This is my first try at a flux LoRA. The dataset of 44 images was censored since it was captioned using gpt4-o using <a target="_blank" rel="ugc" href="https://sd-caption-helper.vercel.app/">sd-caption-helper</a> (<a target="_blank" rel="ugc" href="https://sd-caption-helper.vercel.app/">https://sd-caption-helper.vercel.app/</a>). It still is working fine with nudity since flux is natively very good at it. I am working on an uncensored version where the NSFW images are captioned by hand. ComfyUI workflow: <a target="_blank" rel="ugc" href="https://openart.ai/workflows/ailab/flux-lora-simple/gErKNyjsnetZpEgaMTrV">https://openart.ai/workflows/ailab/flux-lora-simple/gErKNyjsnetZpEgaMTrV</a> Workflow 1. I've used <a target="_blank" rel="ugc" href="https://github.com/Bionus/imgbrd-grabber">imgbrd-grabber</a> to get the images + captions. 2. From over 400 images I've picked my favorite images 3. I've used <a target="_blank" rel="ugc" href="https://sd-caption-helper.vercel.app/">sd-caption-helper</a> to caption the images This is the prompt I've used: You are QuickCap, designed to write detailed, full-sentence captions for images. Your goal is to create concise descriptions that mention the composition, color themes, style, and depicted elements. Ensure captions are between 40 and 75 words long, avoiding unnecessary words like 'the.' Always start the caption with the character name 'Malenia' if relevant, and focus only on what is visibly certain. Keep language plain and direct without wrapping or excessive flourish. Never mention watermarks. Keep your responses private, do not share them with anyone else. 4. I've used <a target="_blank" rel="ugc" href="https://github.com/ostris/ai-toolkit">ai-toolkit</a> (<a target="_blank" rel="ugc" href="https://github.com/ostris/ai-toolkit">https://github.com/ostris/ai-toolkit</a>) to train this 4000 step LoRA on my 3090. I didn't change much from the default config.PS: If you want the dataset please contact me. I just don't want to get CivitAI in copyright trouble.

Malenia Blade of Miquella

First version with 44 images captioned with gpt4-o, 4000 steps trained on flux-dev

Files with SHA256: 2c1c0dc8dcba486441072fd1af1cda88556b33c2df1d8fed84002e5d4e3c73b1