I encountered quite a few issues during my training and use of Slider-LORA, but I found that most of them can be avoided manually. The following summarizes four key points to help you better use Slider-LORA and unleash its potential.
Premise: The Essence of Slider-LORA
Actually, it is just a collection of prompts with adjustable weights. How to understand this statement? In the SD15 era, people would adjust the weight of a certain incantation (such as the classic 1girl:1.5 or 1girl:-1.5), and Slider-LORA is essentially this 1girl, except that its prompt structure may be more complex, and then it is packaged into a preset, so that every time you want to achieve a specific effect, you only need to load the preset instead of memorizing that bunch of complex prompts. Please keep this premise in mind, as it will help you use Slider-LORA better.
Point 1: Effect Reproduction
Why is it difficult for me to reproduce the effects demonstrated by the author? This is a common issue faced by most people who are new to Slider-LORA, and it is usually related to the input prompt. For example, if you need a close-up shot of a character's face, you would describe the character's facial features, but you wouldn't write about what kind of shoes the character is wearing because it has nothing to do with the close-up. The same principle applies when using Slider-LORA (since its essence is a collection of prompts). Take my <ZoomIN&ZoomOUT> as an example, which achieves a smooth transition from close-up to long shot through ± weights.
When I need a close-up, I adjust its weight to +5 and then describe the character's facial features as much as possible. Never write any description that goes against your expectations. This seemingly simple rule is the main reason why many people cannot reproduce Slider-LORA.
Point 2: Weight Dispute
Traditionally, the weights of Slider-LORA should be between ±2, but most people prefer ±5 (mine is like this), because it allows for more precise control of the image. For every ±1 point of weight, you can see subtle changes. Therefore, when you get a Slider-LORA, please read its guide first, which usually indicates the weight range. However, this does not mean that weights outside the range are unavailable; those are just recommended values. The effects of weights outside the range may be more amazing (or more terrifying). Just take a bold try, and all you'll lose is the time for one rendering.
Taking my <body> as an example, it is used to adjust the body fatness or thinness of the character. My recommended value is also between ±5, but in fact, you can adjust it to a larger value. The following image shows the effect of ±7. Yes, you can clearly see the flaws, but the character is indeed thinner (or fatter). If you want this extreme effect in your work, then you need a corresponding workflow.
Point 3: LORA is not a panacea
To unleash the potential of Slider-LORA, you need to create a dedicated workflow. Take my <focus> as an example, which is used to precisely control the focus with a weight of +5, resulting in foreground blurring. The left image shows the result of ZIT's regular workflow, and the right image shows the result of the dedicated workflow I designed for the PhotoBOX series. The effect is obvious, right? So, don't rely entirely on LORA. It's like your right hand, and the workflow is your left hand. Only by working together can you create more beautiful images.
Point 4: Don't be afraid of stacking multiple LORAs
Below are the results of my <ZoomIN&ZoomOUT> + <Dark&Light> + <Warm&Cool> + <Low&High>, with all weights above ±3. Can you see any degradation? The superposition of multiple Slider-LORAs is like the superposition of multiple sets of prompts; it won't cause degradation in your image as long as they are not contradictory.
Mixing various Slider-LORAs together is one of my pleasures. They have the potential to create a brand new visual experience (or perhaps a worse one), but this uncertainty is exciting. This is my collection of Slider-LORAs
Finally, I'd like to add a user's feedback: in extreme cases (greater than or equal to ±5), the character tends to have an Asian face. This is related to ZIT's training, but it's not insurmountable. You can enhance the description of the character's facial features to better meet your requirements. Alternatively, you can add more information through the workflow. The following is the optimization done using the PhotoBOX workflow.
The left image shows the effect before optimization, and the right image shows the effect after optimization.
|🙈🙉🙊OK, that's all for the content
If this article can help you, please give it a like so that more people can see it, which will help more people. Of course, comments are also welcome for discussion
Finally, may you have a great time! AI never sleeps! Peace!