Slider LoRA Targeting¶

In this guide, we'll be training a slider-style adapter in SimpleTuner. We'll be using Z-Image Turbo because it trains quickly, ships an Apache 2.0 license, and gives great results for its size—even with distilled weights.

For the full compatibility matrix (LoRA, LyCORIS, full-rank), see the Sliders column in documentation/QUICKSTART.md; this guide applies to all architectures.

Slider targeting works with standard LoRA, LyCORIS (including full), and ControlNet. The toggle is available in both CLI and WebUI; everything ships in SimpleTuner, no extra installs needed.

Step 1 — Follow the base setup¶

• CLI: Walk through documentation/quickstart/ZIMAGE.md for environment, install, hardware notes, and the starter config.json.
• WebUI: Use documentation/webui/TUTORIAL.md to run the trainer wizard; pick Z-Image Turbo as usual.

Everything from those guides can be followed until you reach the point of configuring a dataset because sliders only change where adapters are placed and how data is sampled.

Step 2 — Enable slider targets¶

• CLI: add "slider_lora_target": true (or pass --slider_lora_target true).
• WebUI: Model → LoRA Config → Advanced → check “Use slider LoRA targets”.

For LyCORIS, keep lora_type: "lycoris" and for lycoris_config.json, use the presets in the details section below.

Step 3 — Build slider-friendly datasets¶

Concept sliders learn from a contrastive dataset of "opposites". Create small before/after pairs (4–6 pairs is enough to start, more if you have them):

• Positive bucket: “more of the concept” (e.g., brighter eyes, stronger smile, extra sand). Set "slider_strength": 0.5 (any positive value).
• Negative bucket: “less of the concept” (e.g., dimmer eyes, neutral expression). Set "slider_strength": -0.5 (any negative value).
• Neutral bucket (optional): regular examples. Omit slider_strength or set it to 0.

It's not necessary to keep filenames matched across positive/negative folders - just ensure you have an equal number of samples in each bucket.

Step 4 — Point the dataloader at your buckets¶

• Use the same dataloader JSON pattern from the Z-Image quickstart.
• Add slider_strength to each backend entry. SimpleTuner will:
• Rotate batches positive → negative → neutral so both directions stay fresh.
• Still honor each backend’s probability, so your weighting knobs keep working.

You don’t need extra flags—just the slider_strength fields.

Step 5 — Train¶

Use the usual command (simpletuner train ...) or start from the WebUI. Slider targeting is automatic once the flag is on.

Step 6 — Validate (optional slider tweaks)¶

Prompt libraries can carry per-prompt adapter scales for A/B checks:

{
  "plain": "regular prompt",
  "slider_plus": { "prompt": "same prompt", "adapter_strength": 1.2 },
  "slider_minus": { "prompt": "same prompt", "adapter_strength": 0.5 }
}

If omitted, validation uses your global strength.

References & details¶

[
  "attn1.to_q", "attn1.to_k", "attn1.to_v", "attn1.to_out.0",
  "attn1.to_qkv", "to_qkv",
  "proj_in", "proj_out",
  "conv_in", "conv_out",
  "time_embedding.linear_1", "time_embedding.linear_2"
]

["to_q", "to_k", "to_v", "to_out.0", "to_qkv"]

["attn1.to_query", "attn1.to_key", "attn1.to_value", "conv_in", "conv_out", "time_embedding.linear_1", "time_embedding.linear_2"]

{
  "algo": "lokr",
  "multiplier": 1.0,
  "linear_dim": 4,
  "linear_alpha": 1,
  "apply_preset": {
    "target_module": [
      "attn1.to_q",
      "attn1.to_k",
      "attn1.to_v",
      "attn1.to_out.0",
      "conv_in",
      "conv_out",
      "time_embedding.linear_1",
      "time_embedding.linear_2"
    ]
  }
}