Thank you for all your support!
Welcome to the Next Era of Finetuning with Flux1!
[ See Before and After ] I’m back and diving into the new wave of model finetuning with Flux1. If you’ve enjoyed my previous work, you’re going to love what’s coming next. To help you get started, use the following workflow with Scheduler Beta at 30 steps: https://rpgv6.byecho.ai/RPGv6-Beta3.json
Want to help? Get free access to Flux1 and GPT-4o by claiming your key on my RPG Discord.
Dive into building entire RPG campaigns, character sheets, artwork, landscapes, and more. Or, join me as we work on training RPG v6 together. We’re a new and growing community, pushing the boundaries of RPG prompting. We're on the lookout for leaders and top-class prompters to join us on our Discord!
Iteration Process
If you've been following my RPG v5 series, you know I’m all about my iterative process. I typically go through around 35 iterations before finalizing a release.
Model Details
Fully trained on 1024-pixel images
100% original content, no merges
Focused on RPG character creation
For Fellow Training Enthusiasts
I’m sharing a Docker template of SimpleTuner for RunPod with all my fine-tuned settings. Perfect for those looking to replicate or build upon my process.
Past release: RPG v5 for SD 1.5
Description
The first 4,000 training cycles on 5 iterations of fine-tuning are now complete, with the Skaven class being used as the baseline.
FAQ
Comments (12)
Is it just me? The download shows at 8mb. That can't be right.
Only one race for now, it's still very early. Here is a before and after the training: https://rpgv6.byecho.ai/01.jpg
Flux LoRAs are surprisingly small.
It's already growing, new update is bigger... :)
@Anashel Should I delete this comment? I honestly first thought civitai was broken or something. Then I realized that flux lora's are just surprisingly small. I appreciate the response, and moreso your effort, i'm a big fan. I didn't intent for this comment to stick around in perpetuity and be a drag, I left it up in case anyone else thought like I did initially, but...that time has certainly passed. Lovin the big release. I'll post some images in the next day or so.
@gurilagardnr Awesome! Thanks for the kind comment. :) You can leave the comment, I don't see it as negative. It's a legitimate question and the information in this thread is relevant.
Can i ask what you used to train these? Hoping its possible on a 3090 in windows lol
I use SimpleTuner, I will upload a template un RunPod with my configuration so anyone can give it a try. Need around 40 GB VRAM and at least 30 GB RAM, but this is my early observation. They have a special update that can bring it down to 20 GB VEAM but was not able to make it work.
I made a full video to help you start, as well as a RunPod template with all my tools and setting on it. https://www.youtube.com/watch?v=UOzPxS-bMFo
I run a 3080 Geforce RTX 12 GB and it works just fine.
Let's Gen This!
Model: Flux dev fp8_e4m
Flexibility: Although I did not struggle to get a skaven to trigger, attempting to get the skaven to show up in various prompts with varying levels of strength and guidance is inconsistent.
Adherence: With the Lora at full strength, the adherence does not appear to be diminished. Although with some bleeding (not at fault of the lora i believe) I am able to get some interesting results with a live rpg battlefield vibe.
Consistency and Variation: This lora is definitely has a good balance of both consistency of style and atmosphere and a good amount of variation between seeds and same prompts. This may be the most important as well for img to video.
It's beta and barely started so I think this is a fantastic start to what will hopefully be an awesome kickass model.
Oh thank you!!! I am hopeful I will soon enough find the right balance in learning rate and dataset variation to get a better consistency.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.