Sign In

SDXL Lora Training: Start until now (FINDINGS) *(Outdated)*

SDXL Lora Training: Start until now (FINDINGS) *(Outdated)*

Originally Published: October 14, 2023

This is an archived SDXL findings article providing insight into the state of affairs for Duskfallcrew and earthnicity's training methodologies pre-December 2023.

Presented without headings, the content reflects a dialogue of unconventional approaches amidst a backdrop of shared experiences:

"Well, they suggested this method, but didn't explicitly advise against another" - a sentiment echoed through various experiments pushing beyond conventional parameters.

Recent observations have been intriguingly paradoxical; the more I revert to 'old habits' reminiscent of SD 1.5 training, the more stability seems to emerge.

Reflecting on a previously penned, admittedly opinionated guide from just a month ago, it becomes apparent:

The only absolute rule amidst the subjective terrain of training is legality. Avoiding illegal activities is a non-negotiable, albeit humorous reminder—steering clear of certain topics is just common sense.

In summary, while opinions abound, and guidelines may vary, adherence to legal standards remains paramount. Now, onwards.

Let's take a look at a couple pictures of what would be considered "STANDARD" SDXL training:

Don't ask me what my settings were, i've been using MOSTLY standard settings via CivitAI trainer laetly.


And as a point of reference, there's POLTERGEISTS XL, which I've been hesitant to introduce to CivitAI because it simply refuses to cooperate. This model was trained by bmaltais back in August.

For context, both models had fewer than 1000 steps each but boasted reasonably sized datasets. The leading contender was likely AdamW8BIT, possibly the second iteration of the miqote retrain.

Now, let's consider THANCRED as a benchmark, representing a fusion of the fundamental approach with occasional variations, akin to a collaborative effort on a Holostrawberry setting for SD 1.5 within the realm of SDXL.

Moving forward, we encounter models around the 2000-step mark, with uncertain repetitions, confirmed using AdamW8bit. Interestingly, there was initial consensus that AdamW8bit was incompatible with SDXL—a notion quickly debunked through practical experimentation. However, it's worth noting that while viable, its usage isn't consistently recommended as of 2024.

It's important to highlight that despite certain limitations, these models exhibit impressive accuracy, particularly in replicating outfits and similar details. Even when faced with challenges like the refusal to render a gunblade as a cohesive unit, their performance remains fairly accurate, even at lower resolutions such as 1-1 or even 0.5.


There's a prevalent notion emphasizing the necessity of uniformity in dataset dimensions, typically advocating for sizes of at least 1024 on all sides or larger. While it's generally prudent not to overlook this guideline, it's worth noting that SDXL might occasionally emphasize the smallest or poorest quality data points. However, there's a workaround: leveraging a pre-existing SD 1.5 dataset can help mitigate this issue, provided one possesses the requisite expertise. Nonetheless, exercise caution and refrain from excessive reliance on this method, as its efficacy isn't guaranteed.

Furthermore, it's crucial to address the minimum dataset size. While it's technically feasible to train models with fewer than 50 images, SDXL exhibits peculiar behavior when confronted with 3D or semi-realistic data lacking in quantity. In such instances, it tends to struggle, manifesting particularly in attempts to distort or warp certain features.

EXAMPLES:

This isn't even the worst I've seen. My mother, who loves horror films, often says, "Please, child, save those pictures so I can laugh at them." It's a comical counter to anyone who claims AI is flawless and that we're all out of jobs. The reality is, AI can't even consistently render hands and faces correctly, let alone replace human expertise.

Here's what I've learned about using SDXL moving forward:

The only strict rules are those dictated by laws, morals, and ethics. Beyond that, you're free to experiment. If you prefer a lighter training regimen and it works well for you, go for it! If you favor a more intensive approach, that's fine too. I'm simply sharing my frustrations from this small study.

Despite the challenges, especially with getting faces and proportions right on SDXL base models, there's hope. As we progress through 2024, trained checkpoints and SDXL loras will improve. This progress will help overcome the initial limitations imposed by external influences, enhancing the model's capabilities.

In essence, while AI development is far from perfect, the journey is about experimenting within ethical boundaries and finding what works best for you.

About & Sponsored

Curious to explore more? We're excited to unveil exclusive content through our Patreon page: EarthnDusk on Patreon.

For those who prefer a one-time show of support, you can buy us a coffee on Ko-Fi:

Dusk's About:

When we say "we," we're referring to our Dissociative Identity Disorder (DID) system, which comprises over 200 alters. Our system embraces diversity, including individuals with ADHD, Autism, and CPTSD. We recognize the potential of AI in breaking down barriers within mental health, but we also acknowledge its limitations and potential hindrances.

Earth's About:

As a diagnosed flavorful UDD system with a variety of additional conditions like BPD, ADHD, Autism, and PTSD, we proudly identify as a system of over 80 headmates. We navigate our journey without shame, embracing our differences and strengths. It was our collective effort that introduced Dusk to AI training and creation. We firmly believe in the power of AI to overcome artistic barriers.

Our website, End Media, invites you to join our forums and share your creative endeavors.

Dusk's real-life photography is accessible for free via Unsplash, offering a treasure trove of imagery for use in Lora or modeling projects.

We're honored to have the support of Pirate Diffusion, a platform that champions creativity and innovation.

Join our Discord community to engage in discussions about the tutorial and explore other gaming adventures: Discord Server.

Connect with us on DeviantArt: DiffusionAI DeviantArt Group.

For Reddit enthusiasts, join our subreddit: r/earthndusk.

3

Comments