[v3.1 is still in further testing. Updates regarding new findings will be updated in the "About this version" section]UrangDiffusion v3.1 (oo-raw-ng Diffusion) is the first UrangDiffusion version that utilize Animagine XL 4.0 as the base.The name “Urang” comes from Sundanese, meaning “We/Our/I.” The history behind the name is to make the model not only suitable for me but also for many people. Another reason is that I use many resources (training scripts, dataset collecting scripts, etc.) from other people. It’s unfair to claim this model as “my sole work.”<h2 id="standard-prompting-guidelines-hlzesdv89">Standard Prompting Guidelines</h2><ul><li>Prompting guide:<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/04b4aacf-0336-4fd5-a221-ea1e29dcf9b5/width=525/04b4aacf-0336-4fd5-a221-ea1e29dcf9b5.jpeg" /></li></ul><ul><li>Default negative prompt: <code>lowres, bad anatomy, bad hands, text, error, missing finger, extra digits, fewer digits, cropped, worst quality, low quality, low score, bad score, average score, signature, watermark, username, blurry</code></li><li>Default configuration: Euler a with around 25-30 steps, CFG 5-7, and ENSD set to 31337. Sweet spot is around 28 steps and CFG 6.</li></ul><h2 id="training-configurations-wag59pcf1">Training Configurations</h2>Finetuned from: Animagine XL 4.0 Base (NOT 4.0-Zero)Finetuning:<ul><li>Dataset size: ~1,600 images</li><li>GPU: 1xA100 80GB</li><li>Optimizer: AdaFactor</li><li>Unet Learning Rate: 1.25e-6</li><li>Text Encoder Learning Rate: N/A (Turned off)</li><li>Batch Size: 48</li><li>Gradient Accumulation: 1</li><li>Warmup steps: 5%</li><li>Min SNR: 5</li><li>Epoch: 15</li></ul><h2 id="faq-um1wr3zu2">FAQ</h2><ul><li>Q: Images are sometimes noisy.</li><li>A: This is a common issue with Animagine XL 4.0 models in general. The base model is trained with only 10 epochs, which lead to the model being undertrained. Unlike Initial N or Initial I model that are trained with more resources.</li></ul><ul><li>Q: Hires fix model?</li><li>A: Check out the cover image metadata, you'll find it there.</li></ul><ul><li>Q: Initial N/Initial I is better.</li><li>A: Just leave and do not use the model. Simple. No need to announce your departure. Except you're willing to leave a constructive feedback or willing to fund future projects.</li></ul><h2 id="special-thanks-dbxwtlfoj">Special Thanks</h2><ul><li>My co-workers(?) at CagliostroLab for the insights and feedback.</li><li>Nur Hikari and Vanilla Latte for quality control.</li><li>Linaqruf, my tutor and role model in AI-generated images, and also the person behind tag ordering.</li></ul><h2 id="license-0mxe0luju">License</h2>UrangDiffusion v1.0-v2.5 falls under the <a target="_blank" rel="ugc" href="https://freedevproject.org/faipl-1.0-sd/">Fair AI Public License 1.0-SD</a> license, while v3.x falls under the <a target="_blank" rel="ugc" href="https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/LICENSE.md">CreativeML OpenRAIL++-M license</a>.

UrangDiffusion v3.1

UrangDiffusion XL v3.1 is fine-tuned from Animagine XL 4.0 Base (not Zero). This 4.0 Base model serves as the base model pre-trained for the final release of Animagine XL 4.0 (not the Opt version).I have received permission from the team to fine-tune the base model using my own method and release it under the UrangDiffusion series.Base model: Animagine XL 4.0 Base Fine-tuning details:<ul><li>Dataset size: ~1,600 images</li><li>GPU: 1× A100 80GB</li><li>Optimizer: AdaFactor</li><li>UNet learning rate: 1.25e-6</li><li>Text encoder learning rate: N/A (disabled)</li><li>Batch size: 48</li><li>Gradient accumulation: 1</li><li>Warmup steps: 5%</li><li>Minimum SNR: 5</li><li>Epochs: 15</li></ul>Due to some quirks of the model, please keep the following in mind:<ul><li>v3.0 may perform better with anatomy</li><li>v3.1 may perform better with more fluid poses</li></ul>If you encounter anatomical issues at 28 steps, try lowering to 27 or increasing to 29. If it improves but isn’t perfect, continue adjusting slightly up or down. If the result worsens, the previous step count was likely the optimal one.UPDATE [19/04/2025]<ul><li>Some generations are stable at step 30++ with Euler a. You might wanna crank up the steps a bit.</li><li>Some generations are also better with <code>realistic, 3d</code> included in the negative. Try that too.</li></ul>

Files with SHA256: 4243efe3ea7bc5cc7822bf6168fffe0889bf628d7266e31efce5e4916a378454