ComfyUI-NunchakuFluxLoraStacker Introduction
ComfyUI-NunchakuFluxLoraStacker is a collection of ComfyUI nodes for building and managing image-generation workflows. Its central tools let you apply several LoRAs in one place, instead of adding a separate loader for every LoRA. The collection also includes tools for describing images, cleaning up prompts, choosing image dimensions, managing groups of nodes, and upscaling finished images.
A LoRA is a small add-on that changes how a compatible base model produces an image. You might use one for a visual style and another for a subject. A stacker gives you a single place to choose those files and adjust how strongly each one influences the result.
The collection offers different LoRA nodes for different workflows:
- FLUX LoRA Loader V2 works with Nunchaku FLUX models.
- LoRA Stacker V2 and V3 work with standard ComfyUI
MODELandCLIPconnections, including workflows using SDXL, Flux, and WAN2.2. - SDNQ LoRA Stacker V2 is intended for SDNQ quantized-model workflows.
- The legacy Nunchaku FLUX LoRA Stack remains available for older workflows and for its separate model and CLIP strength controls.
The nodes are not all dependent on Nunchaku. If the nunchaku package cannot import, as may happen on AMD/ROCm systems, Nunchaku FLUX features remain disabled while the other available nodes can still load.
How ComfyUI-NunchakuFluxLoraStacker Works
Think of your base model as the starting point for an image and each LoRA as an optional adjustment. A LoRA stacker passes the model through a series of selected adjustments, then sends the resulting model to the next part of your workflow. The strength setting controls each adjustment’s influence. For example, you could place a character LoRA in one slot and a style LoRA in another, then change their strengths while comparing results.
The newer stackers let you choose how many slots to show, from 1 to 10. Choosing three slots displays only the controls for those three, keeping the node easier to read. A visible slot does not necessarily change the image: it needs a selected LoRA, a nonzero strength, and, in V3, its enable switches must be on.
Other nodes handle different parts of the creative process:
- Florence-2 can examine an image and produce a caption or perform tasks such as reading text in an image.
- Color Filter can remove monochrome-related wording from that caption before you use it in a prompt.
- Nunchaku Resolution Selector can supply preset or custom image dimensions to a workflow.
- CCSR can upscale an image through its TensorRT-based processing path.
- Fast Groups Bypasser and Use Everywhere help organize and route connections in larger workflows.
These are separate tools. You can use the ones that fit your workflow without using every node in the collection.
ComfyUI-NunchakuFluxLoraStacker Features
LoRA loading and comparison
| Node | Best suited to | Main controls |
|---|---|---|
FLUX LoRA Loader V2 (FluxLoraMultiLoader_10) | Nunchaku FLUX models | A dropdown selects 1–10 visible LoRA slots; each slot has an optional file and strength, defaulting to 1.0. |
LoRA Stacker V2 (LoraStackerV2_10) | Standard ComfyUI model and CLIP workflows | A dynamic 1–10-slot interface for combining LoRAs. |
LoRA Stacker V3 (LoraStackerV3_10) | Standard workflows where you want quick comparisons | The dynamic slots of V2, plus a master toggle_all switch and an enabled switch for each slot. |
SDNQ LoRA Stacker V2 (SDNQLoraStackerV2_10) | SDNQ quantized models | A dedicated dynamic 1–10-slot interface, designed for use with comfyui-sdnq-splited. |
Nunchaku FLUX LoRA Stack (NunchakuFluxLoraStack) | Existing legacy workflows | lora_count selects 1–10 slots. Simple mode gives each LoRA one strength control; advanced mode separates model and CLIP strengths. |
In LoRA Stacker V3, turning toggle_all off passes the model and CLIP through without applying any LoRAs. With it on, you can switch individual slots off while leaving the others active. This makes it easier to compare an image with and without one particular LoRA. V3 uses the same strength value for the model and CLIP; its lora_strength_X controls accept values from -100.0 to 100.0, in steps of 0.01. A slot with no selected LoRA or a strength of zero is not applied.
The legacy node’s advanced mode is useful when you specifically need different model and CLIP strengths. V2 was introduced to support the widget and input handling used by ComfyUI Nodes 2.0 (Desktop); the legacy node remains available so older workflows do not have to change immediately.
<img src="png/FLUX%20LoRA%20Loader%20V2.png" width="400" alt="FLUX LoRA Loader V2 node">
<img src="png/LoRA%20Stacker%20V3.png" width="400" alt="LoRA Stacker V3 node">
Image understanding and prompt preparation
Florence-2 is a family of four nodes for working with images and text. You can load a Florence-2 checkpoint from Hugging Face or from a local ComfyUI/models/LLM folder, optionally connect a supported PEFT LoRA, and use Florence2 Run for tasks including captioning, detailed captioning, OCR (reading text in an image), document question answering, region proposals, grounding, and segmentation. Depending on the task, the run node can return text, structured data, an annotated image, or a mask. Loader controls include precision and attention-mode choices; you do not need to change those to understand the basic captioning workflow.
Universal LoRA Analyzer (UniversalLoRAAnalyzer) helps you inspect a LoRA file without loading it into the image-generation graph. It can report information such as model type, trigger words, base model, and Civitai or Hugging Face URLs when available.
Color Filter (ColorFilter) takes caption or tag text and removes built-in monochrome-related expressions such as “black and white,” “monochrome,” and “grayscale.” You can add your own comma- or newline-separated phrases through exclude_words. For example, when a caption describing a reference photograph says “a black and white portrait,” filtering that phrase can prevent it from unintentionally steering a later prompt toward a grayscale image. The output also combines consecutive whitespace into single spaces, including across line breaks.
<img src="png/Florence2.png" width="400" alt="Florence-2 nodes">
<img src="png/colorfilter.png" width="400" alt="Color Filter node">
Size, upscaling, and workflow organization
Nunchaku Resolution Selector (NunchakuResolutionSelector) provides Flux1-oriented aspect-ratio presets, including approximately 1.0-megapixel options and High variants at approximately 1.5 megapixels. Choose Preset for a listed size or Custom to enter width and height. Its hires_scale control, defaulting to 1.3, supplies a second pair of dimensions for a higher-resolution stage. It also outputs an empty 16-channel latent, a batch size, and an information string summarizing the selection.
CCSR TensorRT provides two nodes: LoadCCSRModelTensorRT loads the engine and its companion weights, and CCSR_Upscale_TRT produces an upscaled image. This path uses a fixed 512-pixel tile and requires an NVIDIA RTX GPU. The repository describes its TensorRT processing as approximately 1.4× the speed of its earlier fp16 PyTorch path; your results may depend on your system and workflow.
Fast Groups Bypasser V2 lets you control groups of nodes with checkboxes. It can filter groups by color or title pattern, sort them, and apply bypass or mute behavior. It is a workflow-management tool, not a LoRA loader.
Use Everywhere supplies nodes for broadcasting and virtually routing values through a workflow, such as hswq_AnythingEverywhere and hswq_SeedEverywhere. For its broader set of routing rules and examples, see the upstream Use Everywhere guide.
The collection also includes ControlAltAI utility nodes. Their individual controls and screenshots are documented in the ControlAltAI node reference.
<img src="png/Resolution%20Selector.png" width="400" alt="Nunchaku Resolution Selector node">
<img src="png/ccsrtensor.png" width="400" alt="CCSR TensorRT nodes">
ComfyUI-NunchakuFluxLoraStacker Models
This extension is primarily a set of ComfyUI nodes, not a single image-generation model. Choose a node that matches the model and connections already used by your workflow:
| Model or model family | Relevant nodes | When you might use it |
|---|---|---|
| Nunchaku FLUX | FLUX LoRA Loader V2; legacy Nunchaku FLUX LoRA Stack | When your workflow loads a Nunchaku FLUX base model and you want to combine compatible LoRAs. |
| Standard ComfyUI models, including SDXL, Flux, and WAN2.2 workflows | LoRA Stacker V2 or V3 | When your checkpoint workflow provides standard MODEL and CLIP connections. Choose V3 when individual on/off comparisons matter. |
| SDNQ quantized models | SDNQ LoRA Stacker V2 | When your workflow uses the SDNQ model path supported by the dedicated stacker. |
| Florence-2-family checkpoints | Florence-2 loading and run nodes | When you want to describe, read, or analyze an input image rather than generate one with a LoRA stack. |
| CCSR TensorRT engine and auxiliary weights | CCSR TensorRT loader and upscaler | When you want to upscale an image on a supported NVIDIA RTX system. |
LoRA files must be compatible with the base model and loader you choose. The context does not identify a complete list of individual LoRA or Florence-2 checkpoint presets, so this README does not recommend particular files as universally compatible.
What’s New with ComfyUI-NunchakuFluxLoraStacker
- V2 nodes, introduced in v1.12: Dynamic 1–10-slot LoRA interfaces were added for ComfyUI Nodes 2.0 (Desktop). The legacy Nunchaku FLUX node was kept for existing workflows and its simple/advanced strength modes.
- LoRA Stacker V3: Added a master switch and per-slot enable switches. You can compare LoRA combinations without repeatedly removing and reselecting files.
- Broader tool collection: The repository includes Florence-2 image-understanding nodes, LoRA analysis, caption filtering, resolution selection, group controls, Use Everywhere routing, ControlAltAI utilities, and CCSR upscaling.
- Florence-2 integration: The integrated nodes include support described by the author for Transformers 5.x and Sage Attention 3. If a Sage mode is selected with a Transformers version older than 5.0, the loader falls back to SDPA and logs a warning.
- CCSR TensorRT path: The collection now provides a TensorRT-only CCSR upscaling path in place of its earlier PyTorch CCSR nodes.
- Resilience when Nunchaku cannot load: Nunchaku FLUX features are disabled without preventing the other available nodes from loading. The author credits the import-guarding approach to 0xDELUXA’s pull request.
These points summarize changes described in the supplied README; they are not a version-by-version release list. See the changelog for the full release history.
Troubleshooting ComfyUI-NunchakuFluxLoraStacker
A Nunchaku FLUX node is unavailable, but other nodes appear
- Check whether your workflow uses a Nunchaku FLUX node rather than a standard LoRA stacker.
- Check ComfyUI’s startup messages for a failure to import
nunchaku. - If Nunchaku is unavailable on your system, use an appropriate non-Nunchaku node for a standard
MODEL/CLIPworkflow. The non-Nunchaku tools are designed to remain usable when Nunchaku features are disabled.
A LoRA does not appear to affect the image
- Confirm that you selected a LoRA file in a visible slot.
- Confirm that its strength is not zero.
- In LoRA Stacker V3, turn on both
toggle_alland that slot’senabledswitch. - Check that the LoRA and base model are compatible.
- Compare the same workflow with the stack enabled and bypassed so you can judge the effect more easily.
An expected LoRA slot is missing
Set the node’s LoRA Count dropdown, or the legacy node’s lora_count, to the number of slots you want visible. The dynamic nodes intentionally hide slots above that count and resize to fit the displayed controls.
Color Filter removed wording you wanted to keep
Review exclude_words first and remove any custom phrase that should remain. The node also removes built-in monochrome-related wording, so use the original caption when that wording is important to your intended image.
CCSR upscaling is unavailable
Check that your system has the NVIDIA RTX GPU required by the TensorRT nodes. The CCSR path also depends on its TensorRT engine and companion auxiliary weights. Its loader uses the engine file ccsr_apply_f16io.rtxplan alongside ccsr_trt_aux.safetensors; the supplied README says missing files can be downloaded on demand and verified when the loader runs.
Frequently asked questions
Do I need Nunchaku to use every node?
No. When nunchaku cannot import, the Nunchaku FLUX-oriented features are disabled, while other available nodes can still load.
Should I use FLUX LoRA Loader V2 or LoRA Stacker V3?
Use FLUX LoRA Loader V2 for a Nunchaku FLUX model input. Use LoRA Stacker V3 for a standard ComfyUI workflow with MODEL and CLIP inputs, particularly when you want to switch individual LoRAs on and off.
Why keep the legacy Nunchaku FLUX node?
It preserves existing workflows and offers advanced controls with separate model and CLIP strengths.
Does selecting more LoRA slots automatically apply more LoRAs?
No. The count controls which slots are shown. You still need to select files and give them active, nonzero strengths.
Does the Resolution Selector generate an image by itself?
No. It supplies dimensions and an empty latent for other nodes in your workflow to use.
Learn More about ComfyUI-NunchakuFluxLoraStacker
- Full release history — details beyond the update summary above.
- ControlAltAI node reference — individual utility nodes, parameters, and screenshots.
- Use Everywhere documentation and usage guide — concepts and examples for virtual routing in larger ComfyUI workflows.
- Florence-2 upstream project — background on the image-understanding nodes integrated into this collection.
- rgthree-comfy — upstream project for the Fast Groups Bypasser utility.
- CCSR upstream project — background on the image-upscaling method.
- Nunchaku FLUX LoRA stacker repository — the project’s source and accompanying documentation.
