TF/TG M2F LoRA — ALPHA

Note: Please read below for usage notes and prompting tips.

This is an experimental model that can be used to create transformation sequence images. In particular, it was trained on male-to-female (m2f) gender transformation sequences -- men getting turned into women. It's fairly good at that, but can also be used to generate similar types of transformation sequences such as breast expansion and bimbofication. It has a strong bias to starting out the image with a male figure, however, so these results will be even less consistent.

Consistency is the most difficult thing to achieve with the current version of this LoRA. Most of the results posted here are at least somewhat cherry-picked. Some sequences are easier to produce than others; likely the ones that are represented in the dataset. I would estimate that I'm happy with <25% of the outputs.

Nonetheless, I decided to post this while I work on an updated version. Enjoy!

Usage Notes:

• Trained on base SD1.5, and it shows its best and most consistent results there, but any realistic model has a good chance of working. (I haven't gotten it working on other kinds of models.)

• Use 0.5-0.8 strength for the LoRA; higher values have a stronger chance of inheriting some undesirable style characteristics. Lower values can still be good as well.

• The aspect ratio of the image has a strong effect on the output. Certain image sizes will gravitate towards generating a certain number of steps. Wide aspect ratio is recommended, and an aspect ratio exceeding 2:1 works well for 5seq.

• ControlNet is really useful for controlling the output of this LoRA. Supply an openpose input to place the desired number of figures in the output.

Prompting:

• Add 3seq or 5seq to the prompt to guide the output towards a sequence of 3 or 5 transformation steps. It was trained on sequences ranging from 2-7, but most of those outside of 3 or 5 did not have a lot of data, so these are the strongest.

• drawing, digital art, 3d render all seem to be the easiest styles for this LoRA to work with, but they also have a higher chance of bleeding over some styles from the training data. I've generated a lot of photo prompts as well.

• Prompt structure that generally works well: <3seq/5seq>, <style>, <description of man>, <description of woman>, <description of scene>, e.g.

  • 3seq, (masterpiece photo), man wearing a t-shirt and shorts, woman wearing a bikini, beach

  • 5seq, (masterpiece photo), man wearing a suit, woman wearing a dress

  • 3seq, (masterpiece photo), skinny man wearing a shirt and jeans, woman with huge tits, black background

• Generally compatible with other style or character LoRAs, though you will have to play with the strengths.