It's practically the complete checkpoint, there are a few imperfections, but I wish they will be corrected in the final version. This model was a lot of work. I had to manually check more than 120k images, half of the dataset. More than 12,000 images have been replaced.
Almost all the concepts of V02 works on XL. See changes section below.
Yesterday I posted the Dev version on CivitAI, but you have the priority of the schnell version, which requires more processing and time to train.
This Lora allows the Bosom V03 checkpoint to work like Hyper SD 1.5. A test that ended up bringing great results.
Based on ByteDance models, but the precision and weights have been fine-tuned by me.
Finally.
This checkpoint is a very strong personal achievement.
After tubulent periods, I make my last act in SD 1.5.
All the concepts of V02 works on V03
Hello everyone, I have some great news.
First of all, my latest dataset for SD1.5 is almost complete.
I've already made the descriptions for all the subreddits. The number of subreddits has also changed: there will now be 581 subreddits. The number of images after running a script to remove duplicates was 238k.
At this very moment I'm running a multi-modal LLM to categorize the quality and aesthetics of each image, to make it easier at the prompt. This proc...
2024-06-26 15:59:55 +0000 UTC View PostFirst of all, I'd like to thank you all for the support you give me. 🥰
I believe the image collection is finished.
I'm thinking for now of making an intermediate model with just the subreddits and SD 1.5, because I want to do some caption tests. The list has been updated and there are now around 500 subreddits. The list of subreddits that have been scraped 2024-06-06 22:39:20 +0000 UTC View Post
This week I got in touch with another content producer to set up a partnership to make a captioner together. Interesting things are coming.
I also finished scraping Reddit this week. 400 NSFW subreddits were scraped, including 300,000 images. We're now at 1.5m images total.
I feel like I'm at the end of collecting the dataset. When that happens, you'll be the first to know.
2024-05-24 19:49:25 +0000 UTC View PostFor this week's status update, I've brought you some news that sounds scary but exciting at the same time: One million high-quality images have already been collected for the JK Bosom V3 / XL Dataset.
I'm afraid of how much time I'm going to use to make all the captions, and that's where the yolo models come in. This time, my yolo training has brought me another fruit: I've created a breast size detector that will identify the size more accurately than ordinary captioners. So far there ...
2024-05-16 00:52:00 +0000 UTC View Post
A more arduous project than I thought. But it will be useful for the next JK Bosom.
⭐Theoretically it works with all versions of Stable Diffusion⭐
AUTO1111 or SD.NEXT (or similars):
Put the ".pt" file on "models/adetailer" folder of your UI.
If ...
2024-05-08 13:14:08 +0000 UTC View PostMore than 350k images have already been downloaded for the next JK Bosom. My goal is around 1 million images.
So I'd like to know which subreddits you'd like to see in the next version of the checkpoint.
If you follow the announcements on Discord, you know that I'm working to optimize the captioning process. This time I'm making Yolo models for captioning (the same ones used to detect body parts in ADetailer) .
Th...
2024-04-30 23:40:09 +0000 UTC View PostI'd like to understand which variant of Stable diffusion I'm going to devote my efforts to.
2024-04-22 18:28:31 +0000 UTC View PostHello,
As you've noticed, I've been away for a long time. Since I announced that I was working on the new version of JK Bosom, a few things have happened in my life that have made me quite depressed. This depression was so significant that I considered the worst. I keep thinking that we all have problems and I'm ashamed that I wasn't professional and continued my work. The delay in updating the status of my projects while I was depressed made me very ashamed and ended up in a vicious cy...
2024-04-22 18:16:51 +0000 UTC View PostI've started making the JK Bosom V03 dataset, this time focusing on:
The goal is to accumulate more than 20k images. I also plan to make SDXL and SD3 versions with the same dataset. So I'll have to exclude a lot of photos that don't match the training resolution.
Some people reported that they had problems training with version V02.5, so version V02.5b is now on V025 OneDrive, which is compatible with the Kohya scripts for training.
Sorry for the mistake. I was so excited about the results of this version that I forgot to make it compatible with training.
JK CHSTR
2024-02-27 23:35:46 +0000 UTC View Post
Some time after I posted V02, I realized how inconsistent it was. Before I got sick, I did some training to reinforce certain concepts, but it wasn't satisfactory. However, this training was not in vain. I was able to use it as a merge element, along with other interesting models.
Obs: Starting in the next few weeks, I'll be working on V03.
Better faces, especially when they're a little further away.
Better hands
C...
2024-02-22 22:53:32 +0000 UTC View PostI'd like to apologize. I was away for a while because I was ill with a disease that is common in tropical countries: Dengue fever. It was a very troubled week for me. It was the first time I had delirium due to the fever.
I'll be active again this week.
Thanks again to everyone who supports me.
JK CHSTR
2024-02-21 19:23:54 +0000 UTC View Post
After repeated tests, I managed to complete a stable JK Bosom V02. Unfortunately, I couldn't finish all the documentation, because of all the details I want to include there, but you can already see a WIP version. I hope that everything will be described in the next few weeks.
More than 80 concepts were trained this time, which are listed in the documentation. (Some are not yet listed in the WIP). In addition, we no longer h...
2024-02-05 19:57:44 +0000 UTC View PostJK Bosom V02 training has begun. In order to remove some biases I will make corrections to the dataset. But I think we'll have V02 next week or two at the latest.
As I always say, check the discord to see how the model is progressing. I've already posted some previews.
2024-01-18 19:05:46 +0000 UTC View PostI currently have 77% of the captions of JK Bosom V02 done. I hope to start training at the beginning of next week.
Some concepts that will be trained (of at least 80):
For more updates, stay tuned on discord.
2024-01-13 21:40:01 +0000 UTC View PostFirst of all, happy 2024!
After extensive image collection and filtering, the images that will be used for the JK Bosom V02 training have finally been defined. The final quantity, after normalizing each concept, was ~8k images. This week I'll be captioning each concept and I hope to start training in two weeks.
As was the case with V01, V02 will present concepts never before trained in SD, so stay tuned.
Thanks to everyone who continues to support me. Sorry for the delay in ...
2024-01-03 08:16:33 +0000 UTC View Post
First of all, thank you for being with me so far.
I've come to update you on my latest activity: the production of the JK Bosom V02 has begun. This time, the focus is on the entire female body. You can be sure of better controls over hips, asses and pussies. In addition, I'm also going to adjust some controls that were problematic in V01, such as the nipple color selection. Likewise, some LoRAs that I've already p...
2023-12-22 10:28:14 +0000 UTC View Post
I'm studying the creation of negative embeddings, it's still giving me a bit of a headache, but this version is acceptable.
What you can expect:
Problems that are still being...
2023-12-13 20:05:32 +0000 UTC View Post
I apologize for disappearing. Making this model made me very unstable. During production I accidentally deleted the dataset and had not made a backup. It destroyed me inside, I blamed myself and was depressed for weeks. However, I managed to use a previously trained version to achieve this result.
Enjoy. I did it with great care and attention. It has my blood, sweat and tears in it.
JK CHSTR
⚖️LoRA prompt weight: 1
⚙️~60% effectiveness
🏫Model used for training: personal modded model
These models was generated from one of my attempts to make ...
2023-11-02 22:09:25 +0000 UTC View PostThe JK Bosom dataset is 50% finished. Entering the descriptions manually is very tiring. I believe this checkpoint will be unprecedented in terms of the possibilities for breast variations. I won't be adding the actions at first. But they are already on the priority to-do list. There will be puffy nipples at the checkpoint.
Tomorrow I'll send a couple of LoRAs from the project that didn't succeed in becoming a single LoRA (breasts separation)
2023-10-31 07:08:13 +0000 UTC View Post😪TL/DR: Because of the biases of the existing checkpoints, I decided to make my own.
As you've noticed over the last few months, the number of LoRAs I produce each week has been decreasing. This is no coincidence. Basically, I'm challenging myself to make more difficult LoRAs, but these checkpoints wear me out a lot because they have biases that make it impossible to create more complex Low Rank adaptations.
Because of this, I'm challenging myself once again, but to make things...
2023-10-24 03:29:48 +0000 UTC View Post
⚖️LoRA prompt weight: 0.8~1
⚙️~40% effectiveness
🏫Model used for training: personal modded model
I know, the previous model was V05. I put the version of the previous model like this to s...
2023-10-17 07:47:18 +0000 UTC View PostSo I've always been curious about the models you use. It can also help me to know which models I need to focus on when creating new LoRAs.
If your ckpt isn't listed below, please don't forget to comment on which one you use.
2023-10-11 20:00:03 +0000 UTC View PostChoose the next Longer Prompters+ LoRA.
I will finalize this poll on Saturday 14/10.
I apologize for not having done this poll before, I was so focused on the macromastia Lora that I forgot.
Plz, if you have any other ideas, don't hesitate to comment.
2023-10-11 10:32:50 +0000 UTC View Post
⚖️LoRA prompt weight: 0.6~1
⚙️~90% effectiveness (V03a) ~60% e...