Nodes/ComfyUI-42lux-Hildegard-Refiner/Hildegard References Split
ComfyUI Node

Hildegard References Split

Split the upscaled image into tiles and build the three Hildegard reference latents in one pass: - tile_latent: VAE-encoded crop of each tile (with optional high-frequency reduction so the model has room to invent texture) - position_latent: VAE-encoded 3x3 position map (current tile in center, 8 neighbors around it, cyan corner brackets ring the center) - global_latent: VAE-encoded thumbnail of the whole upscaled image tile 0 = all tiles (lists), tile # = only that tile.

By 42lux·Created 3 months ago·Updated 2 months ago· 53
Hildegard References Split
  • image
  • dac_data
  • vae
  • TILE(S)
  • POSITION(S)
  • tile_latent
  • position_latent
  • global_latent
tile0
cell_size512
ctrl3_max_size2048
tile_high_freq_reduce0.00
low_freq_radius256
Categoryupscale/hildegard-refiner

Inputs (8)

NameTypeDefaultDescription
imageIMAGEUpscaled IMAGE from Hildegard Plan.
dac_dataHILDEGARD_DATAdac_data from Hildegard Plan. Carries the tile grid, overlaps, and ordering this node has to honor.
vaeVAEVAE used to encode the three reference image types into latents. For FLUX.2 Klein, the Klein VAE.
tileINT00 = process all tiles in the grid (each output is a list of N entries). N >= 1 = process only that single tile (each output is a list of 1). Index follows the same linear/spiral ordering Hildegard Plan picked.
cell_sizeINT51264–2048Side length (px) of each cell in the 3x3 position map. The final position image is cell_size x 3 on each axis. Larger = more spatial context detail, larger position latent. 512 is the Hildegard training default.
ctrl3_max_sizeINT2048256–8192Long-edge cap (px) for the global thumbnail before VAE encoding. The whole upscaled image is downscaled so its longer edge fits this. 2048 is the Hildegard training default.
tile_high_freq_reduceFLOAT0.000–1Attenuate the tile reference's high-frequency band (micro-detail / edges / texture) before VAE-encoding. 0 = full source detail preserved in the reference (tight fidelity, model has little room to add). 1 = soft, low-contrast version with only broad shapes and color preserved. Only affects the tile_latent output. The TILE(S) image output, position_latent, and global_latent are untouched.
low_freq_radiusFLOAT2561–256Gaussian-blur radius (px) defining the cutoff between low and high frequencies for tile_high_freq_reduce. Larger = more of the source's detail is classified as 'low frequency' and survives the reduce. Smaller = only the very-broadest shapes survive; the model regenerates everything else.

Outputs (5)

NameTypeDescription
TILE(S)IMAGE
POSITION(S)IMAGE
tile_latentLATENT
position_latentLATENT
global_latentLATENT