Stable Diffusion - Can't figure out this Style

Beginners
1

Hi. I’ve been playing with SD for about a month now and getting comfortable with it. I’ve got 5 different UIs installed, all sharing the same checkpoints/models to avoid duplication. I’m trying to replicate a style generated from an online “free with watermark” site but I can’t even get close to the same results. I’ve tried hundreds of combinations with SD, SDXL, Flux, and all the most popular safetensors on CivitAI.
I can get digital art results and photorealistic results but nothing close to what this site was generating.

emo4
emo41024×1024 147 KB

only allowed one embed so:

1024×1024 161 KB

What I can’t get from this simple Lara Croft prompt, is the art quality, the colour saturation in the face as an example. I either get an oil painting result at one extreme, or a photorealistic result at the other extreme. I’ve also tried 100 different styles and combination of styles along with the absolutes, juggernauts, realistic, midjourney mimic loras etc. Ofc there is Photoshop, but that’s not the solution.

Any help on how to get to that style would be really appreciated.

1 Like
2

Is it wrong to use LoRA? I think the current general style for serious image generation is to combine the LoRA of the character’s design, the LoRA of the painting style, and various other LoRAs. Especially in the case of copyrighted material. I see a lot of LoRA, especially for Flux, SDXL and SD1.5.
https://civitai.com/search/models?modelType=LORA&sortBy=models_v9&query=Lara

There is also the means of creating a LoRA exclusively for yourself. You will always be able to produce a particular character with the style of painting you prefer. If you decide to do so, there is a lot of know-how and tools in Civitai’s article and in HF’s Spaces.

3

Hi again. No it isn’t wrong to use a Lora, of course not. The issue here is not about copyright, it’s about getting to a certain style. It shouldn’t have to do with training and creating a new checkpoint with Dreambooth as an example. My issue isn’t with consistency of character either. I’ve seen dozens and dozens of images on CivitAI as well as dozens of models and Loras. Some come close to what I’m trying to generate as a style but most are either super realistic photo styles or at the other end fantasy styles.

I don’t want the realistic look nor the fantasy look, but something in between. Digital art is how i would describe it I guess. And it doesn’t have to be Lara Croft. I just want something that is clearly “art” rather than photography and in the style of the pictures i posted.

So what I’m hoping to get is "try this model, with or without this Lora, and using this style in your prompt. Frustrating part is i’ve spent hours experimenting and while i get excellent results regarding quality, I can’t match that particular style.

This is about how close i can get to the desired result, but for me it’s an epic fail.

2024-10-18_00-44-56_4391
2024-10-18_00-44-56_43911024×1024 66.8 KB

4

You’re going for a style that would be called 2.5D or 2.8D. if it were SDXL, it would be easier to find a base model close to that and train on that, wouldn’t it?
There are quite a few, semi-realistic models.
If you don’t use LoRA, or even if you use the same LoRA, the results will be totally different if the base model is different.

5

I understand what you’re suggesting, but with the 100s of models available, both checkpoints and loras, not to mention refiners etc, you would hope to find something off the shelf that you can tweak and fine tune whether it is SD, SDXL, or flux.

Training is very difficult (despite what the clickbait Youtubers advertise) and I’m still new to AI image generation. Given I don’t have an issue with character consistency, training would be the last option. Again, I’m hoping someone will read this post and suggest using a model/lora that is close to what i need and obviously not something I’ve already tried which includes all the major checkpoints such as juggernaut, absolute reality, dreamshaper etc.

6

As for image generation, the current HF forum isn’t very active…
I wonder if communities like Civitai or outside BBS are more active in exchanging opinions. In my country, most of the in-depth information comes from anonymous communities.

The search for a model is always there, but as a last resort, it might be a good idea to elaborate on the prompts. There are still not many taggers who can accurately depict the art style, but there are still individual differences. Here are a few examples from the sample above.

A vivid digital artwork portraying a young woman with auburn hair, set against a backdrop of rugged mountains and a clear sky. She is wearing a green tank top and carries a backpack, with a quiver of arrows slung over her shoulder. The woman's expression is one of determination and focus, suggesting she is on a quest or adventure. The style of the artwork leans towards realism with a touch of fantasy, emphasizing the natural beauty and adventure of the scene.

A highly detailed, realistic digital painting of a muscular, attractive woman with light skin and long brown hair. She's smiling warmly, wearing a green tank top and carrying a backpack with a rifle. The background is a rugged, mountainous landscape with rocky cliffs and a cloudy sky. The woman's physique is well-defined, and she exudes confidence and adventure. The artwork is highly polished with a photorealistic style.

1girl, grin, brown hair, breasts, ponytail, lips, long hair, medium breasts, between breasts, evil grin, green eyes, tank top, necklace, earrings, shirt, sleeveless, looking at viewer, outdoors, sword, mountain, solo, realistic, upper body, blurry, jewelry, freckles, smile, weapon, nose, bag, strap, backpack, sky, teeth, low ponytail, strap between breasts, weapon on back, cloud, dirty, green shirt, day, mountainous horizon, cloudy sky