DucHaiten-Pony-Real
Creator: DucHaiten
The model is trained on pony-no-score, but is aimed at a realistic style.
Updated on July 5, 2024:
Finally, Pony real v1 is complete, earlier than expected since Pony v6.9 is still about 3 months away from release, so further training is no longer meaningful. I will stop and decide to wait for Pony v6.9 to be released. In the meantime, I will temporarily switch to training other models. Although the training of the Pony real model has stopped, it does not mean it has been abandoned. In fact, with a very large data source and thorough training, to some extent it can be considered another Pony base model, different from the Pony base model of PurpleSmartAI. Of course, it is not perfect; no model is perfect. Therefore, everyone can combine it with various Loras while waiting for the next updates. But I can say it is still very flexible and retains the most essential aspects of Pony, while fixing a bunch of bugs from Pony base v6.
I have worked very hard to balance between a realistic style and effects, characters, poses, anime, surreal, or anything related to art in general. It is very difficult, so sometimes I had to make some compromises. It will not be as real as photos you can freely take in real life; this model aims for the realistic feeling you get in movies that use a lot of realistic effects and ultra-realistic Unreal Engine graphics. I have also clearly defined the keywords, so you will need to type detailed and accurate prompts to get clear and precise styles or effects, or what clothes, hairstyles, etc., the characters have. I have trained it in many natural languages and separated two types of captions corresponding to two ways of typing prompts into two different clips, making it more diverse in terms of styles, prompts, and keywords. It won't follow a template, be monochromatic, or single-styled. But of course, it still focuses more on 3D and realism, because that was my initial goal
You can choose not to add tag_score to the prompt, but adding it will still make it look better.
Oh, one last thing, beautiful Asian girls are not produced on the same doll manufacturing line.
The best settings on automatic1111 are:
- Sampling method: DPM++ SDE (slow but makes the image more realistic, with fewer errors, and closer to the prompt)
- CFG Scale: 9 (fewer errors)
- Sampling steps: from 30-40 (starting from step 36 onwards, the image will have more accurate details and fewer small errors. The higher the steps, the better these small details will look. However, the higher the number of steps, the more likely color errors and contrasting color patches will appear.)
Update on June 23, 2024:
Can someone please remind AstraliteHeart to quickly release pony v6.9, I want to use it to train pony real v1 and pony no score v5.
In this Beta 8 version, I improved the colors to be more vivid, with a bit of gold and cinematic depth. More artistic data is trained, improving the natural language part.
Update on June 15, 2024:
Why hasn't the Beta-4 model become as realistic as other pony real models so far? This is entirely intentional, or you could say that I have chosen a much more challenging path: transforming it into a real pony model while still fully retaining surreal details, beautiful effects, and famous characters in the pony base. However, you can completely enhance its realism by using keywords related to realistic style and quality, as well as adding some opposite keywords in the negative prompt section. Additionally, try using natural language prompts combined with some NSFW keywords if you wish.
Update on June 8, 2024:
alpha 7 focuses on making image textures more realistic, which has resulted in some non-realistic pony details being reduced, along with the addition of tag_scores which are now mandatory.
Update on June 3, 2024:
- Alpha 5: A lot of new data has been added, but I focused on fixing bugs and increasing stability first, so the model's realism has only improved slightly.
- Fixed Noise Spots: Resolved the common noise spots issue seen in pony models. This bug occurred when creating images with certain keywords or combinations of keywords that the pony model didn’t like. Yes, no more inexplicable noisy spots in the images.
- Reduced Dependence on tag_score: The influence of tag_score on image quality has been reduced to about 10%. Sometimes, it may even worsen the results.
- Still in Alpha: This is still an alpha version, so many issues still need to be addressed. Please let me know your feedback after trying this Alpha 5 version.
DucHaiten-Pony-Real
Creator: DucHaiten
The model is trained on pony-no-score, but is aimed at a realistic style.
Updated on July 5, 2024:
Finally, Pony real v1 is complete, earlier than expected since Pony v6.9 is still about 3 months away from release, so further training is no longer meaningful. I will stop and decide to wait for Pony v6.9 to be released. In the meantime, I will temporarily switch to training other models. Although the training of the Pony real model has stopped, it does not mean it has been abandoned. In fact, with a very large data source and thorough training, to some extent it can be considered another Pony base model, different from the Pony base model of PurpleSmartAI. Of course, it is not perfect; no model is perfect. Therefore, everyone can combine it with various Loras while waiting for the next updates. But I can say it is still very flexible and retains the most essential aspects of Pony, while fixing a bunch of bugs from Pony base v6.
I have worked very hard to balance between a realistic style and effects, characters, poses, anime, surreal, or anything related to art in general. It is very difficult, so sometimes I had to make some compromises. It will not be as real as photos you can freely take in real life; this model aims for the realistic feeling you get in movies that use a lot of realistic effects and ultra-realistic Unreal Engine graphics. I have also clearly defined the keywords, so you will need to type detailed and accurate prompts to get clear and precise styles or effects, or what clothes, hairstyles, etc., the characters have. I have trained it in many natural languages and separated two types of captions corresponding to two ways of typing prompts into two different clips, making it more diverse in terms of styles, prompts, and keywords. It won't follow a template, be monochromatic, or single-styled. But of course, it still focuses more on 3D and realism, because that was my initial goal
You can choose not to add tag_score to the prompt, but adding it will still make it look better.
Oh, one last thing, beautiful Asian girls are not produced on the same doll manufacturing line.
The best settings on automatic1111 are:
- Sampling method: DPM++ SDE (slow but makes the image more realistic, with fewer errors, and closer to the prompt)
- CFG Scale: 9 (fewer errors)
- Sampling steps: from 30-40 (starting from step 36 onwards, the image will have more accurate details and fewer small errors. The higher the steps, the better these small details will look. However, the higher the number of steps, the more likely color errors and contrasting color patches will appear.)
Update on June 23, 2024:
Can someone please remind AstraliteHeart to quickly release pony v6.9, I want to use it to train pony real v1 and pony no score v5.
In this Beta 8 version, I improved the colors to be more vivid, with a bit of gold and cinematic depth. More artistic data is trained, improving the natural language part.
Update on June 15, 2024:
Why hasn't the Beta-4 model become as realistic as other pony real models so far? This is entirely intentional, or you could say that I have chosen a much more challenging path: transforming it into a real pony model while still fully retaining surreal details, beautiful effects, and famous characters in the pony base. However, you can completely enhance its realism by using keywords related to realistic style and quality, as well as adding some opposite keywords in the negative prompt section. Additionally, try using natural language prompts combined with some NSFW keywords if you wish.
Update on June 8, 2024:
alpha 7 focuses on making image textures more realistic, which has resulted in some non-realistic pony details being reduced, along with the addition of tag_scores which are now mandatory.
Update on June 3, 2024:
- Alpha 5: A lot of new data has been added, but I focused on fixing bugs and increasing stability first, so the model's realism has only improved slightly.
- Fixed Noise Spots: Resolved the common noise spots issue seen in pony models. This bug occurred when creating images with certain keywords or combinations of keywords that the pony model didn’t like. Yes, no more inexplicable noisy spots in the images.
- Reduced Dependence on tag_score: The influence of tag_score on image quality has been reduced to about 10%. Sometimes, it may even worsen the results.
- Still in Alpha: This is still an alpha version, so many issues still need to be addressed. Please let me know your feedback after trying this Alpha 5 version.
