{
  "claim": "GAE-D64-1B is a 0.93B flow transformer on top of 2.56B parameters that are frozen while it trains, of which 2.11B come from two models GAE never trains at all. The learned geometry decoder has zero weights: geometry is read by Depth Anything 3's own DPT head. The learned RGB head has 203.2M, which is 45% of every trained weight in the codec.",
  "method": "Every shipped .safetensors header read directly over HTTP with a Range request (first 8 bytes for the header length, then the JSON header), and the tensor shapes summed. No download, no load. Sizes are the Hub's reported blob sizes. The GAE codec and flow numbers are from TencentARC/GAE-D64-1B; the backbone from depth-anything/DA3-GIANT-1.1; the text encoder from Qwen/Qwen3-0.6B (tied embeddings, counted once as the file stores them).",
  "source": "https://huggingface.co/TencentARC/GAE-D64-1B",
  "captured": "2026-09-22",
  "note": "repa_proj is the projector for L_tok, which the paper says is discarded after codec training; it is still in the shipped checkpoint, and gae/pipeline.py filters it out of the missing-keys check when loading. Every GAE tensor in the repo is F32, as is DA3-GIANT-1.1.",
  "columns": [
    {
      "key": "module",
      "label": "module",
      "align": "left",
      "mono": true
    },
    {
      "key": "role",
      "label": "role",
      "align": "left"
    },
    {
      "key": "params",
      "label": "parameters",
      "align": "right",
      "mono": true
    },
    {
      "key": "trained",
      "label": "trained by GAE",
      "align": "left"
    }
  ],
  "rows": [
    {
      "module": "DA3-GIANT-1.1",
      "role": "encoder + DPT geometry head",
      "params": "1,355,674,125",
      "trained": "no \u2014 frozen at both ends"
    },
    {
      "module": "codec.enc_*",
      "role": "fused 12288-ch tensor to z",
      "params": "126,132,352",
      "trained": "yes"
    },
    {
      "module": "codec.dec_proj/dec_attn",
      "role": "shared decoder trunk",
      "params": "50,453,504",
      "trained": "yes"
    },
    {
      "module": "codec.dec_conv",
      "role": "trunk to 4 x 3072 features",
      "params": "73,509,888",
      "trained": "yes"
    },
    {
      "module": "codec.rgb_head",
      "role": "trunk to pixels",
      "params": "203,215,436",
      "trained": "yes"
    },
    {
      "module": "codec.repa_proj",
      "role": "C-RADIO projector, training-only",
      "params": "1,231,104",
      "trained": "yes, then unused"
    },
    {
      "module": "codec (other)",
      "role": "level statistics + level weights",
      "params": "24,580",
      "trained": "no \u2014 buffers, plus 4 level weights"
    },
    {
      "module": "flow transformer",
      "role": "GAEFlowTemporal, 28x768 + 6x2048",
      "params": "933,675,138",
      "trained": "yes"
    },
    {
      "module": "Qwen3-0.6B",
      "role": "text encoder, cross-attention",
      "params": "751,632,384",
      "trained": "no \u2014 frozen"
    },
    {
      "module": "total at inference",
      "role": "one frame end to end",
      "params": "3,495,548,511",
      "trained": "1.39B of it"
    }
  ]
}