checkout' really only takes a second so if instead of this rant you would have asked us for help, you'd be back up and generating your 20 devianart ai shitposts a day already.

I don't want to be mean to you but my bloodsugar is low and you are really being a hysterical lil girl here. You should also really have more than one install of StableDiffusion at a time, ESPECIALLY if you auto-update.

[E2GO]

It seems most of the people who commented here do not know what open-source development is...

[Koitenshin]

Or some of us just aren't really programmers.

[E2GO]

Indeed... But the thing that most of the stuff done for free to everybody. And the project is not closed - so anyone (who know what to do and have a will to do it) can make own branch from it and continue development...

[pjwaixingren]

controlnet is not supported by xyz plot, when will it be?

[scraed]

F***, we spend months to learn the forge backend in order to port our extension to forge. Now you say it “likely to break almost all extensions”. This is really making us 🤡

[scraed]

"Clone it and use it" makes me more looking like a 🤡. I'm part time extension developer. The reason I write extension for forge because I think forge is a decent interface with many users, which will help us share our research to others. If you say "Clone it and use it", we should have spent our time more wisely on comfy instead of supporting something that breaks all extensions after several month of its release.

PS, The extension I'm developing is very complex, as it alters how CFG scale works. Personally, I think forge is not "adjustment" but "replacement". It is more like comfy instead of A1111, especially about how it evaluate the Unet. Because we are not familiar with comfy, and forge didn't provide any developer guide, so we need to read the new backend line by line using our spare time to understand it. That is very painful.

[jtabox]

"Clone it and use it" makes me more looking like a 🤡. I'm part time extension developer. The reason I write extension for forge because I think forge is a decent interface with many users, which will help us share our research to others. If you say "Clone it and use it", we should have spent our time more wisely on comfy instead of supporting something that breaks all extensions after several month of its release.

PS, The extension I'm developing is very complex, as it alters how CFG scale works. Personally, I think forge is not "adjustment" but "replacement". It is more like comfy instead of A1111, especially about how it evaluate the Unet. Because we are not familiar with comfy, and forge didn't provide any developer guide, so we need to read the new backend line by line using our spare time to understand it. That is very painful.

The "clone it and use it" meant mostly that Forge is still functioning. So it's not like you're suddenly left without a host for your extension. And it'll continue to work for a while, as long as you don't update it to include any new functionality. Hopefully during this time you've learned enough to port it to ComfyUI. Not to mention that maybe someone decides to fork the current version of Forge and update it. But in any case, you still have time to consider your course of action and plan ahead.

Also personally, I doubt you're a 🤡 for having expanded your coding knowledge. Regardless of what happens with Forge, you've learned stuff that will probably be useful to you in the future, and that's not something anyone can take from you.

[scraed]

Of course I know forge is still functioning. The pain comes knowing that my code will break soon in the next version. Now the lllyasviel’s statement about “The logical patching API will not change”&“ only require a few lines of modifications related to Gradio calls”, makes me feel a bit better.

PS. I really hope the knowledge of forge will help us dealing with ComfyUI.

[Koitenshin]

@scraed There's already CHG implemented for ComfyUI, unless the implementation is wrong somehow.
https://github.com/redhottensors/ComfyUI-Prediction?tab=readme-ov-file#characteristic-guidance-prediction

[scraed]

OMG, great news. I don't event know about it.

[jtabox]

I think people here should chill the f out with their entitlement, it's irritating to say the least.

• This is an open source project, and its creator decided to give away their work to other people to use for free. They have literally zero obligations to keep pushing out fixes and updates. Absolutely no one gave any guarantees and/or promises that Forge would continue to receive support indefinitely.
• People here behave like entitled toddlers that throw a tantrum because their favorite toy just broke, going so far as attacking lllyasviel because they dared to decide to move on to something else. This is very ungrateful and honestly quite sh*tty. Get your head out of your behind, no one owes you anything in this case.
• You'll still be able to use Forge by using the current fork, it won't be deleted from your PC.
• Since this is an open-source project, maybe some other person will decide to continue and develop it. Though honestly, after this whole situation here, I'd be very cautious to take on something like this as a dev, since apparently people will take it like a life-long contract signed with blood.

[TheActualDonKnotts]

I see something more akin to mourning a great project than entitlement. I think you're a reactionary seeing what you want to see so you can make your pissy little finger wagging post. Where did you see people attacking lllyasviel?

[dan4ik94]

All the hate comes from the emotive side. When you put someone on good drugs and you decide suddenly to take them away of course there'll be pain and it'll take time to come to normal life. This is the most accurate comparison from my point of view. So the reactions from people in general are normal and expected.

Forge is/was a revolutionary tool for many and it hurts to see this kind of situation happening, that's all. Moreover when the SD future is very uncertain.

[jtabox]

I see something more akin to mourning a great project than entitlement. I think you're a reactionary seeing what you want to see so you can make your pissy little finger wagging post. Where did you see people attacking lllyasviel?

Sure buddy, you got me. If you've seriously scrolled through this thread and didn't see posts attacking lllyasviel for deciding to move on then that says more for you than anything else. I didn't say everyone is doing it, but whatever. You do you, champ.

[lllyasviel]

Hi Forge Users,

Here are some updates regarding the recent announcement:

1. There are no code changes between June 8 and today (June 27). If something is broken, it is likely due to other reasons, not this announcement.

2. We will provide a “download previous version” page similar to most other desktop software. If you are not an advanced Git user, please do not add, modify, or delete program files.

3. The “break extensions” refers to the extensions’ Gradio version. This mainly affects compatibility with A1111 extensions. For newer Forge extensions, upgrading will only require a few lines of modifications related to Gradio calls. The logical patching API will not change.

4. We understand your concerns and hear your feedback, but please do not misinterpret our announcements as an implication to eliminate the repository (which is false).

We recommend users to back up their files because the code base may undergo major changes in a few weeks. Right now, users do not need to do anything. If you are a professional user in a production environment, after the update happens (which will be in several weeks later), we recommend using version 29be1da (which will soon be available on a “download previous version” page) or using the upstream webui if necessary. There is also a possibility that our updates will be seamless, with no noticeable errors, but this chance is relatively small.

Finally, please note that the repository is not being "re-oriented." The original purpose of Forge is to "make development easier, optimize resource management, speed up inference, and study experimental features." This announcement will soon bring Gradio4 and a newer memory management system, which aligns with the original purpose of Forge.

Forge

[Koitenshin]

A Newer memory management system? What will it accomplish compared to what we have now? I can generate much higher resolution images on Forge compared to A1111's UI, but not as high as I can on ComfyUI.

And can we please get the Samplers and Schedulers split like in A1111?
Maybe add a user specific sampler & scheduler file that gets loaded but not modified by the code.
It will make adding more samplers & schedulers for testing extremely easy. I'm not sure if you've seen TCD, AYS or KL Optimal yet, but being able to add those would be nice.

[dan4ik94]

you can add AYS manually, there's a pending pull request.

[Koitenshin]

You simply didn't bother to read the thread. I know how to do it myself, and already did. I even provided proof the implementation is bad as 11 Sigmas stretched over 32 steps provides the incorrect image compared to giving the scheduler the right sigmas.

[koopa990]

I am a little nervous. Am I safe to update this?

[dan4ik94]

It's related to this commit: bfee03d

shouldn't break anything.

[koopa990]

Thanks

[Panchovix]

I'm wondering if it's possible to set a scheduler selection like this on Forge

[Panchovix]

But that makes some sampler miss some schedulers, like Restart.

[koopa990]

But that makes some sampler miss some schedulers, like Restart.

What's unique about restart?

[Panchovix]

I like to use it on hi-res fix steps since it was added into the webui. I like with AYS and it's not possible to use it on Forge. I tried to implement it but failed.

[Koitenshin]

Forge already has this feature. Karras and other types are combined into sampling methods.

Do you have any idea how annoying it is to add new schedulers and samplers to play around with? You have to modify 3 different files in Forge to add a single new scheduler compared to A1111's one file for a new scheduler that you can then use with any Sampler, or vice versa.

[Panchovix]

At the end I managed to apply it following some steps as the A1111 PR AUTOMATIC1111/stable-diffusion-webui@9aa9e98

I have this as a WIP https://github.com/Panchovix/stable-diffusion-webui-forge/commits/dev_schedulers/

[multianimus]

Bruh...

[HITTRAKKZ]

### it says forge will never add any Unnecessary Opinionated changes to the user interface.. it never said it would stay the same not changing at all.. and forge still works perfectly fine and i'm sure it will continue to run fine for months, maybe someone by then will combine a1111 with the best of forge, it's all open source. people should probably just be happy forge was made at all because its been half a year of people being able to run xl models that wasn't possible to run before forge and you can always stay on the extensions your on now using whatever checkout works with it at the moment.. and that's probably the worst case scenario

[C4l1b3r]

If it's about money, let me know what it would cost to fix the issue, I can surely donate

[iamwavecut]

F

[Myoko]

this is the reason we have been waiting for! Personally, I feel that Forge is more suitable for me compared to A1111.

[C4l1b3r]

Agreed and I have a high end video card and still prefer forge over A1111

[crsongbird]

Forge is fast, whatever magic you did in the inference pipeline is brilliant. I can run SDXL on my 1070ti.

The fastest GPU I can foresee being able to get my hands on in the next few years is an 11gb 2080ti so... ultimately, a tool like Forge will still be needed for lots of people. 11GB isn't enough for SDXL on A1111.

I appreciate whatever changes and work you did to this project. I'm still using Forge. I still get good inference speed with SDXL and up to 2 LORAs. I hope your changes get merged into A1111, I don't want to have to learn Comfy. It reminds me way too much of the Unreal4 Shader Editor lmao.

[dan4ik94]

Then you should check the reforge project, it's a great continuation
https://github.com/Panchovix/stable-diffusion-webui-reForge

[Koitenshin]

Another alternative is ComfyUI, which is where the Forge backend improvements came from. The interface leaves a lot to be desired though.

I can not only use SDXL models, but I can merge them with only a 3060TI (8GB of VRAM) & 32GB of RAM.

[crsongbird]

Then you should check the reforge project, it's a great continuation https://github.com/Panchovix/stable-diffusion-webui-reForge

OMFG this is cool!! Thank you!

[dan4ik94]

Another alternative is ComfyUI, which is where the Forge backend improvements came from. The interface leaves a lot to be desired though.

I can not only use SDXL models, but I can merge them with only a 3060TI (8GB of VRAM) & 32GB of RAM.

can't agree more. I use comfyui only for simple workflows like liveportrait and supir but still can't and probably won't make it my main tool. I can't imagine setting up and looking at 100 nodes on my 1080p monitor.

[gohan2091]

Is there a way I can switch Forge to ReForge? or do I have to install ReForge from the beginning?

[dan4ik94]

Is there a way I can switch Forge to ReForge? or do I have to install ReForge from the beginning?

There are instructions in the reforge repo, why asking here.

[lllyasviel]

Update July 21:

Hi Forge Users,

Here are some updates regarding the recent announcement:

1. The "download previous version" is provided now.

Forge