# 6️⃣ Final embedding embed = self.embed_head(pooled) # (B, out_dim) embed = nn.functional.normalize(embed, p=2, dim=-1) # L2‑norm

This is the most commonly cited spec: 1080p (1920x1080 pixels) and 4K (3840x2160 pixels). While a higher resolution is critical for larger screens, a high bitrate is what ensures that the image at that resolution is clean and artifact-free.

def forward(self, x): # x : (B, 3, H, W) (float, 0‑1) feats = self.backbone(x) # (B, C, h, w) feats = self.pool(feats).squeeze(-1).squeeze(-1) # (B, C) return feats # (B, 1792)

: