docs/training.md

# Training and distillation

[← Meridian](../README.md) · [Inference method](method.md)

Training provenance and design are collected here rather than in the model card or research blog.

## Training overview

The following is the release's training description; this repository contains inference code and
artifacts, **not the training or distillation pipeline**.

The teacher was trained on
[MultiCamVideo](https://huggingface.co/datasets/KwaiVGI/MultiCamVideo-Dataset): 13,600 Unreal Engine
scenes, each filmed by ten synchronized cameras over 81 frames. A sample pairs one camera's clip
with a point-cloud render from a second camera; the second camera's actual clip is the target.
Both directions of camera pairs are used, and geometry is reconstructed from the source clip alone.

Later stages added still-frame references, temporally extended examples made by slowing, reversing,
or holding the 81-frame window, and eased sweeps. The student was distilled at 73, 90, and 124 frames,
with 243-frame holds also reported. Training-time temporal augmentation is not a promise that every
time mapping or reverse-playback path is supported by the released interfaces.

## Distillation and checkpoint design

The release has two learned components:

| Component | Role |
|---|---|
| `teacher_lora/` | Rank-128 re-camera adapter on the **unmodified** MiniMax-H3 transformer, reported size 2.5 GiB. Earlier releases shipped this teacher as a fused 61.7 GiB `transformer/`, now kept under `legacy/`. |
| `turbo_lora/` | Rank-128 DMD student adapter distilled against that teacher, reported size 2.5 GiB. The two are loaded together at weight 1.0; neither is merged into the base weights. |

| Mode | CLI settings | Transformer evaluations |
|---|---|---|
| Student, default | `--steps 4 --flow-shift 3` with both adapters loaded | 3 |
| Teacher | `--lora teacher_lora --steps 50 --flow-shift 12` | 49 |

The H3 scheduler counts the terminal zero-noise point in `--steps`. That endpoint does not require a
model evaluation, hence four grid points produce three forwards. Reducing the teacher's step count
is not equivalent to using the distilled student. Forward counts also do not directly translate to
end-to-end speedups: reconstruction, warping, VAE work, and file writing still take time.

The source and point-cloud-rendered references share the selected source timeline. Synchronized
multi-camera supervision is followed by temporal and camera-path augmentation and student distillation.
The training footage is synthetic; real-world performance depends on the scene.

For weight provenance and modification notices, see [`MODIFICATIONS.md`](../MODIFICATIONS.md).