RealFeet for SD 1.5 – Feet and painted toes

The Lora model is specifically trained to render feet for SD 1.5. The same Lora is available for SDXL too.

If you like nylon feet instead, you can download RealNylonFeet for SD 1.5.

The native 1024x1024 was selected to train a performing the XL Lora, with reliable layout.

Due to many requests received, several training runs were performed to assess if an SD 1.5 Lora should be set with a good performance.

So first, if you like this Lora, please review it, use it and share pictures.

Taking into account that with SD 1.5 the AI feet rendering is still challenging, this Lora is at the moment my best result:

please read the instructions in the GENERATION FLOW paragraph at the bottom of page to maximize results!

Triggering prompt keyword:

RealFeet

Lora weight:

A value of 0.7 or 0.9 is highly recommended. Higher values can trigger artifacts.

So a good activation keyword for lora is:

<lora:RealFeet:0.8>

This Lora could be used both with realistic or manga checkpoints (see image below).

Concerning realistic checkpoint Realistic Vision was used to generate the demo gallery.

Sampling steps:

About 45-60 sampling steps.

Additional istructions:

-          There are some different particulars than can be rendered, like painted toenails (preceded by a color name, eg. red painted toenails) or natural toenails that can be triggered by using the following tokens:

color painted toenails

natural toenails

-          You can specify that feet must be rendered on tiptoes including token

on tiptoes

WARNING: The details level of this Lora could drop in case of full body portraits (very little feet) or in case of many different Lora used together. Details can be rescued using inpaint algorithm only on feet, with a very specific prompt focused only on RealFeet triggering words like in the following example.

Raw Photo, RealFeet, red painted toenails, 8k uhd, dslr, soft lighting, high quality, film grain, Fujifilm XT3

Image dimensions

Better result are obtained with a base image of about 1200 px dimension on the long image side with an image width/length ratio of  1.5. After the base generation upscaling is a good practice.

Follow these following tips in  GENERATION FLOW:

1) ControlNet is strongly recommended, during the first txt2img generation:  canny or openpose approach is strongly recommended to avoid gross aberrations.

2) After a first round of txt2img, the best image should be refined using an img2img round, eventually adjusting the strength of some tokens.

3) Finally  inpaint will be fundamental to refine feet details. If checkpoint include a specific inpaint model, its use is recommended.

A specific prompt is recommended for this step, focusing only on feet. For example:

<lora:RealFeet:0.8> RAW Photo, RealFeet, red painted toes

USE DEMO IMAGES PROMPTS AS REFERENCE

Final statements:

1. You are solely responsible for any creative work using this model.

2. You cannot use this model to intentionally produce or share illegal images or content.

3. Furthermore can't use this model for malicious, hurtful, defamatory, fraudulent or political purposes.