Mean Mode Screaming: Mean--Variance Split Residuals for 1000-Layer Diffusion Transformers
Paper • 2605.06169 • Published • 231
import torch
from diffusers import DiffusionPipeline
# switch to "mps" for apple devices
pipe = DiffusionPipeline.from_pretrained("BiliSakura/MVSplit-DiT-diffusers", dtype=torch.bfloat16, device_map="cuda")
prompt = "a red panda climbing a bamboo stalk"
image = pipe(prompt).images[0]BiliSakura/MVSplit-DiT-diffusers
Diffusers-ready checkpoints for MVSplit-DiT (Mean–Variance Split Residual Diffusion Transformers), converted for local/offline use with a project-owned custom MVSplitDiTPipeline.
Re-distribution notice: weights are converted from
StableKirito/mvsplit-dit-1000l. Original work: Mean Mode Screaming. License: Apache 2.0.
Available checkpoints
| Subfolder | Params | Task | Resolution |
|---|---|---|---|
MVSplit-DiT-1000L/ |
1000L | text-to-image | 256×256 |
Each subfolder is a self-contained Diffusers model repo with pipeline.py, model_index.json, and component weights.
Demo
Prompt: a red panda climbing a bamboo stalk — 256×256, 35 steps, CFG 2.0.
Inference
cd MVSplit-DiT-1000L
python demo_inference.py
See MVSplit-DiT-1000L/README.md for full usage and recommended settings.
Citation
@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},
}
- Downloads last month
- 65