ComfyUI_LC_ModelBuilder_nodes Introduction
ComfyUI_LC_ModelBuilder_nodes is a collection of ComfyUI nodes for experimenting with, combining, testing, and saving image-generation models. It lets you make a controlled blend of two models, adjust a LoRA without retraining it, compare VAEs using the same image, or check whether a model repeatedly generates the same face.
The extension is a companion to ComfyUI_LC123_nodes and ComfyUI_LC_AV_nodes. The author keeps the model-building tools in this separate pack so LC123 does not need their additional dependencies.
You can use these nodes for three broad tasks:
- Build: Merge checkpoints or VAEs, bake LoRAs into a model, extract a finetune as a LoRA, or save a smaller copy of a model.
- Adjust: Change which model contributes to particular blocks, or reduce the influence of selected parts of a LoRA.
- Evaluate: Compare VAE results, measure face variety, keep score during A/B tests, and investigate what models or LoRAs may have contributed to a checkpoint.
Several LoRA editing and merging nodes are marked beta. Keep your original files and test edits with a fixed prompt and seed before relying on a saved result.
How ComfyUI_LC_ModelBuilder_nodes Works
A model contains many groups of weights. Think of them as parts of a large recipe: changing one group may affect the result differently from changing another. A block merge gives you a separate mixing control for each numbered group. At a value of 0, that group comes entirely from model1; at 1, it comes entirely from model2; 0.5 blends them equally. Other model components have their own sliders.
A LoRA is a set of changes applied to a model. The LoRA Editor lets you turn parts of those changes up or down while you preview the result. LoRA Bake instead puts the changes into a newly saved model, so you can load that model without separately loading those LoRAs.
A VAE turns a model’s internal image representation, called latents, into the picture you see. VAE Compare decodes the same latents with different VAEs. That makes differences in color, texture, or sharpness easier to judge because the underlying generation stays the same.
Other nodes help you assess results rather than change them. Face Variety Scorer compares detected faces across a batch. Fingerprint tools compare model weights to references or to a library you have built. These comparisons are useful evidence, but they do not replace looking at rendered images.
For a fair visual test, keep the prompt and other generation settings consistent, compare multiple seeds, and change one setting at a time.
ComfyUI_LC_ModelBuilder_nodes Features
Merge checkpoints by block
LC Krea2 Block Merge Advanced blends two Krea 2 checkpoints across a draggable curve covering their 28 main blocks. Ten separately named regions have their own sliders. Choose smooth editing to move a block and nearby blocks together, or spike to change just the block under the cursor. You can save your curve and slider values as a named preset.
The default 50_50 preset blends everything evenly. Other bundled starting points include face_favor, detail_focus, style_favor, sine_wave, resonance_blocks, even_blend, core_focus, and heavy_transplant_anchored. Their names describe the intended blend shape, not a guaranteed visual effect: render a comparison to see what a particular pair of models actually does.
LC Anima Block Merge Advanced provides a similar 28-block curve for Anima, with separate controls for its x embedder, timestep components, LLM adapter, and final layer. Its bundled presets are 50_50, model2_early, model2_look, and model2_late.
LC MOSS Block Merge Advanced covers MOSS Image’s 42 layers and provides separate controls for image input, text input, timestep, shared modulation, and the final layer. It includes 50_50, model2_early, model2_look, and model2_late presets.
LC Model Merge (any precision) is for models saved in different numerical formats, such as bf16 and fp8. Set ratio for an overall blend, or supply block_ratios for individual numbered blocks. Read its report to see the input formats and the resulting format. When one input is full precision and the other is quantized, the node decodes the quantized input and keeps the result full precision. If both inputs are quantized, the result is rounded to model1’s format.
Tune a single model or edit LoRAs
LC Krea2 Semantic Vector Tuner adjusts regions of one Krea 2 checkpoint; it does not combine two checkpoints. Its low, mid, and high controls cover blocks 0–8, 9–17, and 18–27. Use blocks_override for a 28-value, block-by-block setting, the named-region sliders for other components, and base_scale for anything those controls do not cover. Soft Value limits the effect of settings above 1.0; Real Value applies the chosen multiplier directly. These are experimental weight adjustments, so judge them by renders.
The following LoRA tools are beta and still being tested:
- LC Krea2 LoRA Editor applies an edited Krea 2 LoRA in place of a LoRA loader. Its 28 block bars use
1for unchanged, values below1to weaken a block, and values above1to strengthen it. Presets provide starting points for different block groups. You can also adjust attention, MLP, and outside-block layers, or usespike_softento reduce unusually dominant changes. Its heatmap shows where the LoRA is strong before and after editing. - LC Krea2 LoRA Block Scan renders a fixed seed batch with different block groups disabled, then reports which group most affects face similarity. It produces a suggested block curve you can connect to the Editor’s
blocks_ininput. The suggested curve is a starting point, not a guarantee that the face will become more varied. - LC Krea2 LoRA Merge combines two to four LoRAs. For its A/B block bars,
0favors only A,0.5uses both at full strength, and1favors only B. Choosesumto add their changes,tiesto address conflicting changes, orbalancedto keep a stronger LoRA from overwhelming a quieter one. Optional third and fourth LoRAs have their own strengths. - LC Anima LoRA Editor, Block Scan, and Merge provide corresponding controls for Anima’s 28 blocks and layer types. The Anima Block Scan uses an anime-character face comparison rather than a photo-face comparison.
- LC LoRA Save saves the output of an edited or merged LoRA as a standard LoRA file under
models/loras. The edits and strength are baked in, so the saved file can be loaded at strength1.0. Setrankto0to retain each layer’s rank, or choose another rank to refit the layers. Existing filenames are not overwritten.
For any beta LoRA workflow, test the output on the same seeds before saving, retain the source LoRAs, and check settings when reopening a workflow after an update.
LC LoRA Bake 🍞 incorporates up to four LoRAs into a newly saved model. Each LoRA has its own strength; an optional per-block curve changes those strengths across blocks. outside_blocks controls other layers, while device lets you choose CPU or GPU processing. The saved file records the recipe in its metadata. The author recommends baking from a bf16 or fp16 source before quantizing the result.
LC LoRA Extract turns the difference between a finetuned model and its base into a LoRA. Choose all layers, blocks only, or a custom block list. rank limits the compression rank, keep controls how much of each layer’s change is retained, and min_change excludes layers with very small changes. Use bf16 versions of both models where possible: quantization noise can make the extracted LoRA larger and less useful. With applied_loras on, patches already applied to the finetuned model are included in the extraction.
Save smaller models and compare VAEs
LC Model Quantize 🗜️ saves a smaller diffusion model from either a selected model file or the current MODEL input. The wired input can include a merge or baked-in LoRA. Available formats include fp8, mxfp8, nvfp4, int8, int4, and GGUF formats from Q8_0 down to Q2_K. Smaller files generally involve more change to the weights; very low-bit options can be especially hard on smaller models. The node reports drift, a measure of weight difference, but drift is not an image-quality score. Render examples before deciding which format meets your needs.
A file source is read layer by layer, while a wired MODEL must already fit in system memory. Starting from bf16 or fp16 avoids stacking the rounding effects of two quantization steps. The node saves to models/diffusion_models without overwriting an existing filename. GGUF outputs require ComfyUI-GGUF to load.
LC VAE Merge blends the decoder portions of two compatible VAEs while keeping the encoder from vae_a. Five sliders move from coarse image properties toward fine texture: start, coarse, medium, fine, and finest. The report shows how far apart the VAEs are. The author recommends blending only closely related VAEs, roughly under 10% difference in the report; substantially different decoders may produce oversaturated color or noise. For Wan- or Qwen-Image-style VAEs, gradual slider changes are recommended.
LC VAE Compare decodes the same sampler output with up to six VAEs and places the images and 1:1 detail crops side by side. It also reports sharpness and difference from the first VAE. Use it to check whether a VAE change improves the areas you care about, rather than relying on a number alone.
LC VAE Save writes a VAE to models/vae in fp32, bf16, or fp16. Fp32 is the safest option described by the author; bf16 and fp16 take less space. Existing files are not overwritten.
Save files and manage metadata
LC Checkpoint Save saves an all-in-one checkpoint and requires model, clip, and vae. LC Diffusion Model Save saves only the model and has no CLIP or VAE inputs.
Both provide enabled, which skips saving when off, and embed_workflow, which controls whether the node graph is included in file metadata. Their metadata setting can keep everything or use necessary only to retain loading information, architecture, and trigger words while omitting workflow and other descriptive details. The metadata option is also available on LC LoRA Save and LC LoRA Extract. Turn off workflow embedding or use necessary only when you do not want to share your graph.
LC Strip Metadata 🧹 cleans an existing saved model without changing its weights. necessary only retains information needed for loading plus architecture and trigger words; everything leaves an empty metadata header. It can write a separately named clean file or replace the original after the copy completes.
Test results and investigate model origins
LC Face Variety Scorer 🧑🤝🧑 compares faces across a batch, ideally made with the same prompt and different seeds. Its main score runs from 0.0 to 1.0: lower means greater face variety. same_pct counts face pairs above the chosen similarity threshold, crowd_score compares faces within group images, and the faces output shows a grid of what was detected. The score concerns faces, not changes in hair, clothing, or lighting.
Use the photo mode for photographic faces and face_model = anime (CCIP) for anime or illustrated characters. Anime mode measures the share of image pairs identified as the same character; it does not use crowd_score. The required face models download on first use. Photo mode uses InsightFace’s buffalo_l files, which are released for non-commercial research use only; check their license before commercial use.
LC Scoreboard 🏆 counts wins and ties while you compare two models, LoRAs, or settings. Press #1 Win, Tie, or #2 Win after each comparison. The node displays the leader and records the counts with the workflow; scoring does not require running the graph.
LC Model Fingerprint 🔍 compares a checkpoint’s weights with a reference model or LoRA. It can indicate whether models are the same, related, blended by block, or apparently unrelated, and whether a LoRA appears to have been merged in.
LC Fingerprint Library 📚 creates small fingerprints of models and LoRAs in your folders, adding new files on later runs. LC Fingerprint Identify 🔎 compares a model against that library and presents possible ingredients, estimated block ratios and LoRA strengths, confidence information, and a heatmap. It can identify only ingredients represented in your library; an unexplained portion does not, by itself, establish how a model was made.
ComfyUI_LC_ModelBuilder_nodes Models
This extension works with several model families; it is not a download pack of new image-generation checkpoints.
| Family or component | Relevant tools | When to use it |
|---|---|---|
| Krea 2 | Block Merge Advanced, Semantic Vector Tuner, Krea 2 LoRA tools | Blend two Krea 2 checkpoints by block, tune one checkpoint, or edit and combine Krea 2 LoRAs. |
| Anima | Block Merge Advanced and Anima LoRA tools | Blend Anima checkpoints or work with Anima LoRAs. Its Block Scan uses anime-character face scoring. |
| MOSS Image | MOSS Block Merge Advanced | Blend MOSS Image models across their 42 layers. |
| Other supported model families | General-purpose quantization, saving, fingerprinting, and other applicable nodes | Use the general tools where their inputs and model formats are supported; the fingerprint library recognizes families including Flux, SDXL, Qwen, Z-Image, and Wan. |
| VAEs | VAE Merge, Compare, and Save | Test or blend compatible image decoders independently of the diffusion model. |
Face scoring uses separate comparison models, not generation checkpoints: the photo mode uses InsightFace’s buffalo_l pack, while anime mode uses CCIP and an anime face detector. The photo models are placed under models/insightface/models/buffalo_l/; the anime models download to models/ccip/. The author states that these downloads are checksum-verified. The nodes run them through onnxruntime without requiring the insightface Python package or OpenCV.
Troubleshooting ComfyUI_LC_ModelBuilder_nodes
A block merge does not look the way its preset name suggests.
- Confirm that both inputs are the intended models and that
0meansmodel1while1meansmodel2. - Start with
50_50, then adjust one curve region or named slider at a time. - Render the same prompts and seeds after each change. Presets are starting points, not reliable labels for what every model pair will produce.
A LoRA edit does not improve face variety.
- Keep the prompt, seeds, and other settings fixed while comparing edits.
- Try another block-group preset or run Block Scan with a prompt showing one clear face.
- Inspect the scan’s grid and report before using its suggested curve. A LoRA may distribute its face changes across many blocks, so weakening one group may not be enough.
- Keep the original LoRA and test the saved version before replacing any workflow that uses it.
An anime face test gives unhelpful results. Select face_model = anime (CCIP) in Face Variety Scorer. Photo face recognition may miss illustrated faces. For Anima Block Scan, use its anime scoring and judge the rendered images as well as the score.
A VAE merge creates strange colors or noise. Read the merge report. If the decoders are far apart, choose a closer pair rather than forcing a blend. For Wan- or Qwen-Image-style VAEs, try a gradual progression across the five sliders; compare the result on the same latents with VAE Compare.
A quantized model looks substantially different from its source. Check that you started from bf16 or fp16, then try a less aggressive format. The reported drift measures changed weights, not visual quality. On smaller models, especially, compare actual renders before keeping a low-bit version.
A GGUF output will not load. Load it with the Unet Loader (GGUF) provided by ComfyUI-GGUF, rather than the stock diffusion-model loader. For the non-GGUF formats listed above, use Load Diffusion Model.
A saved file contains a workflow you did not intend to share. On a supported save node, turn off embed_workflow or choose metadata = necessary only before saving. For an already saved model, use LC Strip Metadata. Check the node’s selected output name so you know whether it will create a clean copy or replace the original after copying.
New node controls do not appear after an update. Hard-refresh the browser with Ctrl+Shift+R so it loads the updated interface code. Beta LoRA controls or presets may also have changed, so inspect their values in older workflows.
Frequently Asked Questions
Does a lower Face Variety Scorer value mean a worse image? No. It means the detected faces in the tested batch are less similar to one another. It is not a measure of image quality.
Can Fingerprint Identify name every ingredient in a downloaded model? No. It can name only models and LoRAs represented in your fingerprint library.
Will the save and quantization nodes overwrite an existing filename? The author states that LC Model Quantize, LC VAE Save, LC LoRA Save, and LC LoRA Extract add a number when a filename already exists. Check the options on other file-writing nodes, particularly LC Strip Metadata when selecting replacement of the original.
Do I need to retrain a LoRA to soften its effect? Not necessarily. The beta LoRA Editor can reduce selected blocks or layer types, and LoRA Save can write those tested edits to a new file.
Learn More about ComfyUI_LC_ModelBuilder_nodes
- The extension’s
workflows/examples include tests for face variety, mixed-precision merging, LoRA extraction, and A/B comparisons. Some examples also use nodes from ComfyUI_LC123_nodes; the LoRA Test workflow additionally uses LC Vision. - ComfyUI_LC123_nodes includes LC Image Batch From Folder, which can supply a folder of test images to Face Variety Scorer.
- ComfyUI-GGUF provides the loader needed for GGUF models made with LC Model Quantize.
- The InsightFace project documents the photo-face models used by the scorer. For anime comparisons, see CCIP and the anime face detector. Review the relevant model licenses before using their outputs commercially.
