v1 krea2raw Trigger = DCS1F1.
v1 krea2raw Trained @10,000 Steps, Rank 64, Captioned with GPT4.1, Dataset of 524 of my Midjourney Images in a general Scifi, Cyberpunk, Cybernetic robots, androids and people, vast ships and @gossipgoblin from instagram and youtube kinda style all mashed together. Prompt in the way i have prompted or select the prompts from 5 of your favorite images run it through chatgpt, grok, perplexity or open source alternatives for free and ask it to make you 10 more 100-150 word txt prompts in the same style.
This lora is much better at 16:9 horizontal 2k + resolution Images if your pc can handle it.
v1 is a test lora it needs detailed prompting, so large crowds and basic scifi prompting doesn't work that great you'll need decent prompting like [Style] [camera angle and distance] [scale] [subject] [action] [detailed background] [cinematic film, atmospheric haze, color grading] [mood] with futuristic sci-fi tags, with v2 ill try caption so it sticks to the styles a bit better. with a script from an amazing lora creator on civitai here @LolTurdFerguson check her profile out she has amazing loras.
Check my profile gallery or at the bottom of lora page to see lot's more examples for you to use as prompt references. also for other interesting lora's.
All images have prompts and ComfyUI Workflows to recreate it
If you use this lora in merges or upload to other generation sites please mention my civitai page on your post, Thankyou.
https://civarchive.com/user/dkpc69
Description
Trained @10,000 Steps, Rank64, Captioned with Gpt4.1
Dataset of 524 Images
Prompt exactly as i have in the images
works perfectly with turbo
FAQ
Comments (6)
I absolutely love your loras. I don't know if you just know where to find these sick photos or are an artist, but these work a lot like the Vanguard model series or many other detail-type loras -- but are incredibly smart. The sense of scale and detail and variety each of your models bring is awesome. I have a whole power lora loader of just your models and flip em like I'm a mad scientist lol
Thankyou so much it means alot, thats really cool to hear, I consider myself pretty average at this im still learning lol My datasets are really good but krea is an amazing model though, once i have finished training the lora, alot of the time ill see an image i like through insta, tiktok, or youtube and get Chatgpt to write me a prompt so i can try recreate the image, using one or more of the featured prompts/ Captions through my dataset as a reference, what i have noticed is i use alot more steps than most people and larger datasets, which can be bad for certain styles i have alot of failed loras i refuse to release because they're not up to standard, Thanks again though =)
just wish i knew how to finetune a model so i could add all my datasets to a krea or qwen model, I feel like the results will be much better.
@dkpc69 You could use fizgig (https://github.com/shootthesound/Fizgig) as a good start into learning how to prune loras and clean up data you don't want weighted. Locally or runpod or wherever you already train (likely). There's not much difference between finetuning a model or a lora, generally just a flag to indicate the output and differences in configuration. If you're already using larger datasets and longer step counts, that's kind of the default assumption with full blown finetunes. I've seen people have good results simply keeping all the same datasets for their loras and simply throwing it all at a finetune. You could try both: 1) taking the loras you currently have, use fizgig to repair or fix them, and then slowly merge them together one at a time. You'll effectively get the same thing as a full blown extracted model tune. 2) You could throw all your raw data into a fizgig model tune, and do the repair again at the end. You'd have several different ways to compare too.
@makiaeveli Intresting, will definitely look into that sounds promising, have saved a few fizgig Reddit posts over the last month with the intentions to go back and see what it’s all about, Thanks for that bit of info
@dkpc69 I downloaded it a month or so ago and trained a couple loras, and am like halfway through a pretty big dataset. It works quite well. The first time you launch it seems a little sketch cause it launches so silently. But once I used it for a few weeks i really came to appreciate how nice some of the features are. For you it should work at least as a way to explore and compare and repair your current loras. Hope you try it out :)