About this version
### 🐉 Larry Elmore Style Model - Version 2 (V2)
Welcome to Version 2 of the Larry Elmore style LoRA! This version has been completely rebuilt from scratch to capture the legendary, nostalgic 1980s high-fantasy D&D and Dragonlance book cover aesthetic with maximum fidelity.
The flat, digital "cartoon look" from V1 is completely gone! The first version suffered from over-smoothing and flat colors due to legacy AI upscalers (like ESRGAN 4x). For V2, the entire dataset was meticulously re-curated.
---
### 🛠️ Training Specifications
* Dataset Size: 66 high-quality, hand-selected fantasy illustrations
* Training Resolution: 1024 x 1024 px
* Training Length: 27 Epochs (~1800 Total Steps)
* Batch Size: 1
* Optimal LoRA Weight: 1.0 (Feel free to scale down to 0.8 for more flexible compositions)
---
### 🔑 How to Use
* Trigger Word: painting style of larry elmore
* Prompting Advice: This model responds beautifully to descriptive, natural language (Descriptive Prompting). Avoid heavy tag lists.
Description
FAQ
Comments (8)
It absolutely nails the color but something about the characters seems more cartoonish than the original artwork. I'm not sure why. Any chance you would post the dataset?
OK I played with your LoRA quite a bit. It looks like you did a great job on captioning and dataset. I mean this to be helpful not as criticism. You super overcooked the LoRA. The example images I posted were ran at only 0.4 strength. If I use 1.0 it's cooked hardcore and simplifies the style. I don't know how many steps you did, but probably only need about half of them. All in all the it's very good.
Yep, just read "48 Epochs (Exactly 2,544 Total Steps)" Since you did batch size of 2, that's 5,088 steps. You probably only needed 2000 steps with a batch of only 1. 48 Epocs is too many. You probably only need about 22.
I just wanted to add on here, I agree that he could keep the same settings, batch 2 is fine, and then just save epochs at 10 , 15 and 20. I'd like to see a v2 with that. Krea 2 seems to learn things faster than other models ive trained as well. Great lora man!
OK, I've trained a V2 and uploaded 2 variants under the optional files. This time, I used a batch size of 1 and full 1024px resolution. One version has 1248 steps, and the other has 1976 steps. Feel free to try them out and let me know what you think! One of these will become the official V2. Thanks!
@rhin_tin I don't see the variant, i only see a dataset