Hi to everyone!
For not making it extremely long and boring this will basically a place where I will show images models/emb. I'm working on right now, other exclusive models not released here, personal thoughts on certain settings (why not?), things planned to train, etc. (All related to AI since I suppose drawing, materials, 3d models and so on are a bit nonsense in CivitAI U^^) Feel free to comment whatever you want since I'm always receptive to ideas. Will be basic my diary.
Will try to maintain it organized and will edit it according to my doings, etc. Don't expect anything too fancy but yes coherent. And for the ones who want to know a bit about me.
Brief introduction and how this started
******************************************
David, 36 y.o., professional artist from Spain and I started in April-May past year first with DiscoDiffusion generations (the cover is my first generation in DD, awful), then came MJ and then SD. I started training textual inversions on August last year and since then I haven't stopped, being Dreambooth, textuals, LORA, LYCORIS, etc.
For the curious ones (and if you are bored enough) you can check out my progress on my Instagram since I used to post it there. The rest is story, a lot of reading and most of all experimenting (I'm extremely proud of all my Dreambooths done in the past), meeting my awesome friend Dervlex, creating our Patreon and getting ready for SDXL which I have great hopes... and if not, making all on my hand to deliver quality here and in Patreon.
So... back to the main theme. what I'm doing these days?
Dense Blue (Patreon exclusive LORA alternate ver. of Bubble Drip with intentional differences on color schemes altering dataset and also using a heavy multiresnoise scale):
I did several testings with repeats and multiresnoise but this time playing with iteration steps (None of those were released, not passed test/prompt control/quality). Chambercore/Bluecore:
However not everything was bad since this made me focus more on Duotones training and created Poison Cloud (exclusive Patreon LORA). I will revisit this style since I enjoyend it a lot, I think it fit comics a lot.
And since a week ago more or less I finished SeekGold (texture morphing focused and no variants exclusive planned) and SeekAstral which I'm working on a re-vamped and more concept based version (the one in Civit is more lineart like), you have a preview here (Unreleased but planned for Patreons).
Edit 04/07/2023: Trained SciPink a dual LoCon/Lora (both different) Patreon exclusive to create scifi pink aesthetics with certain styles embedded.
This is how the style adapter works:
Lora Ver:
LoCon ver:
Edit 07/07/2023: Started training Pill a dual LoCon for CivitAI, some previews:
You can see the style change here:
Edit: 09/07/2023: "Pill" was released. You can download it here
As you can see its red and blue colors duo-tone (with gradients) with a graphic/vector embedded aesthetics. Super cool for vector and comic styles and with some prompting you can take it to other styles,
Edit: 11/07/2023: Started training tests for Ibai Llanos (a famous person in Spain and in the world). Only for personal learning purposes.
Edit: 20/07/2023: Started testing SDXL trainings. This is a WIP training for both, dual SD 1.5 and SDXL release LORA. For now, WIP.
Edit: 27/07/2023: After some days testing SDXL trainings I released Coinmaker SDXL 0.9 based LORA (just the same day SDXL 1.0 is released, lol):
Edit: 01/08/2023: August. Some days off but still training, yup. Trained a Sun]flower LORA (Exclusive Patreon LORA) and since I wanted to be on CivitAI too will train a SDXL version, this time I'm approaching using LOKR (LyCoris) IA3 to see if I can get less size and same quality?
Edit 08/08/2023: After almost two weeks I release finally Sun]flower SDXL LORA, excited about my first SDXL style LORA. Still digging into more alternatives for SDXL as LycoRis, LOKR, etc.
And for those who wondering, YES, I'm using both ComfyUI and WebUI, I mean, why not?
Edit 16/08/2023: Double release day (after an entire month of work among Sun]flower). Exclusive LORA for patreon (commission) called SeekFantasy90 - Tome I (there will be more) and public version of Vulcan.
WIPS: Tunning Halftoned versions of Vulcan for Patreon and do the SDXL version of Vulcan.
Edit 19/08/2023: With the new Vulcan 1.5 and SDXL double release (non stop baby). I saw that filesizes are a thing to (maybe) "fix". For now everything I tried, failed. From changing optimizers, using IA3, dadapt, lion, reducing Dim/Alpha etc. Will elaborate more if I solve this.
As important thing I did a mini test that some of you noticed. Some of the SDXL images are 512x768 hi-res fix applied (you should know that quality is better if 1024x res.) but at the same time I found that this way the image were less "painted" (if this makes sense...). This is a nice collage.
Edit 31/08/2023: "Holidays" are gone. In these days I kept searching for things I liked to train and decided to train some personal works with another mix from other artists. The result is something I'm still polishing but looks kinda cool. My new 1.5 LORA "Schematics" will be available soon in CivitAI for free. SDXL will come next and Patreons will have some custom versions for the support.
Edit 07/09/2023: Schematics LORA (SD 1.5 ver) was released (which almost 400 downloads in 24 hours, for me that's a lot, thanks!) the alternate version is on the making and I went back again to SDXL trainings. This time I'm just experimenting, I will only plain to release SchematicsXL but I don't know when and where, we will see. Testing Adafactor/Adamw and still making my mind for which direction to go...
Edit 11/09/2023: I have started working on Ice Age XL LORAand Ice Age SD 1.5 LORA's. Both performs very good and I really landed the aesthetics. Just I need some tweaks here and there and potential retrainings. This time I will release here the SD 1.5 ver. Also working of Schematics XL LORA for also releasing it here. Schematics will take a bit of time though. Ice Age will come before, I think.
Edit 15/09/2023: Ice Age 1.5 LORA was released on CivitAI
In my opinion is better than my super and ultra underrated D3fr0
Still I think I like the old one however this is much more worked and better on everything.
Edit 19/09/2023: I'm working on Schematics XL and will be ready for the next week, more or less. Final tests needs to be done and I want to test certain type of prompts on more advanced syntax but it works very good at the actual state, oh, and yes, will be free on CivitAI.
On a more personal plane I'm doing more tests with XL. Trained Ice Age XL for Patreons and kinda happy with the results but still learning. SD 1.5 and SDXL works on different way. Still getting used to ComfyUI (Finally switched to it).
Edit 25/09/2023: Schematics XL was released. After some minor test done the LORA behaviour is up to my standards so free to go.
Edit 03/10/2023: I have been testing Dreambooth trainings and for styles the differences I noticed on training a LORA vs. a model checkpoint is minimal. And yes, I tried extracting the LORA's and play with conv. ranks but meh, not worth, IMO.
In my current settings it's 5 hours (DB) vs. 3 hours (LORA) so I will stick for now to LORA's. 19 GB VRAM currently of LORA Vs. 23.7 GB VRAM of Dreambooth.
On the other hand I have been testing different styles to train in XL. Some examples here bellow on 4 different personal XL styles (Unreleased and testings). Deconstructed/Disfigured photography, Cartoon Exaggerated style, Unsize photography (Sci-atmospheric photography) and Styilized 2Dviews (different aspect views of stylized 2d painting).
Edit 11/10/2023 (Part 1): This week I was busy with other thing but still got train to train and doing sme testings here and there. First of all I'm training "B3L". B3L is a LORA which aims to simulate the Cinema 4D octane render works and I use several of my own works (and other great artists too) to train it. Still is in WIP but I can say the XL version (which was trained on different data on purpose) is almost finished and the 1.5 will come to CivitAI next week. B3L won't be the only model following this structure cause I have a lot of my works to train as I spent almost 2 years daily rendering with Cinema 4D and Octane.
1.5 Ver.
XL Ver.
Edit 11/10/2023 (Part 2): And after several of you asked me to do this, I did it. Yeah, I trained a 100% one artist pure style (without anything of mine, non mixing or anything from other artists...) I decided to give it a shot and as usual, first trainings were horrible but finally I got a "clean" simulation of Aaron Horkey's popular ink/drawing style. This was just a test but in the future I will love to do more of this (even won't be public to download until remixing it with other artists/styles and/or personal works too). If you can, support him too please.
Edit 19/10/2023: Insane week to be honest. First of all, yesterday I released B3L LORA for SD 1.5 and for SDXL. However the SDXL is only available in Patreon as part of thanks to everyone supporting our labor there. The LORAs were done under different datasets and that's why even they go after the same task (Cinema 4D abstract renders), they behaviour in a different way. The XL is more intricate, flexible and maybe "realistic" version to it's 1.5 partner due to how SDXL understand prompting single alone.
SD 1.5
SDXL
More important... I released my AI Style training guide on dataset preparation
This was a thing I wanted to go long time ago since I feel most of time we spend more time tweaking DIM/Alpha, rates, clip skip, models, etc. than actually focusing on the main problems. Maybe this will give a different insight and way-to-go for people (or not). Also included the experiment I did with the known artist Jeff Easley. Feel free to read it cause I think it's good information. At least something I wish I was told before doing insane amount of trainings in the past.
Edit 25/10/2023: This was an interesting week. I took some time to do some sketching after some weeks without barely moving the pencil and only datasetting/curating/training. Well... I'm working on a SDXL Warhammer type-XL LORA (with the help of LDWorks Dervlex my fellow team mate) . More news soon.
Edit 29/10/2023: Paused a bit things to enter in the CivitAI art style contest. Back to normal in a few days. I'm training a Dominic Qwek LORA for SDXL. Soon more Warhammer and a preview of my Moon XL/1.5 I'm working on but for now... time to enter the contest!
Edit 31/10/2023: Lets see what happens. I entered the CivitAI contest with my new Dominic Qwek's art Style for SDXL. It's based on his latest works (from 3 years ago onwards) such as these
or these ones.
This time I changed my strategy to train and caption and it's sorta like an experimental LORA. We will see if I keep using this method of captioning plus settings and preps or the old one with additions. Samples here.
Edit 07/11/2023: I returned to "normal" trainings and I'm working on a double release and an experiment. First of all, I have changed my strategy towards training in SDXL, for not making super large feel free to read it on my original Reddit post and to summarize it, I have started to making more complete trainings with Dreambooth, extracting the LORA's and merge them but in my case I do TWO full trainings with differents settings under samedataset so the mix goes toward a main target. You can see some results of my future patreon's Moon XL LORA here:
And you can peep the SD 1.5 version here:
Edit 24/11/2023: Yeah yeah, late as ever, sorry. Well, for every soul following my steps around here, I released 10 days ago Moon (SciWhite) LORA so all of you can use and make cool things with it. If you want the XL version you can get it on Patreon (and at the same time you will be supporting me). By the way, some of you may noticed, all I did in this month was experimental trainings (hope you liked them!).
And also released a few minutes ago something that "probably" is the last thing till December. GTA v Synthwave edition Style for Stable Diffusion XL (request):
Edit 10/12/2023: Hello again and sorry for the late. I have been digging myself extremely deep with nodes and ComfyUI. In fact the moment I saw MagnificAI I wanted to do somethig very similar (but capped with the current tech I have). I'm working on several image enhancement workflows, custom detail LORAs (different methodology to the ones already here on CivitAI), upscale methods, etc. I plant to schedule the worfklow into 4 parts:
0 - Image caption (optional) Clip/Blip to identify details (testing VRAM usage for workflow)
1 - Image enhancement via denoise
1.5 - Preserve the integrity of the original image (target or txt2img)
2 - Upscale details via Segmentation
2.1 -Upscale details with Self Attention or gradual upscaling replicating method in ComfyUI
3 - Second pass via ControlNet (1.5) or using other method in XL
I suppose a lot of methods will be releasing from now on bases into this but I wanted to do one myself.
And of course... LORAS! I have one custom with my personal style in the kitchen and another holding a bit which is a version of HR Giger (yet, another one, yes) for XL and probably 1.5. We will see, Bear with me cause I'm currently doing a lot of things, and please support us.
On another point, Dervlex release his upscaling method for 1.5 models and some wallpapers in our Patreon. Cool content!
Edit 31/12/2023: HAPPY NEW YEAR!!!
Sorry, been very very busy with the MagnificAI project and current released couple workflows, extremely experimental here, in our Patreon. You don't need to pledge anything, just download them. Feel free to contribute if you want cause it's always good for the cause.
Here is the slicer worfklow
Here is the normal workflow
And the making off of the slicer worfklow
Remember that all are very experimental and suscept to changes.
For 2024 January are coming the rest of things that were scheduled for December Aka HR Giger month. Not normal Giger, though ;). See ya later!
Edit 18/01/2024:
Back to my "normal" life. I was contacted by some groups regarding the Magnific posts I did on Reddit. To be fair all I knew was already explained on the post and I don't have plans to go further cause I lack off knowledge. The last update on this aspect is that I developped a way to replicate their tiling method (based on what laidawang, an user of Reddit showed me in the generations/upscales).
If you're curious about the process, you have the whole post here and if you want to download the ComfyUI JSON file, you have it here. In a nutshell they changed their tile system so they can drag part of image info for re-upscaling on every step to the point all parts are connected. You just have to see the intersections.
More important. This finishes all my part on the Magnific replicate project and I will allow others to get every info, JSON file, etc. to do whatever they want or develop anything but please, think in the open source community too. Thanks.
So, yes, I'm back to my training field. I started training 3 days ago (after almost 13 days of curating the dataset), the H.R. Giger rainbow LORA for SD 1.5 and XL (will need to decide which one goes to patreon and the other free).
Another important news is that @LDWorks Dervlex (my pal in arms since looong time) is back and can train in SDXL (he couldn't before) so we will be doing way more content on CivitAI and Patreon too. Please support us for quality content. So this January, all will start over. Thanks for your time!
Edit 19/01/2024:
Good to be back. Both LORA's of Giger Rainbow have been trained after 2 weeks of curation in the dataset. The SDXL is being tested at the moment but the results are good. So for now I will release the XL version on CivitAI as soon as I had done more testings (only vanilla SDXL fully tested) and also checking my "modules" as I trained splitted painting and watercolor so it has a deeper impact on the generations. The SD 1.5 is good and will be released in my Patreon at the lowest tier possible and fullfilling a request (well, not exactly a "pure" Giger style LORA cause have my touchings, additions, works, etc. but is very good!).
Some visuals of the SDXL Vanilla ver:
Edit 23/01/2024:
And finally my H.R. Giger Rainbow X LORA is released on CivitAI (in the case you want Patreon 1.5 version you will need to go my Patreon). Super happy with the results.
This is the 1.5 version:
This is the XL version:
In the next month I want to do something more anime/manga/japanese. We will see how it goes.
Edit: 12/02/2024:
Well, a lot of stuff going on, jobs, side projects but still here for the community and trying go keep making quality things.
I have been working on a request I had in Patreon 2 months ago but as always on my own terms. Avoiding to be a LORACopy-Style. So this month you should be seeing here my new "Midori No Kaze" which means Green wind. And that's what this thing is about since I added elements as dynamic movement (flowing), wind patterns, green/white (forced training on color schemes) and eastern style (which is not the same same as Anime but has certain components).
And since I'm a freak of cybernetic stuff, yup its there. You can prompt it or not to enable/disable but the more you tweak the text encoder towards it, you should be seeing it more clearly. Some table charts of the SD 1.5 (1) and the XL version:
We could say that Dynamic-white/green scheme-eastern-75%anime-cybernetic stuff I love, is what this thing could be. I suppose one version (1.5/XL) will be for here and the other for Patreon to give supporters something but will work so both will be very good and up to my standards. And for the curious ones, this was the very early 2-3 trainings (which I didn't like them so started tweaking from there to get something in more trainings to this and started this as a solid point at least)
Releasing this for February, hope you like it. ^^
Edit 20/02/2024:
Released Midori no Kaze for SD 1.5 in CivitAI.
To be honest I had some problems on this release but quite happy with the results.
You can see the previews here:
There is also a thing related to the VAE and as O pointed in the original post... care with dynamic range colors VAE. You can even not use them if you want, like here:
The first one on the left has VAE ON, rest nope. Figure it yourselves ;).
Even I'm not a big fan of anime I have to say that this works very very good (and maybe even better than expected). I'm doing optimizations in the SDXL version which will be released for all our Patreons (as giving them something, you know) for their support the next week.
So, see you soon, thanks to everyone for your support!
Edit: 29/02/2024:
Midori no Kaze XL version was released in Patreon for all of our supporters. You can see some examples HERE:
If you ask me, I like this version more and gives me room to more things but I also loved the 1.5 one so both are def. good. See you on March guys!
Edit 11/03/2024:
Time flies. This time I got myself involved into weather LORA. Raining. But... yes, challenge is you get rainy weathery, wet surfaces, water, etc. without calling for it. Also I added an extra "module" (let's call that) in the 2nd training so it can bring Depth of Field (DOF) to the scenes without again, calling for it. The name is "t1hal".
About the making: Quite "simple" to build, hard to pull off. This is just a LORA trained (and retrained 2/3 times more, first training kinda meh) on wet textures/surfaces, water, waterdrops, raining weathery, etc. Some syntethic images, some photographs and some Substace works that I had over there + DOF on the mix.
I noticed that this LORA works on CFG 2-2.5 (wonder if that was cause my settings?) and when it goes a bit higher, the rain/water/drops, get way too messy (you will see when it's on CivitAI, in the next 2-3 days probably). Overall happy with the result and learned couple things so worth.
Samples here:
Edit 13/03/2024:
Under the Rain was released for Stable Diffusion 1.5 HERE (CivitAI) and also the SDXL version for oour Patreon Supporters.
You can see the previews here of both versions:
SDXL:
On a side note of course I'm doing an Akira Toriyama rendition, I'm just creating and gathering the data. Will take some time though.
Edit: 26/03/2024:
Closing the month and getting ready for April.
For the first time I have dropped an experimental "mini Dreambooth" which is based on my own trained and tunned 1.5 checkpoint (avoiding merges and other things that bias the models) but with the same t1hal inner aestetics.
With this I cover the triple release I was planning to do this month. 1.5 LORA in CivitAI and XL LORA + Dreambooth oriented model in Patreon. Some shots of some random prompts:
Edit 16/04/2024:
Hello again and for those who were asking I'm not dead (yet, lol). This month I'm working on my second release of my C.A.H. (Concept Art Helper) series and I'm still doing the training for the past week so I think it will be ready soon. Not aimed to a single artist style and more as a global style. I'm working on the aesthetics right now.
As usual I will publish some versions in Patreon (maybe 1.5, maybe SDXL, still don't know).
Edit 23/04/2024:
As professional artist I can't be more happy with the results of my next LORA. I think it's going to be my best LORA and I really feel proud cause my art is there too. It's like seeing a son of you swimming for the first time in the pool. The name will be C.A.H. (Concept Art Helper) - Quick Designs. I'm working on the 1.5 cause the XL I'm totally happy with it.
Edit 25/04/2024:
Today is the release date of my Concept Art Helper LORA. Sad to say that It will only a Patreon release for the time being at leats. Did my best to port to 1.5 but the release it's not up to my standards so probably will put other thing instead, sorry for that. Keep tunned cause I deliver as always! ;). I prefer to give you another thing that is cool instead of something I can regret later.
Also I'm digging myself into PONY XL (v6/v6.2) and even I'm not sure if it will go good with styles, I saw that it absorbs very good character concept (although it works better with anime aesthetic). Will keep you informed.
Edit 22/06/2024:
Yeah! Hello there all! A lot has happened and I can't be so focus on CivitAI stuff as I used to be but still I will try to deliver things. Where to start?
I released in May Devil Within XL LORA for Patreons, take a look at the images here:
I was aiming to create black and white ink designs, cybernetic, robotic stuff, creatures, manga characters but with the most important thing of all: Super detailed lines with complicate and intricate designs. And also to maintain a clean background.
Continueing with the goodies... I released on Patreon too SeekFantasy90 Tome II for PONY XL (first PONY XL LORA I have ever done publicly). Take a look here:
And last but not least I started to do finetunnings, currently on PixArt Sigma.
And I released NINETIES it in Patreon too (expect some goodies to at CivitAI too, is just I have been very busy with projects) at lowest tier so everyone can enjoy it.
Nineties is a personal project and my purpose is to capture the illustration style of the fantasy art of the 90's. For now is minimal tunning on PixArt Sigma base moedl but expect a lot more to come. This is the different of the base vs. finetunned model. Image.
Some outputs too:
Edit: 30/08/2024 - Holid.. what?
Well, sorry for the lack of new to the few ones still wanting to see what I'm doing. I received some messages over here if I left AI, etc. On the contrary, I started with some private projects and very little time to "play around". Still I will try to deliver content here and in Patreon.
As a personal side note I have been more focused for the last 4 months on study and reading papers, doing optimizer tests, testing different arguments, new trainers, creating customization on Python code so I can be better at wham I'm doing.
And here goes a bit of what I have done (excluding private projects and main WIP workflows, you can find more info about that in Patreon):
First LORA's tests on SimpleTuner and AiToolkit (this had a wrong thing on the noise [old noise implementation] so the image are not crispy and good enough, we were testing guys).
The next I tried (aside more personal projects was a Flux-Dev) try on Thanos character training (which still had the old noise implementation so it's a good LORA but will give you oversharpening when now theorically not).
You can download it here, yes.
As one of a random thing... I tried merging elements from one LORA to another and the same time training some likeness test. (Old Noise implementation here too):
I started doing some LLM testing, LLM workflows auto captions (you can find more of that in Patreon, yes) and while I was working on FluxKraft, I trained more stuff so I can see the latest git pull changes. Here I had some interesting results cause were testing how LORA strenght is correlated with words on training.
Along my pal Dervlex (he released very cool FLUX LORAs too here in CivitAI) we started digging into some cool configs and I started seeing more changes that were better in my trainings. However, I'm still testing more in AIToolkit but this one comes from Kohya and honestly, awesome lol.
Ironic thing is this set comes from the old ones from flux and performed better, cleaner and IMO more stylized.
Recently I have started to cloud computing (also with my 3090 at home doing works) but this will level up the works.
I'm wondering also about training pure artists 100% based styles but I like creativity so I don't know if that will come or not. Just ideas.
Future plannings: Keep making embeddings XL & 1.5. Both have their area and niche. Keep training personal styles and also 100% styles.
Hope you have enjoyed!
Image random (daily Weekly):
Thank you guys!
PS: To see the images larger click on the link in blue and please, support me in my Patreon for more quality content. Thanks!