fix app.py

app.py

@@ -60,6 +60,7 @@ if "latent_mean" in config.model:
     latent_std = torch.Tensor(config.model.latent_std)[None, None, :].to(device)
     assert latent_mean.shape[-1] == config.model.generator.in_channels
     perchannel_norm = True
+latent_nf = config.model.latent_nf
 
 config.diffusion.pop("timestep_respacing")
 config.model.pop("vae")
@@ -114,14 +115,14 @@ def process(input_image, input_num_steps=25, input_seed=42, input_cfg=6.0):
             final_samples = samples
         recon_param = final_samples["sample"].reshape(inf_bs, config.model.num_prims, -1)
         if perchannel_norm:
-            recon_param = recon_param / config.model.latent_nf * latent_std + latent_mean
+            recon_param = recon_param / latent_nf * latent_std + latent_mean
         recon_srt_param = recon_param[:, :, 0:4]
         recon_feat_param = recon_param[:, :, 4:] # [8, 2048, 64]
         recon_feat_param_list = []
         # one-by-one to avoid oom
         for inf_bidx in range(inf_bs):
             if not perchannel_norm:
-                decoded = vae.decode(recon_feat_param[inf_bidx, ...].reshape(1*config.model.num_prims, *latent.shape[-4:]) / config.model.latent_nf)
+                decoded = vae.decode(recon_feat_param[inf_bidx, ...].reshape(1*config.model.num_prims, *latent.shape[-4:]) / latent_nf)
             else:
                 decoded = vae.decode(recon_feat_param[inf_bidx, ...].reshape(1*config.model.num_prims, *latent.shape[-4:]))
             recon_feat_param_list.append(decoded.detach())