Slider LoRA
A slider is a LoRA whose strength you can adjust: weight +1 goes towards “positive”, −1 towards “negative”, and 0 changes nothing. Examples include lighting, expression, age and amount of detail. Both the SDXL family and Anima can be trained.
When generating, a negative weight moves in the negative direction.
Two ways to teach
Section titled “Two ways to teach”On the training page, set “What to train” to “Slider”, then choose under “Training method”:
| Method | Description |
|---|---|
| Words | Write the words to add for positive and negative; no images needed. The direction comes from how the model understands those words, which often changes the style too |
| Image pairs | Prepare a positive and a negative version of the same image; the direction is the difference between them, so other things are less likely to change |
| Item | Description |
|---|---|
| Neutral (what gets changed) | The subject being adjusted, one prompt per box, up to 4, for example 1girl, solo. With only one short prompt, a slider usually works only on short prompts; add prompts you normally generate with (longer, different scenes) so it still has an effect in real use |
| Positive (+): words to add | Words added when the weight is positive, for example smile, happy; appended to each neutral prompt |
| Negative (−): words to add | Words added when the weight is negative |
| Anchor (optional) | Things the slider shouldn’t affect, for example background, scenery. Training is a little slower when filled in |
| Base images | The images used for training. By default they’re drawn from the neutral prompts; choose a dataset when you need a specific style or character |
- “Quality tags for base images and samples” and “Avoid NSFW” are on by default. Neutral prompts are usually short; without quality tags the base images tend to be grey and blurry, and very short prompts often lead anime models to draw nudity.
- Write only the direction you want to change, and leave out anything that shouldn’t change with it (for example hair colour). Some tags bring other things along, so choose words with a simple meaning.
Measured on an RTX 5080, 400 steps take about 11 minutes for SDXL and about 17 minutes for Anima.
Image pairs
Section titled “Image pairs”| Item | Description |
|---|---|
| Positive images, Negative images | Images with the same file name in the two folders form a pair (extensions can differ). Each pair should differ only in what you want to adjust; the more the composition, character and style match, the better |
| Prompts | The .txt next to an image is the prompt for that pair; pairs without one use “Prompt when there is no .txt” |
| Steps | 200 by default |
Making pairs with Qwen edit
Section titled “Making pairs with Qwen edit”If you don’t have pairs yet, open “Make pairs with Qwen” and let CoCo make them. A Qwen Image 2.1 model is required.
| Item | Description |
|---|---|
| Image source | “Draw base images” (recommended): base images are drawn with the base model chosen above and then edited, so the pairs share the base model’s own style and training is less likely to pick up another style. “Images in a folder”: edits existing images; if they all share one style, high weights will bring that style out |
| Prompts for base images | One per line, drawn in turn. The more varied, the better |
| Method | “Edit both” is best: small changes caused by editing appear on both sides and cancel out in training. “Original as negative” or “Original as positive” edits only half, twice as fast, but editing artefacts may be learned too |
| Positive edit instruction, Negative edit instruction | For example Change the lighting to dramatic lighting with strong contrast and deep shadows. Keep everything else the same. |
Press “Make pairs”; each image takes about 25 seconds, and you can’t generate meanwhile. When finished, the positive and negative folders above are filled in automatically. Finished images are skipped, so pressing it again after an interruption continues where it stopped.
Samples and choosing a checkpoint
Section titled “Samples and choosing a checkpoint”Every save (every 100 steps) draws a row of −2, −1, 0, +1, +2 for each sample prompt and puts them in a comparison set, so you can check the direction is right.
- Sliders often change other things as well, so weights of around ±1 are recommended.
- If the side effects are noticeable, choose an earlier checkpoint.