Hildegard References Split

This is the node that makes Hildegard Hildegard. References Split slices the upscaled image into tiles and, for each one, builds three reference latents - the tile itself, a 3×3 position map of its neighbors, and a thumbnail of the whole image. FLUX.2 Klein is natively a reference-editing model, and this pack repurposes that: instead of ControlNet Tile (which has gone quiet for the new-generation models - check the KB's controlnet essay if you've noticed the stall), it feeds Klein spatial context it already understands. The matching LoRA (42lux/hildegard) teaches the model to actually read those three slots.

Why this beats each-tile-in-isolation: the classic tiled upscale failure is tiles drifting into patchwork because each one is generated blind. Here every tile carries its own crop (fidelity), a position map showing where it sits among its eight neighbors (context), and a global thumbnail (color and composition anchor). Tiles stop hallucinating structure that isn't there, and seams stay invisible.

How it works

For each tile it builds, in one pass:

• tile_latent - VAE-encoded crop of the tile. Optionally frequency-reduced (below).
• position_latent - VAE-encoded 3×3 map: the current tile in the center cell, its 8 neighbors around it, with cyan corner brackets ringing the center so the model knows where the cell boundaries are. Off-canvas cells stay neutral gray.
• global_latent - VAE-encoded thumbnail of the whole upscaled image, long edge capped at ctrl3_max_size.

All four outputs are index-aligned: element i is the same physical tile across TILE(S), tile_latent, and position_latent. The author describes it as "a pure reference-based upscale that samples from an empty latent" - you run your sampler once per tile, feeding the latents through the ReferenceLatent nodes plus per-ref weight controls (Flux2KleinRefLatentWeight), then stitch with Combine.

The frequency knob - the part people argue about

tile_high_freq_reduce (0.0–1.0, default 0) subtracts part of the tile's high-frequency band before encoding, with low_freq_radius (default 256) as the Gaussian blur that defines the low/high cutoff. At 0.0 the reference keeps full source detail - tight fidelity, little room for the model to add. At 1.0 it's a soft, low-contrast shell - only broad shapes and color survive, and Klein invents fresh micro-detail. Only tile_latent is affected; structure and global context stay locked. Reach for it when the source is noisy or over-sharpened and you want the model to re-synthesize clean detail. This was the feature one reviewer called the key innovation of the whole pack.

Inputs and outputs that matter

image (the upscaled output of Hildegard Plan), dac_data (its grid geometry), and vae (for FLUX.2 Klein, the Klein VAE - the fl2vae in the example workflow). tile (default 0) processes all tiles; set it to a number to process just that one. cell_size (default 512) is the side of each position-map cell - 512 is the LoRA's training default, leave it. ctrl3_max_size (default 2048) caps the global thumbnail; the shipped workflow still uses 768, and the author says for modest 1k–4k upscales you can disconnect the thumbnail entirely.

Outputs: TILE(S) and POSITION(S) image lists, plus tile_latent, position_latent, and the single global_latent.

Install

Same pack as Plan - ComfyUI Manager (search "ComfyUI-42lux-Hildegard-Refiner") or:

cd ComfyUI/custom_nodes
git clone https://github.com/42lux/ComfyUI-42lux-Hildegard-Refiner

Restart. No pip deps. Then the actual requirement: download the Hildegard LoRA from huggingface.co/42lux/hildegard and load it (e.g. LoraLoaderModelOnly) before sampling. Skip it and nothing will happen - the model won't know what those three reference slots mean. Prompt with the RFNTILE. trigger phrase from the README or one of the LLM prompt templates in the pack's llm_prompt_templates/.

Gotchas

Keep the index order intact - if you process a subset, feed Combine the same tiles in the same order. If someone hits a tensor size mismatch on the reference latents, it's almost always a hand-tuned graph feeding a reference at the wrong size; wire this node's outputs straight into the ref slots. And expect this node to be the slow, VRAM-hungry step of the pipeline - it's doing real VAE encodes, per tile. A 12GB card will top out around a 2048-tile pass regardless of what you do downstream.