Skip to content

vae-merge.md

Latest commit

 

History

History
66 lines (47 loc) · 2.22 KB

vae-merge.md

File metadata and controls

66 lines (47 loc) · 2.22 KB
Raw

VAE Tools (MEC)

Four VAE-focused nodes covering merging, latent diagnostics, and model introspection.

VAEMergeMEC

Merge two (or three) VAEs using one of eight algorithms:

  • weighted_sum — out = (1-α)·A + α·B
  • add_difference — out = A + α·(B - C) (requires vae_c)
  • tensor_sum — element-wise mean of A and B
  • triple_sum — equal mean of A, B, C (requires vae_c)
  • slerp — spherical linear interpolation of flattened parameters
  • dare_ties — DARE/TIES with sparsity drop and sign-resolution
  • block_swap — replace whole blocks of A with B according to per-block weights
  • clamp_interp — like weighted_sum but bounds the result to the per-tensor range of A and B

Per-block alpha

Pass either a JSON object or a comma-separated list to override the global alpha for individual blocks. Recognised names follow the SD/SDXL VAE block layout (block_conv_in, block_0..3, block_mid, block_norm_out, block_conv_out).

Brightness / contrast

After the merge, two scalar tweaks can be applied to the decoder.conv_out weights only — useful for nudging output luminance without retraining.

The merged VAE is returned as a fresh deepcopy; the inputs are never modified. The merge runs on CPU in float32 and is cast back to the source dtype before being installed into the wrapper.

VAELatentInspectorMEC

Per-channel min / max / mean / std / abs_mean stats across the latent, plus NaN/Inf counts and an overall verdict:

  • corrupt — non-zero NaN or Inf count
  • saturated — channel abs_mean exceeds 30 (suggests clipping)
  • low_contrast — channel std below 0.05
  • healthy — none of the above

Use fail_on_corrupt=True to raise hard on corrupt latents (useful as a sentinel in long batched runs).

VAESimilarityAnalyserMEC

Computes cosine similarity between two VAEs:

  • Globally (over all common parameters)
  • Per block (using the same SD/SDXL layout as VAEMerge)
  • Optionally per tensor

Reports tensors that exist only in one model — handy for debugging mismatched checkpoints.

VAEBlockInspectorMEC

Per-block weight statistics (mean / std / abs_mean / count) for a single VAE. Useful to spot blocks dominated by NaN/zeros, or to compare fine-tunes against a reference.