Hello all!
We've just hit a milestone today - the 500th model has been uploaded!
I am quite happy with where I am today :)
I was here on Civitai pretty much from the start. My first uploaded model has an ID of 51 ( https://civitai.com/models/51/felicia-day ) and it was uploaded on the 11th of November.
The first-ever model (id=1) was uploaded on November 4th. I remember I was quite agitated since the guide I tried to follow (Nerdy Rodent's - https://youtu.be/w6PTviOCYQY) wasn't working for me out of the box. I spent 3 days (Friday-Sunday) until I managed to install all that stuff and get no errors. The guide was released on 3rd October, my very first model was created on the 10th (I still got it!). There were some trials and errors, underfitting, overfitting, bad training datasets, and so on but on the 10th of November I had a first model I was pleased with. And to be honest - some of the generations from the 2nd published models aren't that bad either (https://civitai.com/models/69/miley-cyrus).
I remember that there was a big pushback towards Dreambooths for single characters, but there were enough people encouraging me to continue so I did.
But I was not uploading all that much. Late November and December was a time when I was training my colleagues from my company for the 2023 calendar which was quite successful :)
I was uploading sporadically and there was usually someone who would point out that Dreambooth is no go for a single character.
So I tried Textual Inversions and they worked fine on the model they were trained on (here is my first TI: https://civitai.com/models/6614/chrissie-mayr) but not on other models so I stopped focusing on them.
Soon after (February) I found out that one can extract trained concepts out of the Dreambooth into a LoRA. I tried that and I really liked the results, my first extraction was of Chloe (https://civitai.com/models/8244/chloe-grace-moretz) and I think she turned out quite all right :)
In April I found out that it was possible to extract the content into a LyCORIS so I tried that and LOVED the results from the get-go (https://civitai.com/models/28369/gillian-anderson)
So I kept doing LyCORIS and people seemed to like them. That gave me a boost to do more and more of them. In the meantime, I started writing some code to automate some parts of this process.
But not everything can/should be automated. To this day - I manually select the training datasets. There was a lot of trial and error there. Some models I never uploaded because I wasn't happy with them. But some I've uploaded knowing that they are not the best (usually noted that in a comment) and people agreed. Since I was able to do quite a lot of them - I picked those that were good and discarded the bad (they had to go back to the "drawing board" which usually meant revisiting the datasets)
Around May I started tinkering with other concepts and I started my "Concept" series which was mainly focused on improving certain elements of anatomy or adding some cool effects.
Mid-April I released my process as a guide (https://civitai.com/models/45539/dreambooth-lycoris-lora-guide) and I'm very surprised at how popular it is (21k views right now).
I'm glad to see that some of the best creators got inspired by it!
Around June I realized (thanks to the great feedback) that LyCORIS is not everything and that sometimes less reality and more flexibility is what people want so I started including more and more LoRAs (while still pumping LyCORIS). I've also launched my Famous People series where I upload Dreambooths of requested people (since there are still some people who value quality over filesize :P).
I've also started writing some articles and recently I've showcased a way to boost the likeness of a person by mixing two or more models of the same person. I really love the results and I figured others would do too so currently I'm pumping more LoRAs than LyCORIS but do not be afraid - I still want to keep uploading LyCORIS models :)
And then there was SDXL. I have plans and ideas for SDXL, I tried some stuff already. I guess I can say now that I was in the beta during SDXL 0.9 and I hope my contributions helped at least a little bit :)
I was actually very surprised (pleasantly) that my training parameters were included as a template in the Kohya SS GUI :)
I think SDXL merits a separate article. I will just say that I won't abandon the 1.5 branch but I also want to make content for SDXL. There are some things in the pipeline that can make it much easier but I don't want to jinx it yet so I'll keep it a secret :)
What is not a secret is that I want to create a service where people would be able to train using my parameters as a baseline so they could get similar quality results. There are many services like that already so I need something special to differentiate from the rest and that takes time :)
So, that was pretty much a summary of the past, now, for the future!
I have not been slacking, I really have a lot of models to upload :-)
I've been focusing on the LoRAs recently but soon I will equalize that and you will see again new LyCORIS models :)
We are celebrating our 500th model, and let me tell you the good news - I have more than 500 models waiting to be uploaded. I will list them below so that you may pick the ones you want to be uploaded first and you can write them in the comments :-)
I will try to do my best, I am aiming at uploading around 25-30 models per week.
And now is a time for a shameless plug :-) If you want to see some (or most) of the models sooner, you can do that by supporting me on my coffee page ( https://www.buymeacoffee.com/malcolmrey ).
All the supporters are getting a megaupload link where many of those models are available for download earlier. It is so much easier to just upload a model file without samples and that's a nice gift for those who support me :)
But no worries, I do not gatekeep any content, all those public models will eventually be uploaded to Civitai! And yes, by public I mean models of known people because you can contact me and request to make a personalized model just for you :)
Okay, the advertisement is done, now to the meat :-)
Here is what is already waiting for upload:
https://rentry.co/upcoming-lycoris-and-lora-models
And here is what I'm going to train soon (TM):
https://imgur.com/gallery/WpbAG37
And as a reminder, here is my TODO list (does not include yet all the image sets I've downloaded) so you can see what will be in the pipeline (and you can make suggestions still): https://docs.google.com/spreadsheets/d/1fOa6Jp_32DPM4Dc-oG2ILJOs0Xn1BydgbK2fbkp9k-c/edit#gid=0