<h1 id="20241130v2.6-wo4q9uvdm">20241130v2.6</h1>Due to "flagged for review", many images will be delayed by many hours before they can be viewed.Miiyamoto Hikari from SUMMER LESSONTrigger Words: miiyamotohikariTrained on Pony Diffusion V6 XL checkpoint.I have set a moderation on the gallery, so the NSFW content might not appear.<h2 id="update-m89b56psx">Update</h2>Improved the character features to make them more similar to the original style. Overcame the flatness issue caused by changing checkpoints in the previous version. If the facial lighting is chaotic, you could include 'backlighting' in the negative prompt.However, the current tone is slightly yellowish, which can be addressed by using other checkpoints.<h2 id="usage-tips-wvmsoaqyn">Usage tips</h2><ul><li>You could find the example prompts from my post.</li><li>adding such as "realistic", “photorealistic”, "photorealism" to the positive prompt and even making its weight higher to maintain the original style and features as much as possible.</li><li>My prompts are basically composed in the order of [character traits] + [style] + [expression] + [clothing] + [camera and action] + [background], and you can delete or modify them as needed.</li><li>Recommended weight: 1~0.6, adjust as needed until the character's appearance meets your requirements.</li><li>Upscale value recommendation is around 1.3~2.0, denoising strength is 0.2</li><li>Facial distortion may easily occur in situations such as full-body shots. If there is facial distortion, consider using ADetailer for repair.</li></ul><h1 id="20241101v2.5-peu7zzf53">20241101v2.5</h1>This version has better consistency, but it has become a bit flatter.<h1 id="20240819v2-p27ss6bug">20240819v2</h1>Trigger Words: miiyamotohikariTrying to redo it. And adding some new outfits.The training of this model and the images it generates are solely for learning purposes.You could find the example prompts from the images above.Recommended weight: 1.0~0.6, adjust as needed until the character's appearance meets your requirements.Upscale value recommendation is around 1.5, denoising strength is 0.2You could add "3D" in the negative prompt to reduce the model's 3d style.If not added, it can make the result more closely resemble a in game style.When the character's features become very flat or there is a loss of detail, consider adding such as "realistic", “photorealistic”, "photorealism" to the positive prompt and even making its weight higher to maintain the original shape and features as much as possible.If the style of the 3D model is very rigid consider reducing the model weight.If you are also interested in this character and have clearer screenshots of her other costumes, please consider sending some to me for dataset.The version this time has reached a balance that I am satisfied with, both in 3D and non-3D styles.<h2 id="log-pkgttaahi">LOG</h2>I added two AI-generated images to the dataset. I've tried this method before, but I added too many images, which led to the model overfitting quickly, resulting in a very rigid and lifeless style.In the following two versions, I didn't add any AI-generated images, and the styles in the dataset were basically consistent with 3D styles. However, the trained model was very stiff in 3D style, and the style became very flat after putting 3D in the negative prompt, losing a lot of details.My goal is to train a model that can maintain most of the character's features without being too stiff.Therefore, some images of other styles still need to be added to the dataset, but how much to add is a question.The current problem with this model is: When "3D" is added to the negative prompt, the images generated by the model will lose a lot of details, and the line style is obvious; it is suggested to put tags such as "realistic" and "photorealism" into the positive prompt, and to lower the model's weight, around 0.8 is better;When this character grins, the teeth are very prone to problems.There are still many issues, and I believe that future training based on flux will yield better results.<h1 id="20240807-v1-61pd9d74e">20240807 v1</h1>I have tried repeatedly modifying the tags of the dataset, and controlling variables by adding or deleting some images.In the end, one of the the factors that had a more significant impact was setting of the value of Scale weight norms. This time I set it to 1, and I had tried values of 2.5, 3.6, and 5 before. The results showed that the fitting phenomenon intensified with increasing values, and conversely, the model's output results became more and more flat.The LR Scheduler chose cosine with restart, with 3 restarts. Currently, this model has achieved a relatively good balance, but I don't like the current style of the model. It feels a bit greasy, insubstantial, and lacks stereoscopic.However, if want to make further improvements, I believe that changing to a different checkpoint would have the most noticeable effect.<h1 id="20240804-v0.5-hrngz0vdy">20240804 v0.5</h1>This is a version that is not user-friendly.After accumulating some experience in training models, I felt that I could train the past models better, but the reality proved that I was still naive. This model has not yet met my expectations.The main problems with this model currently are:<ol><li>When "3d" is added to the negative prompts, the model's output results become very flat, and the character's features are easily lost;</li><li>Without adding "3d" to the negative prompts, the model's output results are very rigid.</li></ol>In the future, if my knowledge has improved, I may continue to train this model.I cleaned up all the blurry images, inconsistently styled images, and AI-generated images (these types of images will overfit quickly) in the dataset, and then divided the dataset into two parts, one part is 3D models, and the other part is more game-style oriented. I set the game-style oriented part's training repeating as 2, with the aim of maintaining the character's features;I optimized the caption and fixed many errors in the past.

Miyamoto Hikari PonyXL

1101miiyamotohikari.safetensors

Files with SHA256: c18f229fb1ebd4fa28c7eeb55d47e927d2b1a68e7c19fc9e15159d934e2c09ce