By downloading this model, you agree not to use it for any commercial purpose, including selling images generated using this model. If you are interested in commercial licensing or commissioning a model, please contact me and we can discuss your use case.

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!

Overview

This is a multi-concept amputee model for Flux.2 Klein 9b.

How to Prompt (Please Read)

Klein 9b strongly prefers prose based prompts.

This model does not utilize a trigger word, and in general.

Describe limb differences in a semi-clinical fashion. "She has bilateral hip disarticulations" works much better than "She is a dhd amputee" since common abbreviations like "dhd", "dae", etc don't encode to vectors with reasonable relationships during text encoding but the text encoder absolutely understands hips, shoulders, and disarticulation as English words.

Plan to spend a decent amount of words and sentences on the limb schema. Generally each caption contains 1-3 sentences of information, and sometimes more for complex/inconsistent schemas.

Limb Differences

Here are some examples of how to describe some limb differences:

• She has bilateral {hip|shoulder} disarticulations

• She has a {left|right} {above|below} knee leg amputation

• She has a right {hand|foot} amputation

• She is a limbless amputee. She has bilateral hip and shoulder disarticulations.

You can try describing the length of residual limbs, or where they end. Generally the model gets it to a point, but the more picky you are the lower the hit rate.

Fixing Extraneous Amputations

The model sometimes removes more limbs than you asked for, especially when you request asymmetric schemas like ooe, since like most models, Flux.2 Klein expects lateral symmetry in human bodies.

If you find this happening, the fix is pretty simple: describe the intact limbs clearly. Where are they? Is the elbow bent? Etc. Add details about them until they are showing up reliably. With a good prompt, I can usually get an 80-90% hit rate on asymmetric schemas.

Lefts and Rights

Unfortunately, like many other models, Flux.2 Klein is not good at lefts and rights, and I have been unable to repair this via fine tuning, despite careful captioning and consistent use of stage directions, which is the best practice. This likely won't be fixable with Flux.2 Klein 9b. You can always flip the image horizontally after the fact if you need a certain outcome.

Scarring

While this model was trained on some details related to scarring, it is in general less reliable with scars than my Qwen model. The Flux.2 model seems to have priors about scars baked in that are more graphic/recent than typical amputation scars.

Here's some captioning advice:

• For disarticulations, refer to the scarred areas as "amputation sites".

• For limb or hand amputations, refer to the scarred areas as "residual limbs"

This was done consistently, and helps reinforce the differences between disarticulations and amputations with a residual limb.

Sample phrasing:

• She has bilateral shoulder disarticulations with faint scarring at her shoulder amputation sites

• She has bilateral hip and shoulder disarticulations with faint scarring at her amputation sites

• She has bilateral above-knee amputations with short residual limbs. She has bilateral above-elbow amputations with very short residual limbs. She is supported by distal ends of her residual arms and legs.

• Her amputation sites are smooth.

• Her residual limbs are smooth.

Finally, some images in the training set were described as "Electronic Surgery Images" if they were built in photoshop. I have not experimented much with prompting for this, but it's something you can play with if you have some nostalgia for that vibe.

Prosthetics / Assistive Devices

Unlike my previous models, this dataset does include wheelchairs, crutches, and prosthetics. I wouldn't say that it's exactly great at making all of those things fully consistently, but many people asked for them so I included them in the captioning pipeline and dataset.

Photorealism

Flux.2 Klein is decent at photorealism if you prompt for it well, and this Lora generally biases towards a photorealistic style. It combines well with Loras like SNOFS and Ultra Real if you want their brand of photorealism.

About the Sample Images

All sample images use the simplest Flux.2 Klein 9b ComfyUI workflows, generally at 12 steps with Euler or Res Multistep samplers. Lora weight 0.8-1.0.

Training Details

This model was trained with a new, ~30k image dataset. Captioning was performed by an automatic system that involved a new limb schema classification model that I built along with other data sources / priors. The dataset was balanced and overfitting was monitored to avoid identity memorization. In general, it is significantly more flexible than models trained on the older datasets that I used for Qwen/Flux1 training.

This model was trained with a new regularization set that was balanced to include non-photographs, images with text, low quality images, etc as well as some basic anatomical NSFW concept awareness (though this is not a good model for NSFW alone, training on some concepts increases its potential when combined with other Loras).

Limitations

Flux.2 Klein 9b is a great model, but it has its limitations. With my Qwen model I made a lot more progress at improving the base model's behavior while adding the concept. With Klein, this more or less behaves like Klein but with amputee capabilities. Klein has general challenges with anatomy, background consistency, and facial variety and generally benefits greatly from using LLMs to expand your prompts and flesh out details. Nonetheless this model is a lot of fun, and when combined with the speed and aesthetic capabilities of Flux.2 Klein 9b, you will likely get better results faster when compared to Qwen Image.

Commercial Use Policy

By downloading this model, you agree not to use it for any commercial purpose, including selling images generated using this model. If you are interested in commercial licensing or commissioning a model, please contact me and we can discuss your use case.