Elastic-forcing-14B

Non-EMA generator from step 80 of 20260912T074000Z_krea_official_wan21_480p8k_keepall, the checkpoint used in the existing 14B evaluations.

Only generator model weights are included: 1095 tensors in original FP32 precision. Every tensor was verified exactly equal to the source checkpoint. No optimizer, RNG, FD statistics, training counters, EMA weights, VAE, text encoder or loss encoders are included.

The generator weights are split into standard safetensors shards to meet the upload endpoint's per-file limit. This does not change tensor values or precision. This is not a standalone Diffusers pipeline; use the matching Elastic-Forcing/Self-Forcing generator runtime.

Load the generator

import json
from pathlib import Path
from huggingface_hub import snapshot_download
from safetensors.torch import load_file

folder = Path(snapshot_download('liuyueyi-8/Elastic-forcing-14B'))
index = json.loads((folder / 'model.safetensors.index.json').read_text())
state = {}
for shard in sorted(set(index['weight_map'].values())):
    state.update(load_file(str(folder / shard), device='cpu'))
generator.load_state_dict(state, strict=True)

# Optional: recreate the generator-only .pt format for an existing CLI:
# import torch
# torch.save({'generator': state}, 'model.pt')