Elastic-forcing-1.3B-three-encoder

Non-EMA generator from experiment 10351, step 150. Trained using DINOv3, V-JEPA2 and VideoMAE distribution matching; these loss encoders are not needed for inference.

model.pt contains only {'generator': state_dict}: 825 tensors in the original FP32 precision. Tensor values were checked against the source checkpoint. No optimizer, RNG, FD statistics, training counters, EMA copy, feature encoders, VAE or text encoder are included. This is a generator-only checkpoint for the corresponding Elastic-Forcing/Self-Forcing runtime, not a standalone Diffusers pipeline.

Load with the existing inference entry point and select generator (do not select EMA). To inspect directly:

import torch
weights = torch.load('model.pt', map_location='cpu', weights_only=True, mmap=True)
generator.load_state_dict(weights['generator'], strict=True)

Checkpoint step: 150. SHA256 of model.pt: d4c7dc95f84f200b669af7dfac0c2cdd9ed6bfc1435b48cbc3f9a040aba4a454.