ComfyUI Krea2 Merge

ComfyUI Krea2 Merge

<p align="center"> <img src="assets/icon.png" alt="ComfyUI Krea2 Merge icon" width="200"> </p>

ComfyUI Krea2 Merge is a standalone set of nodes for exactly combining two to four Krea 2 LoRAs without loading or modifying the base model. It supports the PEFT/Diffusers lora_A + lora_B keys used by Krea 2 LoRAs, while retaining support for Kohya lora_down + lora_up checkpoints.

This fork uses its own package name, extension name, node IDs, category, and output folder. It can therefore be installed alongside the original comfyui-merge extension without replacing its nodes.

Features

• Merge two to four Krea 2 LoRAs without loading a base model.
• Support PEFT/Diffusers lora_A + lora_B state dictionaries.
• Retain Kohya lora_down + lora_up compatibility.
• Infer missing alpha values from the LoRA rank.
• Handle negative merge weights correctly for both lora_B and lora_up.
• Exactly merge LoRAs with equal or different ranks without unwanted cross terms.
• Optionally cap the output rank with svd_truncate, an optimal SVD truncation that reports its approximation error.
• Enter positive or negative weights from -4.0 to 4.0.
• Report unsupported files and incompatible tensor shapes clearly.
• Load .safetensors, .sft, and supported PyTorch checkpoints through ComfyUI's safe loader.
• Save only inside ComfyUI/models/loras/krea2-merged-loras, including subfolders.
• Keep alpha in float32 even when the factor tensors use FP16 or BF16.
• Keep merge results independent of input order for equal weights.

Installation

1. Download or clone this repository into:

   ComfyUI/custom_nodes/comfyui-krea2-merge
   

2. Restart ComfyUI.

The extension requires PyTorch (provided by ComfyUI). The safetensors package is required for loading and saving .safetensors files and is included in normal ComfyUI installations.

Nodes

All nodes appear under Krea2 Merge / LoRA:

• Krea2 Merge • Load LoRA — loads a LoRA as a state dictionary.
• Krea2 Merge • Merge LoRAs — combines two to four LoRAs.
• Krea2 Merge • Save LoRA — saves the merged state dictionary.
• Krea2 Merge • Apply LoRA — loads and applies a LoRA to a connected model.

The internal node IDs are prefixed with Krea2Merge_, so workflows do not collide with the original extension.

Basic workflow

1. Add two Krea2 Merge • Load LoRA nodes and select your Krea 2 LoRAs.
2. Connect them to Krea2 Merge • Merge LoRAs.
3. Set weight1 and weight2. Optional third and fourth inputs use weight3 and weight4.
4. Keep the default exact_concat for a mathematically exact weighted merge. Choose svd_truncate instead when you want a smaller file: it limits every module to target_rank (see the compatibility notes). legacy_linear remains available only for compatibility with the original approximate factor-space behavior.
5. Select the output precision. fp16 is the practical default; use bf16 if that matches your Krea 2 setup.
6. Connect Krea2 Merge • Save LoRA and choose a filename ending in .safetensors or .sft, optionally inside a subfolder such as characters/my_merge.safetensors.
7. Set allow_overwrite to yes when repeated runs should replace the same output file. Keep it on no to protect an existing merge.

Since version 1.2.1, output paths must stay inside krea2-merged-loras. Absolute paths, .., and symlinks leading outside that folder are rejected. If an older workflow uses an absolute output path, replace it with a filename or relative subfolder. Supported output extensions are .safetensors, .sft, .pt, .pth, .ckpt, and .bin.

On Windows, overwrite first moves the previous output to a temporary backup. If the new save fails, the original file is restored automatically. If Windows has the destination open or memory-mapped, use another filename or restart ComfyUI.

An example is included at workflow-examples/krea2-lora-merger.json.

The example includes Show Text from ComfyUI-Custom-Scripts to display the saved path. The Save node is also registered as an output node, so saving still works when Show Text is removed from a custom workflow.

Compatibility notes

• exact_concat is the default. It concatenates complete A/B or down/up pairs, avoiding the cross terms produced when factors are added separately. It also supports different ranks: a rank-4 plus rank-16 module becomes rank 20.
• Concatenation increases the output rank. Combining two rank-32 LoRAs produces rank 64; four produce rank 128. This increases file size and the temporary work needed while ComfyUI applies the LoRA.
• svd_truncate first composes the LoRAs exactly like exact_concat, then limits each module to at most target_rank (default 32, only used by this mode). It keeps the best possible rank-target_rank approximation of the merged weight delta (an optimal SVD truncation), so the output rank no longer grows with the number of inputs. Modules whose combined rank is already within target_rank stay exact. The result is deterministic, and the console reports the relative error of the truncated modules. The error is the share of the merged delta that lies beyond target_rank: it is small for LoRAs with a decaying spectrum and large for LoRAs whose rank is fully used, so pick a target_rank close to the summed input ranks when fidelity matters. Only linear and 1x1-up (LoCon-style) modules are supported; use exact_concat otherwise.
• legacy_linear preserves old saved workflows and factor-space results. It is approximate, introduces cross terms, and requires matching shapes and ranks.
• exact_concat and svd_truncate apply weights directly and ignore force_same_strength. For an even blend, start with weight1 = 0.5 and weight2 = 0.5.
• All modes operate only on the LoRA files; none changes or saves the Krea 2 base model.
• When a PEFT checkpoint has no embedded alpha tensors, the extension uses alpha = rank (unit scaling). This is a fallback assumption: the original training alpha cannot be recovered from the factors alone. Files that rely on a separate non-default training alpha need that scaling supplied first.
• Alpha is always saved in float32. save_dtype controls the factor tensors; it does not reduce alpha precision.
• DoRA, LoCon lora_mid, LoHa/LoKr, diff/diff_b, factor biases, and other unrecognized adapter keys are rejected in both modes instead of being silently omitted or merged incorrectly.
• Inputs should use the same base model and module naming convention. A warning appears when two inputs have no shared module names; separate target layers can be valid, but differently named aliases for the same layer are not converted or combined. Matching shapes alone do not establish compatibility.
• A missing LoRA filename in a saved workflow is retained so ComfyUI can report the missing file. Select its replacement explicitly in the Load or Apply node.
• A connected third or fourth LoRA with a zero weight is ignored with a clear console warning.

Attribution

This project is derived from LingSss9/comfyui-merge and retains its MIT license. Krea 2 support, separate node IDs, validation, tests, and the release branding were added in this fork.

Krea 2 and its model names belong to their respective owners. This community extension is not an official Krea product.