LyCORIS for "Swamp Escape Additional Assistance" was uploaded to civitai and is still unpublished. If you need it, please download it from HERE or https://civarchive.com/models/51564/swamp-escape-additional-assistance
This is the final model from mmtrs-test.
https://civarchive.com/models/18734/mmtrs-test
I got tired of the repetitive work (swamp) of merging and training of mmtrs-test, so I will complete with this model.
There are many other great photorealistic models available at the moment, so there may not be a demand for this model now, but please use it if you like.
The difference between the standard and PLUS versions is slight, but you can use whichever you prefer... By the way, I like the standard and don't use the PLUS version.
Description
Models that have completed testing
FAQ
Comments (32)
One of the things I usually do when I'm either making a new TI or testing out a new LoRA I downloaded is to try to put them all to use at the same time... and I have to say I was having a hell of a time the other day with some new stuff.... and this + mmtrs-test were basically some of the only Models that actually got everything to come out looking right. So some credit is owed for the versatility and how nicely the model plays with other stuff you want to use. A+ on this, can't wait to use it more for more stuff!
Thank you very much for your valuable feedback and appreciative evaluation. I really wanted to create a convenient model that could easily switch between "animation" and "realistic" with a single prompt such as "illustration," but it didn't work out so well, which is why I decided to release it officially. I am looking forward to your continued growth and release of more wonderful works in the future. Thank you for your continued support.
Don't know... seems to me that the v2.x and the LORA v3 versions, were better than this final swamp. It would be interesting having a comparison grid with the same prompts among v2,v3 and swamp.
I thought this at first too after I downloaded Swamp to test.... but I think it has to do with just there always being some amount of randomness with the image generation process. It was by no means an exhaustive study.... but I did in fact do a comparison grid test that I ran with a handful of different prompts comparing the Swamp final versus the last mmtrs test, and the results were ultimately fairly similar looking, but more often than not the Swamp versions of the images did come out taking to the prompt better. It was a definitive enough result for me to delete the mmtrs model and only keep the Swamp Escape.
@Balthazar99 Ok, so it appears that the difference in quality is due to the specific renders and not to the models. Can you publish the comparison grid with the same prompts ?
@ritcher1 Unfortunately I can't. The assets I was using was for a likeness of a person I was doing for a project for, so it's not someone whose likeness I am able to post publicly.
It was a pretty small difference. I'm going off memory here, but I think I ran maybe 5 different prompts, and 3 different seed generations, and of those, Swamp Escape had less errors / bad generations than the mmtrs-test model. I want to say it was maybe only like 2 or 3 less. But it was less, nonetheless. So it cooled my previous concerns about the model being a downgrade over the test model.
Now, granted, the metric that I was measuring by was how accurately the models were taking to the complex prompts that involved a TI + 2 or 3 LoRAs.... there very well could be other metrics to measure by that could be different - such as overall aesthetics or style.
I to be honest didn't even actually realize at first using the two models, that they defaulted to a more Oriental appearance (similar to Chillout Mix for example) for people generated. I was using a TI that was not someone of that ethnicity, and the likeness came back pretty convincing of the TI I was testing.
But it's such a good model, I'm of course going to keep it and use it for other stuff as well. It's rare to find too many models that make such pleasant looking generations AND take so well to TIs and LoRAs so effectively.
@Balthazar99 Ok, so the swamp version is better compared to the previous ones. Can you confirm that this great model is also able to make non asian faces ? If so, can you publish some european renders ?
@ritcher1 Haha, you're more than welcome to do so yourself. Test it out. My opinion is that it's a great model, though I haven't used it all that much yet. My early impressions were of how versatile and receptive to prompts it was. Give it a whirl yourself and see how you like it. :-)
@Balthazar99 Just collecting models and can't make renders for hw limitation. No problem.
I think it is a difficult problem to choose which model is best, because the feeling of the different details and personalities in each model will vary from person to person who views them.
Also, please consider mmtrs-test0→1→2→2.1→SwampEscape→SwampEscapePlus not as version upgrades, but as different models due to the difference in merge rate and merge method, respectively. SwampEscape is the model to finish that endless swamp work. There is still room for improvement if we pursue the ultimate, but it is a waste of time, effort, and electricity cost, so we need to give up somewhere.
In the meantime, we have uploaded comparison images of each model for your reference.
@mmtrs Thank you ! One last question: is this model also able to make non asian faces ?
@ritcher1 In my experience, yes. I was using a TI of a non-asian face and it was extremely accurate to the person's real life likeness.
@ritcher1 Non-Asians are not impossible, but since we are learning additional Japanese women for the base model, they will be rather Asian-leaning unless we specify a prompt.
@mmtrs So it will be sufficient making the right prompt to also get non-asian faces or it will necessarily requires some embedding (TI or LORA) ?
@ritcher1 I think it is possible to generate a woman for almost any model, not just my model, by putting (desired nationality 1girl:1.8) in the prompt.
Example: (Polish 1girl:1.8) etc.
If that does not generate the desired nationality, then I would suggest using a TI or LoRA of your choice...
@ritcher1 Matched the topic. TEXTUAL INVERSION is now available. I tried a few, but I didn't see much of an effect. But the idea is great. 3 TIs are published and the location of the most important file for neg is hard to find.
https://civitai.com/models/50755/style-asian-less
Also, apart from that, I Generated a multiracial race with only prompts, so you can refer to it.
@mmtrs Magic ! Thank you. This make the Swamp much more versatile than what appears from the renders.
@ritcher1 I hope at a certain point, since you say you're just collecting models and such - and I see you e v e r y w h e r e on so many different model comment sections asking for a fp16 pruned version.... I hope since you say you're hardware limited, that you'll check out a Google Collab or some other means of actually using Stable Diffusion and show us all the awesome images you've had cooking in your mind that you can finally show off and use with all the different models you've been collecting.
@Balthazar99 As I've already said, I'm just collecting and giving feedback/suggestions to the authors. I don't make renders.
Where is gone the "swamp-escape-additional-assistance" ?
"This model is not publicly available and cannot be viewed by the community due to violations of the guidelines. The description of the resource is inadequate. Please include information such as training instructions, recommended weights, or additional information to help users better utilize the resource. If you have adjusted your model to comply with the guidelines, you may ask one of the moderators to review it. If you believe this was done incorrectly, you may file an objection."
The above warning has been received and the publication has been suspended.
I guess some people don't find it interesting to be published... I'm starting to distrust people.
@mmtrs You are doing great models, so I don't understand what was different with this last one suspended. Just a lack of description ? Now I'm curious; what that addon would do ?
@ritcher1 I created and published this with the intention of adding SwampEscapeAA to the prompt to take it to another level.
I added a note to the description and now have a warning "This model is currently being reviewed by our moderators. It will be visible to the community once it has been approved." It will be visible to the community once it has been approved.
@mmtrs Well, I think it's now standard procedure. We have to wait.
@ritcher1 I uploaded the file to HuggingFace because it was still in the state of being Unpublished forever.
@mmtrs Thanks. It is a LORA, right ? Can you anticipate a little description of what it adds to the full Swamp model and how to use it ?
@ritcher1 To use it, simply add it to the prompt.
The effect is to slightly increase the photorealism. However, if the weight is increased to about 0.4 or more, the composition of the original picture will change, so I enjoy Generate by adding it to the prompt with a weight of 1 from the beginning, rather than adding it to the prompt afterwards.
As an aside, I was curious to see what would happen if I used it on an animated model, so I used it with "AnythingV3.0" and found that it did not improve the realism, but rather tightened up the blurry animated picture, which I thought was interesting.
I also found that using "SwampEscapeAA" on the many photorealistic models that are currently available with a perfect sense of realism does not make any sense at all, and the picture is burnt and distorted.
@mmtrs Interesting experiments. Thank you for the clarification.
Really excellent model for LORAs and NSFW support. It can use both anime and realistic character LORAs. One of my favorites from this entire site
Thank you for your high rating. We look forward to working with you in the future.