If you use my models, please take the time to post some images to the pages so I can see how they are being used and make improvements in the future. Also feel free to reach out if you have ideas for future models!
What this is for
This model is a concept-injection model intended to help you create images with Puffy areolae — raised, dome-like areolae that project forward from the breast surface as a soft mound around the nipple.
While you can use it alone, it’s best used in conjunction with a base model that provides the style and compositional capabilities that you need.
How to Prompt
Flux.2 Klein 9B prefers prose based prompts.
Training used a mix of vague, medium, and detailed captions, so the model should be flexible to levels of detail. In all captions, this concept was woven into the prose using the term “puffy nipples”.
I strongly recommend combining this model with a fine-tune style LoRA like SNOFS.
What is a concept injection model?
There’s a fundamental conflict between broad fine-tune style models and maximizing concept fidelity for a narrow concept. Larger scale fine-tunes always blur the concepts more than you’d like.
Broad finetunes can fix composition, house style, and add new capabilities, and maintain flexibility, but don’t maximize creative control over fine-grained topics.
Narrow, focused models, what I am calling “concept injections” can reproduce detailed concepts reliably, but can’t effectively target the broader concerns, and reduce flexibility around that concept.
I think there is value in training both types of models (and I do), and from a research standpoint, I’ve developed training processes that work for both.
The main challenge with concept injection models is “off-target effects”. Whether this is facial biases, glamourization, or compositional inflexibility. Some amount of this is going to come with the territory, especially for concepts for which finding or generating diverse data is difficult–but we do want to minimize it. By separating training and inference work into two categories: a broad base model, and narrow concept-injection models, which are then combined at inference time, we can produce better outputs with less overall compute investment.
All of that is a long way of saying: for best results, don’t use this model alone. Pair it with a great fine-tune style LoRA like SNOFS.
Training Details
This model was trained using ai-toolkit with a hand-curated dataset.
I first developed a generic approach to concept injection training. Over 100 models were trained in order to sweep hyperparameters, dataset prep approaches, adapter types, prompting/captioning styles, etc. 95% of the compute in this project went to experiments, not the final models, which can are each able to be trained in 60-90mins on a single RTX6000 Pro. This framework was designed to support training and evaluating concept injections as efficiently as possible, while also retaining enough repeatability that concepts can be re-trained predictably as new experimental results become available.
An experiment framework was developed where candidate models are evaluated using blind pairwise comparisons and then ELO ranked for both on-target and off-target effects. In each case, the pareto-optimal model was selected. This was applied in aggregate at the start to develop the procedure, and then again at a smaller scale for each concept to narrow in on the highest quality model.
The training settings are not exotic. Many oft-recommended techniques were evaluated in isolation including LoKR, adaptive optimizers, LR schedulers, batch size, and regularization regimes. While these can be very useful techniques when doing a broad fine-tune over a long period, they were all measured as statistically detrimental in my experiments. To some extent this makes sense–sharp concept injections are at odds with stability and regularization. Most of the interesting levers pulled related to captioning, trigger word selection, and image curation.
This model was trained exclusively on imagery depicting adult human subjects. It is intended only for generating images of adults. Any use to the contrary violates both the model’s intended purpose and Civitai’s policies.
Description
FAQ
Comments (2)
do you have training tips? i tried to train a small breasts lora and always got some normal default ai looking flux breasts. 2500 steps rank32, body shots.
Ideally you want training data that looks like what you're trying to make. Focus on diversity of subject/setting. No one person repeated more than once plus avoiding "mistakes" that are off-target, or otherwise introduce badness. Generally favor 100-200 images for this kind of Lora to prevent overfitting or memorizing faces. I trained this one for about 4k or 5k steps at bsz=1.
Also, think about the flexibility you want it to have. If you want to support both short prompts and verbose prompts, create that distribution in the training data. If you want it to work paired with other loras, eval it paired with others. If you want it to work at different resolutions or aspects make sure that's represented in your training data. If you want to support racially diverse subjects, train on a mixture with some diversity. Etc.
This is kind of a principle of all NN training--make your training data look like the task you're going to do with the model afterwards. Anything you do (e.g. training only on body shots) that departs from the task is going to cause trouble.
I generate thousands of eval images across ~10-20 candidate models per concept and then blind-rank them and publish the model with the best ELO. I've built systems to automate this, but I always do the ranking and dataset-eyeball by hand. The results often surprise me. I might feel that going from 2k..5k..10k..25k steps is improving, but the ELO winner checkpoint ends up being like 8k or something, and all the >8k checkpoints are doing is overfitting and losing diversity. I've done 100k+ step training runs and ended up choosing relatively early checkpoints. If you're not doing honest experiments you just can't know what's going on. I
I made a lora for your topic, It's not quite as good at this one, mainly due to some biases in the dataset, and I haven't bothered to go back and rebalance/fix that. Also I feel like CivitAI has a hair trigger when it comes to images of women with smaller breasts and I don't enjoy being moderated.