SD1-KOHYA_LYCORIS_LORA Components Params

SD1-KOHYA_LYCORIS_LORA Components Params (class SD1-KOHYA_LYCORIS_LORA Defaults Hyper) is the per-component weight node for LyCORIS networks on SD1 - the kohya-layout config for LoKr and other LyCORIS variants (sd1-kohya_lycoris_lora). Like the other components params nodes it exposes clip_l, diffuser, and vae floats, and emits a single param recipe that applies each weight to its own part of the network.

If the kohya-LoRA node is for the standard LoRA case, this is its sibling for the newer, denser family of low-rank networks. LyCORIS - and LoKr in particular - has become the default for character training in parts of the community because it packs more representational power per parameter than a plain LoRA, often converging better at low rank. Since mecha can merge and compose LyCORIS networks, it also gives you component-level control over them, which is exactly what you want when a character's "look" lives mostly in the UNet while a style's tagging lives in the text encoder.

How it works

LyCORIS networks have their own key extraction logic in sd-mecha (each algorithm - Lokr, LoHa, IA3, and friends - defines how to extract weights from its factorization). The sd1-kohya_lycoris_lora config is an auxiliary config layered on the sd1-kohya base, so its components mirror the base's: clip_l, diffuser, vae. This node maps your three floats onto those components and ORs them into one param recipe, exactly like its LoRA sibling - the difference is purely which key layout the config describes.

The inputs that matter:

• clip_l - weight for the network's text-encoder tensors.
• diffuser - weight for the UNet tensors.
• vae - weight for VAE-side tensors (usually absent in practice).
• recipe (output, param space) - the combined param for sd1-kohya_lycoris_lora.

Wire it into the alpha/weight input of an Add Difference or Weighted Sum whose inputs are base + a LyCORIS network, and you get per-part strength: full UNet, quarter text encoder, VAE untouched.

Install

Part of the Mecha Merge Node Pack (ljleb/comfy-mecha):

• ComfyUI Manager → Install Custom Nodes → search "mecha" → Mecha Merge Node Pack, or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/ljleb/comfy-mecha.git
pip install -r comfy-mecha/requirements.txt

Restart ComfyUI. Only dependency is sd-mecha==1.1.7; no model downloads.

Common issues

• The config must match what was loaded. If your file is a plain kohya LoRA, you want the SD1-KOHYA_KOHYA_LORA node; if it's a LyCORIS/LoKr network, this one. Plugging the wrong config's params into the wrong merge space is the main way people get confusing "component does not exist" errors - check which config Mecha Lora Recipe reports.
• LyCORIS extraction is key-count heavy. Merging a LoKr network means reconstructing tensors from factorizations, so don't be surprised if it's slower per run than a plain LoRA merge. That's the cost of the denser representation, not a bug.
• 0 still means "no override." Per-component params scale the method's treatment; they don't exclude parts. Use Omit Component for hard exclusion.

The three SD1 param nodes - LDM, KOHYA, KOHYA_LORA, LYCORIS_LORA - are the same tool in four costumes. The skill is picking the costume that matches your recipe's config, and this one is the costume for SD1 LoKr and friends.