Motif-Technologies

/

Motif-VAE

like 12

Follow

Motif Technologies 395

What

Apply the modeling refactor merged in
MotifTechnologies/motif-vae-public#1:
clearer, Motif-specific block/function/output-container names, plus removal of
leftovers the inference checkpoint never uses (tiling / temporal-chunk /
quant-layer flags, the get_last_layer GAN helper, the 0.18215 latent-scale
fallback).

Safety

modeling_motifvae.py is byte-identical to the GitHub PR head (d6a045f).
No architecture or weight change: the published checkpoint loads with no
missing/unexpected keys via diffusers.AutoModel.from_pretrained, and a full
forward pass (encode -> decode) on the real weights succeeds.

Test

from diffusers import AutoModel
import torch

m = AutoModel.from_pretrained("Motif-Technologies/Motif-VAE", trust_remote_code=True)
out = m(torch.randn(1, 3, 1, 64, 64))  # ForwardOutput, sample shape (1, 3, 1, 64, 64)