File size: 2,234 Bytes
3f9453c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
---
license: apache-2.0
pipeline_tag: text-to-image
---

# MVSplit-DiT (1000 layers)

This repository contains the weights for the 1000-layer Diffusion Transformer (DiT) presented in the paper [Mean Mode Screaming: Mean--Variance Split Residuals for 1000-Layer Diffusion Transformers](https://huggingface.co/papers/2605.06169).

[Project Page](https://erwold.github.io/mv-split/) | [GitHub Repository](https://github.com/erwold/mv-split)

## Introduction

Scaling Diffusion Transformers to extreme depths (hundreds or thousands of layers) introduces a structural vulnerability known as **Mean Mode Screaming (MMS)**. In this state, token representations homogenize, and centered variation is suppressed, leading to model collapse. 

MVSplit-DiT addresses this by using **Mean-Variance Split (MV-Split) Residuals**, which combine a separately gained centered residual update with a leaky trunk-mean replacement. This architecture enables the stable training of DiTs at boundary scales, such as the 1000-layer model provided here.

## Usage

To use this model for image generation, please refer to the official [GitHub repository](https://github.com/erwold/mv-split) for installation instructions and requirements.

### Sampling

You can generate images using the `sample.py` script. The model requires the DiT checkpoint from this repo, a FLUX.2 VAE, and a Qwen3 text encoder.

```bash
# Custom prompt example
python sample.py \
    --checkpoint_path /path/to/model.pt \
    --flux_vae_path   /path/to/flux2_ae.safetensors \
    --qwen_model_path Qwen/Qwen3-0.6B \
    --prompt "a red panda climbing a bamboo stalk" \
    --output_dir ./samples
```

### Key sampling flags

| Flag | Default | Meaning |
|---|---|---|
| `--image_size` | 256 | Square output side in pixels. |
| `--num_inference_steps` | 35 | Euler steps for the flow-matching ODE. |
| `--cfg_scale` | 2.0 | Classifier-free guidance. |
| `--time_shift_alpha` | 4.0 | Time-shift in the flow schedule (must match training). |

## Citation

```bibtex
@article{lu2026mms,
  title   = {Mean Mode Screaming: Mean--Variance Split Residuals for 1000-Layer Diffusion Transformers},
  author  = {Lu, Pengqi},
  journal = {arXiv preprint arXiv:2605.06169},
  year    = {2026},
}
```