In other words, anyone managed to get the exact same results in Forge & SwarmUI with the same prompt data? (it's my own prompt data, so I know nothing is missing/wrong, etc).

What exactly needs changed? (I've already tried changing a bunch of more hidden away default settings to fix it, but it still wasn't reproducing the exact same image results).

I'm assuming it's different optimisations/way data is rounded or handled slightly different as a result, but there must be a way to get them reproducing the exact same results...?

ShadowCell A lot of it comes down to how the software ties the model to the settings mathematically. A different version of libraries used to handle tensors through the software can introduce more than enough difference to change the end result even if all the more obvious settings are the same. I'm not sure there 'must be' purely because they're not the same software.

But - there might be something about your specific use-case that we can help with. Without telling us what the actual content is, can you tell us more about the problem you're trying to solve? The solution to the underlying task you're trying to do might not be about having different software do the same thing.

I have...I dunno, 50k+ worth of generated images (last time I even checked, which was ages ago lol). I've always used multiple different back-ends to generate, but I've customised some of them so much with depricated add-ons & prerequisite dependencies (that I manually added in) to the point its become a total mess.

I wanted to consolidate it all; but obviously when I go to import image/prompt data....it's just not the same. Unfortunately, close enough, just isn't good enough ;( Even tiny changes just throw the image composition, faces, etc, off to the point it goes from an S tier image set (at least in my books), right down to "whatever, meh" when I want to make alternate versions from that data.

I loved generating variations of that 'perfect' dice roll image to see what else I could get....but it's just not the same when I import it elsewhere ;( I'm just really nostalgic about preserving them, so it's really disapointing to see my perfect prompt data I spent hours working on for a particular set of images just get utterly 'butchered' like that when I import elsewhere ;(

PS - I was in part using it to create self portrait oil paintings of myself/likeness, so that's why I really don't want to part with it ;(

PS - Why does Swarm/Comfyi seemingly look so junk by comparison btw.....?  Like, I can't be the only person thinking/feeling this way lol.  I know it processes things differently, & I always kind of roll my eyes when people say things like "new checkpoint sucks", etc, because they're just using the exact same prompts/methods, on something that works differently (& has to be used differently)..... but I've fiddled around tailoring things in Swarm/Comfyi, & it just never seems to produce outputs anywhere near the quality level of even Automatic111.... (which is ancient/practically obsolete)....

It's like there's some kind of secret sauce, voodoo magic going on in Forge that just makes outputs better.  Wish I knew exactly what it was; because I'd be more happy even just capturing this magic pixie dust, even if it still didn't replicate things exactly.  (& yeh, I know it handles/parses data differently, & tried changes to reflect that...but to little avail).  I've even tried absolutely cranking quality settings with comfyi, & it still ends up producing inferior outputs in terms of general style!

With Forge, I can often just spam out c'est magnifique sets with the same seed/prompt, & only very minor changes!  Like, it will spam out half a dozen S+ tier images in a row..... that just never seems to happen in Comfyi.  It's almost like it handles prompts/tokenization differently, or something.... (Tried changing VAEs, etc).  

There a comprehensive, authorative technical guide on exactly all the processing differences somewhere that's actually good?

To elaborate a little further - I've manually updated to latest CUDA/optimizations of things which aren't officially supported, etc, & it still keeps producing magic, inspite of all the unsupported changes, but Comfyi just 'doesn't seem to get it' quite the same way in terms of understanding style. Like, why does comfyi seem to produce way more derpy outputs with bad anatomy, boring facial expressions, etc, in otherwords? I'm sure most people would be happy with my outputs....but it's just not on the same level! It just never seems to generate 5+ random amazing images in a row like Forge, for instance. I seem to have to do way more wrangling to force it to produce really great images.