Stack LoRAs, pick their trigger words visually, and apply them — including automatic block remapping so LoRAs trained on Anima still work on Anima 2.9B / 3.8B.

What you get

| Node | What it does | Outputs | |---|---|---| | LoRA Stack Auto Triggers | Pick up to 4 LoRAs and choose their trigger words in a visual picker. Trigger words are fetched automatically from Civitai and from the training metadata inside the model file. | LORA_STACK, LORA_PROMPT, LORA_MetaData | | LoRA Stack Apply | Plain applier: takes any LORA_STACK and applies it to a model. | MODEL, CLIP, INFO | | LoRA Stack Apply (Anima Remap) | Same applier, plus cross-generation block remapping for Anima models. | MODEL, CLIP, REMAP_INFO |

Why two appliers? So you don't need a third-party "apply LoRA stack" node to use this pack, and so Anima users can swap in the remapping version only when they need it.

Install

Via ComfyUI Manager / Registry — search for LoRA Stack Auto Triggers.

Manual

cd ComfyUI/custom_nodes
git clone https://github.com/opaiopaio/ComfyUI-LoRAStackAutoTriggers.git

Restart ComfyUI. No extra Python packages are required (requests ships with ComfyUI).

Quick start (5 steps)

Add LoRA Stack Auto Triggers.
Turn on lora_1 and pick a LoRA from lora_name_1. The summary line updates to something like 28 layers · 4 variants / 1749 training tags.
Click ⚙ Triggers 1. In the dialog:
- Variants – Civitai usually publishes whole trigger sentences (one per costume). Click one to load it.
- Segment chips – after loading a variant it is split on commas. Click a chip to drop that part (e.g. see-through mouth mask), click again to bring it back.
- Training tags – the raw training-set tag dump, searchable. Click to append.
- Free edit – edit the text directly, or select part of it with the mouse and press Apply selection to keep only that part.
Press Apply all (or Apply selection).
Feed LORA_PROMPT into your text encoder, and the stack into an applier (LoRA Stack Apply / (Anima Remap), or any other node that accepts LORA_STACK).

The trigger picker

| Tab | Use it for | |---|---| | Variants | Pick one of the Civitai trigger sentences | | Segment chips | Remove optional parts of a variant without retyping anything | | Training tags | Pull single tags out of the training set (searchable) | | Free edit | Hand-edit the final text, or keep only a mouse-selected portion |

What ends up in LORA_PROMPT

If trigger words are available: just the trigger words (no <lora:...> tag).
If nothing is found and fallback_to_filename is on: <lora:name:weight>.
fallback_to_filename off: that LoRA is simply left out of the prompt (it is still applied to the model).

Nothing is hidden from your workflows: the selected text lives in a trigger_selector_* field that the UI keeps out of the way, so existing workflows and the legacy slice syntax (0, 3, 5:8, :) keep working.

Node reference

LoRA Stack Auto Triggers

Four identical slots. Per slot:

| Field | Meaning | |---|---| | lora_N | Enable this slot | | lora_name_N | Which LoRA | | model_weight_N | LoRA strength on the model; 0 = no model effect | | clip_weight_N | LoRA strength on the text encoder; 0 = no CLIP effect (model still applies) | | ⚙ Triggers N | Opens the picker | | trigger_weight_N | Wraps every selected segment as (segment:weight); 1.0 = no wrapping | | trigger_preview_N | Read-only one-line summary |

A slot is active when it is enabled and a LoRA is selected and at least one of the two weights is non-zero. Model and CLIP weights are independent — setting CLIP to 0 gives you "model only" instead of disabling the LoRA.

Bottom settings:

| Setting | Meaning | |---|---| | save_hash | Add the file's SHA256 (first 12 chars) to LORA_MetaData | | force_fetch | Ignore the local cache and re-fetch from Civitai (refreshes all slots) | | fallback_to_filename | Emit <lora:name:weight> when no trigger words exist | | lora_prompt / lora_metadata | Prefix incoming text, for chaining nodes |

Optional inputs lora_stack / lora_prompt / lora_metadata let you chain several stack nodes together.

LoRA Stack Apply

| Input | Notes | |---|---| | model | Where the stack is applied | | clip (optional) | Without it, CLIP weights are ignored | | lora_stack (optional) | From this pack or any node that outputs LORA_STACK | | bypass | Pass model/CLIP through untouched | | weight_multiplier | Scale every weight in the stack at once | | apply_clip | Off = never touch the text encoder |

Outputs MODEL, CLIP, INFO (a per-LoRA log of what was applied or skipped).

LoRA Stack Apply (Anima Remap)

Everything above, plus:

| Input | Notes | |---|---| | enable_remap | Off = plain applier (no remapping) | | target_model | Auto (detect) / Anima (28 layers) / Anima 2.9B (40 layers) / Anima 3.8B (52 layers) | | source_model | Which generation the LoRA itself was trained on | | remap_mode | Keys only (recommended) / Extend to new layers (experimental) | | extend_strength | Only used by Extend mode; 1.0 copies the source delta as-is |

Outputs MODEL, CLIP, REMAP_INFO, for example:

target: Anima 2.9B (40 layers)
applied: 1 lora(s)
zzz_remielle_anima_v2-rift1000: 28->40 remapped (840 moved, 0 dropped)  [w=0.85, clip=0]

Anima cross-generation compatibility

Anima generations are depth expansions, not retrains: 12 new blocks are spliced into the middle of the previous layout.

28 layers (Anima base):  0  1  2  3  4  5 ...
40 layers (Anima 2.9B):  0  1 [2] 3  4 [5] 6  7 [8] ...   [n] = newly inserted
                         ^  ^     ^  ^     ^  ^
                         0  1     2  3     4  5            <- old blocks shifted

A LoRA key says blocks.2. On a 40-layer model blocks.2 does exist (it is an inserted block), so nothing errors — the deltas just land on the wrong layers. This node rewrites the LoRA's block indices before applying it.

Two dropdowns, on purpose. target_model is usually detected correctly from the connected model, but source_model is inferred from the LoRA's highest block index — which mis-reads adapters that only touch early layers. When in doubt, set it by hand.

Two remap modes

| Mode | Effect | |---|---| | Keys only | Indices are moved; the 12 newly inserted layers get no delta from this LoRA, so its effect is weaker than on the original model | | Extend | Additionally copies each source block's delta onto the layers inserted from it (scaled). Best effort — those layers did not exist when the LoRA was trained |

Direction is one-way. Older to newer only (28→40, 28→52, 40→52). A LoRA that needs more layers than the model has raises an error instead of silently applying half of itself.

Coexistence. Two community plugins (ComfyUI-Anima-2.9B-loraPatch, ComfyUI-Anima-LoRA-ControlNet-Patch) patch comfy.lora.load_lora globally. This node does not patch anything — it rewrites its own copy of the state dict, so both can be installed side by side without double-remapping. You only need one of them anyway.

Language

Node titles, categories, descriptions and tooltips follow the system language (Chinese or English). The picker dialog follows the browser language.

Force a language with an environment variable before starting ComfyUI:

OPNODES_LANG=zh    # or en

Dropdown values (Auto (detect), Anima 2.9B (40 layers), …) are intentionally not translated: ComfyUI stores the selected value inside the workflow, so translating them would make a workflow saved under one language unreadable under another.

Troubleshooting

| Symptom | Cause / fix | |---|---| | Picker button missing, raw trigger_selector_* fields visible | The frontend extension did not load. Hard-refresh (Ctrl+Shift+R); if it persists, check the browser console for [LoRAStackAutoTriggers] lines | | Trigger words are empty | The LoRA has no Civitai entry and no ss_tag_frequency metadata. Tick force_fetch once, or type words by hand in the picker | | Wrong words after replacing a LoRA file | The cache is keyed by LoRA name. Toggle force_fetch to refresh, or delete loras_tags.json | | Grey / black image with no error | A NaN weight reached the applier. Current versions treat NaN as 0 and skip that entry — update if you are on an older build | | Anima LoRA looks weak on 2.9B / 3.8B | Expected in Keys only mode (the 12 new layers get nothing). Try Extend with strength 0.3–0.8 |

Credits

Trigger-word extraction is ported from idrirap/ComfyUI-Lora-Auto-Trigger-Words (MIT, © konohana) — itself a fork of Extraltodeus/LoadLoraWithTags.
The stacked-loader layout follows jakechai/ComfyUI-JakeUpgrade.
The Anima block-mapping tables are the ones community-verified by sparklingcoffee777/ComfyUI-Anima-2.9B-loraPatch and INuBq8/ComfyUI-Anima-LoRA-ControlNet-Patch (exhaustive torch.equal verification). No code from those GPL projects is used here.

License

This pack is MIT. Note that Anima models themselves are non-commercial, and remapped LoRA files produced from them count as derivatives under the CircleStone Labs license. Images you generate can still be used commercially.