discussion Current state of SD Forge, "old" vs "new"

[TheLowEnder]

Well guys, I'm in need of advice.
While learning Forge I learned that the "new" version (from early august on) has broken many extensions, thus explaining (at least in part) my failures with regional prompter.
So I'm asking you, can you suggest a stable, fully working commit of "old" Forge, or do you recommend sticking with the "new" one, suck it up, and rely on the much slower A1111 when I need more control over my gens?

 

[User_124703]

I´ve got no clue. And dunno if @encassion or @Mr. Hyde or @Stuffy or @Strut use SD forge. (Sorry for calling you if not. )

 

[User_89876]

Sorry no.

 

[TheLowEnder]

Sorry no. My beast of choice is ComfyUI

Thanks anyway. ComfyUi it's next on my "to learn" things. Know that I might bother you in the near future.

Below is the old stable version I use. The extensions I have installed are working but I can't speak for all of them.

You must be registered to see links

The bleeding edge version is actually pretty buggy for me. The image previews break when I use the "Send to X tab" functions (img2img, ect.). I have the old and new versions installed and switch between them sometimes.

You must be registered to see links

Thank you very much. The only extension that I care for (at least for now) are adetailer and regional prompts. As long as those two, img2img and impainting work, I'm happy.

 

[Stuffy]

Thanks anyway. ComfyUi it's next on my "to learn" things. Know that I might bother you in the near future.



Thank you very much. The only extension that I care for (at least for now) are adetailer and regional prompts. As long as those two, img2img and impainting work, I'm happy.

Also a ComfyUI user.

Though I occasionally load up Stable Swarm (now called Stability-AI)

You must be registered to see links

I use that because it has comfy integrated and I can build a comfy workflow and convert it into a Stable-type interface. If I don't feel like looking at noodles.

But I mostly use that because it saves all the metadata on images, and all I have to do is click on an old image that I liked the settings for to reload everything including prompt, seed, model etc

 

[Stuffy]

Sure. For me it's been a rabbit hole of endless possibilities and permanent experiments. I'm glad to answer questions to the best of my (still limited) knowledge

I've been playing with ModelMergeSDXL recently, but haven't found a good info source on the individual blocks.
Latent has some vids I managed to get a little from but still have gaps.
Any recommendations?

 

[TheLowEnder]

Though I occasionally load up Stable Swarm (now called Stability-AI)

You must be registered to see links

I use that because it has comfy integrated and I can build a comfy workflow and convert it into a Stable-type interface. If I don't feel like looking at noodles.

But I mostly use that because it saves all the metadata on images, and all I have to do is click on an old image that I liked the settings for to reload everything including prompt, seed, model etc

Thanks, I will look into that, hopefully it won't be too taxing on my poor faptop (VND™) with only 8g VRAM.

Sure. For me it's been a rabbit hole of endless possibilities and permanent experiments. I'm glad to answer questions to the best of my (still limited) knowledge

It looks really interesting, the UI is amazing and way faster than A1111. Now I have to learn how to handle nodes and get the kind of control over the outputs I have on Forge/A1111.

 

[Stuffy]

Ha. First question, first miss. I never dove into model merging other than using basic model merge nodes for model and clip. But to test this or to even find improvements by adjusting the settings is too time consuming. So I simply stick with my favorite checkpoints from Civitai (ponyRealism and cyberrealisticPony)

I use both of those, mostly cyberrealisticpony v63, for its unique advantages, but its lacking in things like background, lighting, details, proportions. I usually toss in albedobaseXL for creativity, and realitiesedge or quantumfox.

If I can find more info on blocks I'll try to share

 

[User_3471]

Recent advancements in AI rendering and software render multi-threading make complex graphics rendering easier with less cpu/gpu load. Upgrading software has always been a game of balancing new technology with backwards compatibility. On the one hand you want the best leading edge software you can get but on the other hand you want to reach a wider customer base. At some point you have to ask yourself one question. Am I willing to sacrifice "quality" in the interest of "quantity"?

 

[TheLowEnder]

I use both of those, mostly cyberrealisticpony v63, for its unique advantages, but its lacking in things like background, lighting, details, proportions. I usually toss in albedobaseXL for creativity, and realitiesedge or quantumfox.

That's odd. I've been experimenting with cyberrealisticpony v65 for the past couple of days and I loved lighting and background. Not to mention that Adetailer is almost useless... Now I have to check your works to get an educated guess on your standards of quality compared to mines...

 

[Strut]

Below is the old stable version I use. The extensions I have installed are working but I can't speak for all of them.

You must be registered to see links

The bleeding edge version is actually pretty buggy for me. The image previews break when I use the "Send to X tab" functions (img2img, ect.). I have the old and new versions installed and switch between them sometimes.

You must be registered to see links

First sorry for the late reply @TheLowEnder Iv not be around much.

But yeah, as enc said

You must be registered to see links

was the last stable version that I used.

Both the stable one and latest ones have pros and cons, For example, I myself have found the latest one too generate pictures faster and maybe using a bit less system resources, while the old one, was more stable (until I messed somthing up and broke it and have been too lazy too reinstall it)