A behind the scenes look at Delavestra's creative and production process: History and journey through professional AI art creation

(Video attached for the intro)

Hi all, I get asked every day about how I make my images and animations. I have no intention of making technical content, but for those of you who are interested or may want to know what I go through and the rigorous work and effort that goes into my RnD and production... here's a quick guide with some images on how I do it...

In late 2023, I saw a Black Desert Online creator turn one of his BDO characters into a photorealistic version of that character. I thought it was incredible and had to learn how I could do this myself. That lead to a two year journey to today, and it all started with learning Stable Diffusion with the A1111 (Automatic ‘eleven-eleven’ UI). There are mountains of tutorials on youtube, so you’re welcome to go there for the basics.

That soon turned into FORGE UI – a better/faster version of A1111 which allowed for faster image creation and more control/options over my creations. I soon found out that prompt engineering is less than half the battle when it comes to high quality AI art. Models are only capable of reproducing the generalizations of the data they’re trained on. A LORA, (Low Rank Adaptation Model), is like a mini-model that you can apply to a main model, like SD1.5, SDXL, Flux etc, and get a new look/characteristic/quality out of an image.

I would spend (and still do) hours balancing and testing the weights to apply to different LORAS, and in what quantities would yield new effects and styles that I liked. Fast forward to blending 10+ LORAS for a single image, training my own LORAS, merging my own models, and I came up with my original ‘Delavestra’ set just over a year ago. The image below is a black canvas of Forge UI:

Up until early 2025, I was only using FORGEUI to create my images. Starting around March of this year, I switched over to using COMFYUI (youtube tutorials will explain its details) and have gone into an endless pit of upgrades in an attempt to perfect my images. This is what a workflow looks like that I use today:

This workflow incorporates 5 different models, along with custom LORAS, to create, upscale, polish and finalize an image. It uses a base model plus 5-15 LORAS to generate the base image at 832x1216, then does an upscale with that model, and another small upscale with an upscaler model, then proceeds to a beefy tiled upscaling with a controlnet and medium denoise to add refined final texture and quality to the image, then it utilizes 3 detailers to enhance the hands, eyes, and face, in that order. Finally it sharpens the image and out pops and completed iamge (sometimes).

I'm still between using ForgeUI for it's qualities and benefits and ComfyUI for what it can do, and I've tested hundreds of workflows now by this point to see all of the wonderful things it can do.

Here's a quick list of the things i've tested that aren't worth going into detail now:

Custom Flux Face Lora creation per character set

Base Model selection, testing, experimenting, iteration, selection

Upscaling and enhancer nodes and processes like daemon detailer, hires fix, image upscaling, flux upscaling, sigma upscaling workflows with high steps, one step upscalers, tiled upscaling (all with different models and controlnets), and the list goes on and on.

Long story short, I've paired this with extensive hardware upgrades where I now run my 5090 GPU with an HVAC line to pump out the hot air away from my desk. I use combinatorial generations to run overnight batches of 400 to 1600 images over night, and then must review, sort, name, file, store, watermark, and post all my sets.

I've created 5 custom pieces of software shown below:

Here's a snip of the code from the image sorting tool which allows me to quickly review images and with a single button press, allocate those images to the appropriate folder/file structure, rename, and copy and delete the original, while maintaining history in case I miss click.

When it's running, it looks like this:

I used a combination of ChatGPT, Grok, and Claude to create these programs, which all take 4-6 hours for the version 1 development. Then as I use them, I tweak them as needed.

All of that now leads to animations, a whole new beast that's far more tricky to get consistent.

The animations I create are produced with the WAN2.1 video model and require specific LORAS for specific fun visuals to take place effectively. Just like an image can have a LORA for a style or a person, an video WAN LORA can encode an visual act...

I had zero experience with comfyUI and generating videos before 2025, and this is what my current workflow looks like where i've meticulously tested EVERY lora to ensure I'm using the highest quality training data, (and most up to date, they change weekly), the correct prompts, and lora weights, along with balancing the Model sizes (GGUF versions) along with Speed enhancing loras, like the the lightspeed self Enforcing x2v model that really only works for SFW boring animations (doubles the generation speed), and setting up Sage Attention and Triton on my PC which is a real bitch to get working:

I tested 5 different Image generation workflows and came to the conclusion after making 500 5 second animations that this one is indeed the best. You've seen the varying quality of my animations over the past 5 months, so you now know why. Also, with my current heavily optimized setup, it takes me about 6 minutes to produce a 5 second animation at the best quality. Around 33% of those are useable... Meaning with the best GPU, optimized setup, and brute force trial and error to get the best of the best, it takes 18 minutes to generate 5 seconds of animation if you already know you have the right prompt, input image, models, loras, settings etc.

I'm purposefully glossing over 90% of the details because this would take literally hours just to explain this one workflow...

Anyways, there's a very rough and quick overview of a part of my production process and what it takes to produce the content I make for all of you. This is the first time I've created anything like a tutorial or explanation of my process, so let me know if you find this interesting or if you care at all. I know what people are here for, so be honest if you'd rather not worry about details like this.

For those of you reading to the end, I appreciate you greatly and can't wait to make more incredible artwork and eyecandy fun for years to come.

-Del

PS. When Dark Elf Maidens and Ashe Animations come out in the coming days (I've ran over 700 (seven HUNDRED) 6 second animations already), you'll now understand why it took some extra time and the effort I've put in to the results and quality I now deliver. There are still kinks, and ai output is sometimes inconsistent or glitchy, but it's a journey I'm on and I'm thrilled you're here, fueling the ride.