HandEnhance

It is very difficult to control the fingers in SD1.5.

I had given up on the fingers in SD1.5 in the past, so many of the models I released at one time had really bad fingers. Recently, I tried various ways to improve them, and this LoRA is what I put together as something that could be used generally.

I had tried many prototypes of LoRA for support before, but most of them were not very effective, or were better not to use. Of course, none of them were released. This LoRA uses the same method as the model I recently released, so I decided to release it thinking that it might have some effect.

The final conclusion may be that the fingers during image generation should be corrected with ControlNet, but this LoRA was originally created to adjust models with extremely broken fingers, so please think of it as something that can be used during image generation as well.

The appropriate strength varies depending on the model to which it is applied. A rough guide is around 0.2 to 1.5, and it is better to start with around 0.2. In many cases, adding a small amount is appropriate. There is no benefit to using too many, as the hand may break or fingers may grow in unnecessary places. This LoRA does not create perfect fingers. Think of it as just one more solution using a different method from previous ones. That is how difficult it is to create fingers with SD1.5. If you use this type of LoRA, it may be a good idea to use the Finger Count Slider in conjunction with it.

Below is a comparison with BetterHands and GoodHands, which are similar LoRAs, and BadHand4 and BadHand5, which are Negative Embeddings.