@MagicArt35 It's all good, feels like a very well taken care of model with a lot of work into it!

In his defense, or her defense, the last 3 models have used the same settings and seeds for the example pictures, so it's no what has changed, they seem like completely different models under the photoMovie umbrella.

@Zenyth True, but what even is that umbrella? Is it just random merges or is there some logic going on here?

@OppyLog There is, if you analyze the model's compositions newer models are updated versions of old ones. I stopped my tests after I realized what made the photoMovie models unique is lost after each update, because it becomes similar to what is being merged with, and this isn't specific to this model, it happened to EpicPhotogasm, Analog Madness, and even to RealisticVision and Deliberate 3, the best are the early models because they're the most unique, the most recent version make you look at the generations and think "where have I seen this before?", specially samefaces (DreamShaper8 is a big exception to this rule, I'd say, other models are similar to it, because when they do something similar, D8 will make it better).

And then we have models like YiffyMix or NuipeniMix (it took me a while to find them because they're sold as furry models, but in reality they perform great as general toon models) or MistoonEmerald (with 90s on the prompt!) and they are unique and blow my mind away , so I'm now searching for those unique models I'm missing instead of checking new versions of my favorite models (in this case, AmovieX, the first version of PhotoMovie, is still the best from this family in my book.)

@Zenyth thanks, now I have to check all these out. I've always been intrigued by the /trash or furry models because they are immensely capable and do some special stuff but have never had any luck when prompting them

@omenizer Oh, also, it's very important what version you use of them, for YiffyMix I wasn't impressed with earlier versions, the one that hit hard was v31, but then v32 dropped the ball, I made a post about it:

https://civitai.com/images/3140946

This isn't about Judy Hopps, the subject, but she's hard to get right, so a model that does it, does a lot of other hard things right too. Though my favorite thing is it doesn't have the sameface problem, each girl has her own face, and they don't look like generic danbooru, it reminds me of the Dalle 3's face variety, it's like a rose on a desert of sameness. Like... PinappleAnimeMix, another that looks unlike the others, though, my spectrum is limited to toon models because reality looks like reality.

@kubilayan There should be a way to prune these models down to 2.0GB without loss of quality, but it's not a trivial task as one has to download UIs and scripts and such. If you need a smaller filesize and don't mind it being slightly different, grab your favorite model that is already 2.0GB and on A1111 use SuperMerger's Weight Sum Train Difference with Alpha 1 and save it as fp16, so you'll end with a 2.0GB model that has been trained to give the outputs of these bigger models.

@Zenyth Thank you for tips :)

@kubilayan Just get the model toolkit extension, it's a matter of seconds to make a 2gb novel from it

@omenizer I have just done. There are only minor differences from the original model. But it can be tolerated. Thank you very much.

@kubilayan yes the results are slightly different but not worse

@Zenyth I would like to do this, but not set up with A1111, and no desire to try it. With the extreme wealth of great pruned models in any style, sadly I don't have a reason to try one of these unpruned models ever again. A bit of a shame on this one, looks kinda nice.

What error are you getting? If it's a Header error you need to download it again.

if you are getting 'Nans" on the weights that means, while you're merging the checkpoints you are overwhelming the weight nodes.

If you are still looking to merge it, use triple sum and normal method