# LoRA Loader Applies a whole stack of LoRAs to a model (and optionally CLIP) from one compact node. Each row is one LoRA: an on/off pill, a searchable picker, and strengths you can drag like a slider. Trigger words from every enabled row come out of one socket, ready to wire into a prompt. ## Rows - **grip** (the dotted handle): Drag up or down to reorder the stack. LoRAs apply in row order, so the order is part of the recipe; the move commits when you drop, and a cancelled drag changes nothing. - **toggle**: Enable or disable the row without losing its settings. - **name**: A plain click opens the picker — type to search, **Arrow Up/Down** to highlight, **Enter** to select, **Escape** to close; the list is refreshed from `models/loras` every time it opens. Dragging left/right on the name instead **scrubs the model strength**, and the bar behind the text rides along. Rows show just the file name by default; the full path lives in the tooltip and the picker keeps its folder groups. A name too long for its row is cut at the end; gear menu → **Long names** → `middle` cuts the middle instead, so the checkpoint number at the end (`…_000004000`, `…_epoch_10`) stays visible in a narrow node. The picker opens wide enough for long names; when one still does not fit, it shortens the folder first and keeps the end of the file name. - **strength**: Drag left/right on the number to scrub it, click to type an exact value, use the chevron arrows, or the arrow keys. Hold **Shift** while scrubbing or stepping for fine 0.01 moves. Range is -10 to 10. - **i**: Info card with the preview image, base model, file size and modified date, and trigger words for that LoRA. Right-click a row to insert an empty LoRA above or below it, move, duplicate, or remove it. **Delete all LoRAs…** is a separate red, trash-marked action with confirmation; it clears this node's rows and keeps all files on disk. Cancel or Escape keeps the stack. Inserting keeps the existing rows and their settings in order; the new slot uses the gear menu's default strength (initially 1). The stack holds up to 64 rows. One strength drives both model and CLIP by default; the gear menu's **Separate model / CLIP strength** switch shows a second box per row. The switch is per-node — new nodes always start unified — and an absorb that imports a row with unequal model/CLIP strengths flips it on automatically so the difference stays visible. ## Strength bars Every named row paints a center-zero bar behind its name: teal grows right of center for positive model strength, muted red grows left for negative, a 1px tick marks zero and a brighter 2px cap marks the value's edge. One shared scale across the stack: the field edges represent the largest absolute strength among the **enabled** rows, floored at 1.0 — so an everyday stack of 0.5s reads absolutely as half-bars, one row at 2.0 rescales every bar rather than clipping, and switching that row off hands the scale back to the strongest row still in play (the dimmed row's own bar clamps at the edge meanwhile). The scale in play is named in each row's tooltip. Gear menu → "Strength bars" turns them off; "Scrub strength on the name" turns the name-drag off. ## The control bar The bar rides the empty middle of the output-slot band, so the node spends its height on rows, not chrome. Left to right: - **▤ templates** — save the current stack under a name, or apply/delete a saved one. Templates are kept in this browser and shared by every workflow opened in it. - **master pill** — cycles *mixed → all on → all off → back to the mixed setup it destroyed*. Every hand-made row toggle refreshes that memory, so an accidental master click is always one more click from home. - **summary** — how many rows are on. - **↻ reconnect** — re-checks every row against a fresh `models/loras` list and repairs what it can: a row whose file moved folders (or came back after being missing) is rewritten to the resolved path. ComfyUI's own **R** refresh runs the same pass quietly. - **⚙ gear** — settings, and the **Absorb chain LoRAs** action below. **+ LoRA** sits inside the stack container, pinned to its bottom edge. ## Absorb the loader chain Gear menu → **Absorb chain LoRAs**. Walks the model chain on BOTH sides of this node — upstream through the `model` input, downstream from the `model` output — and for every loader it recognizes (core `LoraLoader` and `LoraLoaderModelOnly`, rgthree Power Lora Loader and Lora Loader Stack, JPS Lora Loader, `PixaromaLoraLoader`, another AusBoss LoRA Loader; Reroutes are walked through) lifts its rows into this stack in application order: upstream loaders, your existing rows, then downstream loaders. The original loaders are set to **bypass** only when their connections can be preserved. Repeated filenames remain separate rows, with their own strengths and toggles. Details that keep the absorb faithful: - the walk follows connections, not canvas position, and stops at unknown nodes and downstream forks - JPS rows switched off stay disabled; an rgthree Lora Loader Stack strength applies to both model and CLIP - names are resolved against this install's `models/loras` — exact match first, then a unique basename match ignoring folders and extension (for workflows saved on another machine's layout); unresolved names import verbatim and the rows tint red - `LoraLoaderModelOnly` rows import with CLIP strength 0 - repeated LoRAs are kept: applying a file twice is part of the original recipe - a shared model path, including a branch hidden behind a reroute, blocks absorption instead of changing another branch - when CLIP patches move, CLIP must follow the same serial chain as MODEL; rows from a loader without a connected CLIP import with CLIP strength 0 - connected auxiliary outputs, such as trigger words, block absorption because bypassing their source would break those connections - linked LoRA settings and stacks beyond the row limit leave the sources active - already-bypassed or muted loaders contribute no rows - a toast summarizes what was imported, bypassed, remapped, and skipped ## Moved and missing files Rows are checked against the install's LoRA list. A row whose file moved folders gets a dashed border and a tooltip naming the file the run will use; a row with no match anywhere tints red; an ambiguous basename tints red with advice to re-pick. Thumbnails, the info card, and range lookups all use the resolved name, so a moved file keeps its previews and Civitai data. The run side matches: a stale saved path resolves by name at load time (exact → unique case-insensitive path → unique basename, with a one-line console note when it remaps). A LoRA that is genuinely missing **stops the run before anything is sampled**, and the error names the file — a picture that rendered without its LoRA is worse than no picture, above all for a script or an agent that would happily judge it. The error also says where the switch is. Gear menu → **Stop on missing LoRA** (on by default) turns that off for this node: a missing row then warns once in the console and is skipped while the rest of the stack applies, which suits a person at the canvas who wants to see the rest of the picture. In API graphs the same switch is the `on_missing` input (`"error"`, the default, or `"skip"`); graphs saved before the input existed get the default. ## Preview thumbnails Hovering a row's name — or any entry in the picker — floats a small preview beside the cursor when a sidecar image sits next to the LoRA file (same basename: `.preview.png` / `.preview.jpg` / `.preview.jpeg` / `.preview.webp`, or the bare `.png` / `.jpg` / `.jpeg` / `.webp`). No sidecar, no thumbnail — nothing appears and nothing shifts. The gear menu's **Preview thumbnails** switch turns the behavior off. ## Trigger words The info card gathers words from three places: the LoRA file's own metadata, a standard `.civitai.info` sidecar when one sits beside the file (other Civitai tools write these; the pack reads them and never goes online itself), and words you add yourself (remembered per LoRA across workflows). Click a word to toggle it into the row; enabled rows' selected words are joined into the `triggers` output, deduplicated, in row order. ## Outputs - **model**: The model with every enabled LoRA applied in row order. - **clip**: The CLIP with every enabled LoRA applied (unchanged when no CLIP is connected). - **triggers**: The selected trigger words from enabled rows, deduplicated, joined with the gear menu's separator (a comma by default). Rows with a strength of `0` (both model and CLIP) are skipped at load time but still contribute their trigger words, so you can park a LoRA at zero while comparing. A row skipped because its file is missing contributes none: that LoRA is not in the picture.