| LoRA Stack Container | Collects a selection of 5 LoRAs (+ weights) into one LORA_STACK. Applies nothing to the model—only the data. | | Load LoRA (StackInput) | Takes model, clip, lora_stack and applies the entire stack. The only note touching the model. | | Trigger Filter | Accepts trigger words from container(s), removes duplicates, returns one ready-made line for the prompt. |

============================================================================

1. LoRA Stack Container - `LoraStackContainer5`

Noda-container for 5 slots. Each slot has 4 lines, with an empty line between slots:

Compact tumbler slotN_enabled (standard view of ComfyUI).
⚙ Slot N Info button — opens a pop-up with complete information about selected LoRA (description below).
Select LoRA — slotN_lora_name.

slotN_strength_model. 5. slotN_strength_clip.

Inputs: lora_stack (optional) - Connect the output of another here container to combine more than 5 LoRAs into one stack.

Outputs:

lora_stack — combined stack, to be passed to Load LoRA (StackInput).
trigger_words_1 … trigger_words_5 — one per slot, typ LORA_TRIGGER_WORDS (only compatible with nodes of this pack, normal It is physically impossible to connect STRING'). Each line is a **complete** list tags of this LoRA (without clipping), sorted by decreasing number training images, and always ends with a comma ("cat, dog, blue eyes,") — so that when several such lines are connected in a row, the last tag of one is not merged with the first tag of the next one.

============================================================================

2. Load LoRA (StackInput) — `LoraStackInputLoader`

The only purpose is to apply the stack to the model.

Inputs:

model, clip are mandatory.
lora_stack' is optional; if empty, node just skips model/clip` unchanged (except for the CLIP layer, see below).
stop_at_clip_layer is an integer -1..-24 (as in standard ComfyUI CLIPSetLastLayer node), default is -1. Applies to clip always, whether lora_stack is connected or not.

A node does not have its own LoRA selection — all selection is done in the Container(s).

============================================================================

3. Trigger Filter - `LoraTriggerFilter`

Takes trigger words from one or more containers, removes duplicates, etc adds your own words with priority.

Inputs:

trigger_words_1 … trigger_words_5 (type LORA_TRIGGER_WORDS, optional) — are connected only from trigger_words_N outputs of LoRA Stack nodes Container. None are required - noda works with even one.
extra_words (ordinary STRING) — text field right at the foot, you can enter manually, or connect by wire from any text node, including the filtered_words of another Trigger Filter (filter chain).

Output: filtered_words is one finished string.

Filtering logic

extra_words has absolute priority. Its words are processed first; if the same word occurs further among trigger_words_1..5, this later take is simply ignored.

Among trigger_words_1..5 is a simple string comparison without counting register A word that occurs in more than one place, completely hides from the result (there is no option left). Reinforcement in styles (tag:1.3) are not recognized here - LoRA tags are always simple words without that syntax, they don't need that logic.

Live preview without running Queue

Until the graph is run, the node's Python code has not yet executed—but the Trigger Filter has holds the hidden service field text, which updates itself: as soon as you change the LoRA in the connected container or the text in extra_words, the node in ~0.25 s turns to POST /lora_stack_pack/filter and receives the same result that the real execution will give (the logic is shared, lora_utils.filter_tag_texts, so they cannot diverge).

This allows third-party nodes, which read the connected text widget nodes live (for example, CLIP Text Encode mod with Tag Toggles Reader), see tags without triggering generation.

If a wire is connected to the extra_words pin, the preview takes the text from the connected one nodes, and not from their own field. Supported: node with text widget (text, string, value, prompt, tags or the first line field), pass through Reroute, and another ``Trigger Filter'' (its live result is taken — yes you can build chains).

Limitations of live preview:

Only understands direct Container → Trigger Filter connections. If between they are other nodes — the preview is empty (the actual execution works at the same time as usual).
Text, which is only computed in Python at startup (does not lie in the no widget), is not visible in the editor.
The `text' field is not saved in the workflow — it is recalculated when every discovery.

============================================================================

Popup ⚙ Info

Opens with a button on each container slot. Shows everything that was successful find about the selected LoRA.

Top:

Name, Base Model, Clip Skip, Resolution.
Notes — auto-detected source link (Civitai page, if found) + your own notes: button ✏ Edit → text field → Save saves them to a small sidecar file next to LoRA, notes experiencing a ComfyUI restart.
Preview image. If several application images are found - sub a row of thumbnails appears as the main picture, a click switches the preview. The "Use as preview" button saves the selected image locally (<name>.preview. png/jpg/webp) - since then it is always shown, without referring to the network. The old local image is deleted automatically.

Tags:

The search field — the entered text highlights the match in yellow and mutes the rest, automatically scrolls to the first match.
Tag grid, sorted by descending number of trainees images. Default is top 200; if there are more tags - a button "Show all (N)" loads the full list.
A separate field Trained Words is an official, curated list of trigger words from the sidecar file (if any), regardless of tag statistics from the dataset.
Copy Selected / Copy All (a comma is added at the end of the copied text) / View raw metadata / Close.

Control: is closed with the Esc key, a click outside the window or with the Close button. Adapts to the size of the screen - completely scrolls as one whole on small windows.

Where do popup data come from?

A normal LoRA file (.safetensors) does not contain trigger words, images or model references are technical training parameters only. These data are taken in one of two ways:

Local sidecar file next to LoRA (<name>.civitai.info, .civitai.info.json or .json) - if the previous one left one LoRA manager.
Live search on Civitai by file hash if there is no sidecar file: the node calculates the SHA256 of the entire file and asks https://civitai.com/api/v1/model-versions/by-hash/<hash>. The result cached in <name>.civitai.info next to the file - next time read instantly and offline.

Works for any LoRA that was once on Civitai. If LoRA is trained locally and has never been published - there will be no data, it is expected. The first a new LoRA request requires Internet access and may take several seconds (a pop-up shows a loading indicator).

The numbers next to the tags in the tag grid are the number of training images, in signatures of which this tag occurred (data from ss_tag_frequency, which stores kohya_ss coach himself right in the file).

Lora Stack Pack

Nodes (3)

1. LoRA Stack Container - LoraStackContainer5

2. Load LoRA (StackInput) — LoraStackInputLoader

3. Trigger Filter - LoraTriggerFilter

Filtering logic

Live preview without running Queue

Popup ⚙ Info

Where do popup data come from?

1. LoRA Stack Container - `LoraStackContainer5`

2. Load LoRA (StackInput) — `LoraStackInputLoader`

3. Trigger Filter - `LoraTriggerFilter`