How to Train Signed MiniMax H3 Sliders: The Multiplier-Matched Method

Most video "slider" LoRAs fail in two ways: they either bake a single static look regardless of weight, or they completely ignore negative strengths (-1.0). When building camera and motion controls for MiniMax H3, a true slider must act as a signed, seed-consistent dial—where +1.0 delivers one exact unit of the attribute (e.g., +20 degrees clockwise camera roll), -1.0 delivers the exact opposite (-20 degrees counter-clockwise), and 0.5 gives half.

In this technical breakdown, I share the exact Multiplier-Matched Training workflow behind my MiniMax H3 tools on Civitai.

---

Featured MiniMax H3 Models by Cosmic_Crafter

Before diving into the dataset math, you can grab the production-ready MiniMax H3 LoRAs directly here and add them to your workflow:

1. H3 Dutch Angle Slider (Model ID: 2952026) — Signed camera roll / Dutch angle dial calibrated so strength 1.0 = 20 degrees tilt across both positive and negative weights.
2. Painterly H3 by Cosmic (Model ID: 2903606) — Temporal painterly brushwork and illustrative motion styling engineered specifically for MiniMax H3 video generation.
3. Cosmic MiniMax H3 Cinema Suite Collection — Bookmark the full collection for upcoming H3 camera moves (roll-in, zoom, and tonal sliders).

---

Why Multiplier-Matched Training Works

In ai-toolkit, each dataset folder can define a network_weight scalar that is applied as the LoRA multiplier during that folder's forward pass:

```python
network.multiplier = batch.get_network_weight_list()
```

Instead of relying on text trigger words (which fight the base prompt), we:
1. Generate 24-40 Clean Anchor Scenes at 1344x768 (giving rotation/geometry plenty of headroom without edge clipping).
2. Grade Each Clip Synthetically Across N Signed Levels: For example, -40, -30, -20, -10, -5, +5, +10, +20, +30, +40 degrees, with unit = 20.
3. Use Identical Captions Across All Folders: The captions describe the scene geometry and subject, never the attribute itself.
4. Set network_weight = level / unit: Because the scalar multiplier is the only variable separating the folders, gradient descent forces the LoRA matrices (lora_A / lora_B) to encode a pure directional vector whose magnitude scales linearly with weight.

Note: Always skip level 0 in your folders. A multiplier of 0 turns the LoRA off and produces zero gradient—the base H3 model already serves as your 0.0 anchor for free.

---

Measured Calibration Results (Step 1200)

When evaluated across held multi-seed test scenes before and after post-training lora_B scaling (x0.8 calibration factor):

| LoRA Strength | Measured Camera Angle (Mean) | Per-Clip Hold Stability (SD) |
| :--- | :--- | :--- |
| -1.5 | -29.0 deg | +/- 0.3 deg |
| -1.0 | -20.8 deg (calibrated) | +/- 0.3 deg |
| -0.5 | -9.0 deg | +/- 0.2 deg |
| +0.5 | +12.0 deg | +/- 0.2 deg |
| +1.0 | +20.8 deg (calibrated) | +/- 0.3 deg |
| +1.5 | +34.0 deg | +/- 0.3 deg |

---

Combining H3 Dutch Angle Slider + Painterly H3

For dramatic graphic-novel or neo-noir motion sequences in ComfyUI:
- Load H3 Dutch Angle Slider at +0.75 or -0.75 for a classic 15-degree tension tilt.
- Stack Painterly H3 by Cosmic at 0.85 to transform the tilted cinema shot into living oil-and-ink concept art.

Drop your renders in the gallery on either model page.