Modelpublic

Owen777/z-image-cfg-distilled

sourceHugging Faceupdated 4mo agoView on Hugging Face

2likes

Model Card

Z-Image CFG-Distilled EMA Checkpoint

记录日期：2026-06-04

HF repo：

text

Owen777/z-image-cfg-distilled

Checkpoint

这是 Z-Image CFG distillation run 的 step 3000 EMA transformer 权重。

本地源 checkpoint：

text

runs/zimage_cfg_distill_tcfg2p5_1024_1node_gbs16_20260603_234415/ckpts/step_00003000

融合后的 EMA init 目录：

text

runs/zimage_cfg_distill_tcfg2p5_1024_1node_gbs16_20260603_234415/ckpts/step_00003000_ema_init

HF 上传结构：

text

transformer/
  config.json
  diffusion_pytorch_model.safetensors
metadata/
  cfg_distillation_config.yaml
  training_config.yaml
  ema_meta.json
README.md

transformer/diffusion_pytorch_model.safetensors 是由 ema_rank_00.pt 到 ema_rank_07.pt 融合得到的 bf16 safetensors。

融合验证：

text

stage: zero2
world_size: 8
n_params: 521
EMA decay: 0.995
EMA optimization_step: 3000
bf16 tensor size: about 11.46 GiB
output file size: 12,309,880,368 bytes

Training Source

源 run：

text

zimage_cfg_distill_tcfg2p5_1024_1node_gbs16_20260603_234415

核心参数：

text

trainer: zimage_meanflow_zero2.train_zero2_cfg_distill
method_type: ZImageCFGDistillation
base model: Tongyi-MAI/Z-Image
resolution: 1024
global batch: 16
max_train_steps: 3000
teacher_cfg_scale: 2.5
flow_shift: 6.0
sigma_sampling: shifted_uniform
loss_reduction: mse
uncond_loss_weight: 0.0
student_aux_time_mode: none
optimizer: AdamW
learning_rate: 3.0e-5
mixed_precision: bf16
EMA: enabled, decay 0.995

The exported config identifies the transformer class as:

text

ZImageTransformer2DModel

This is a CFG-distilled Z-Image transformer checkpoint, not a full pipeline directory. Use it together with the original Z-Image pipeline components.

MeanFlow Follow-Up Run

This EMA checkpoint is also used as the initialization for a new 10K-step MeanFlow run with teacher CFG reset to 1.0:

text

Slurm array: 3149472
actual first task job id: 3149495
nodes: 4 x GB200 nodes, 16 GPUs total
output:
runs/zimage_meanflow_hsg_b200_4node_fa4varlen_1024_bs4_10k_from_cfgdistill3000ema_tcfg1_array_3149472
config:
zimage_meanflow_zero2/configs/ablations/zimage_fd0p005_1024_10k_gapend0p3_fm0p25_teachercfg1p0_bs4_cfgdistill3000ema.yaml

The MeanFlow run keeps the previous 1024/fa4-varlen/bs4/global64/10K setup and uses teacher_cfg_scale: 1.0.