A picture kept as a combo keeps its nodes, every value as it was

Keep -> Its combo now saves the nodes upstream of the picture's samplers' model and prompt
inputs (loaders, LoRA chains, CLIP skip, both prompts) as a node-structure combo, values
untouched; models are named by where they are here, missing ones kept by the picture's name.
A workflow without a sampler still gives the simple combo.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

ARCHITECTURE.md

@@ -527,6 +527,11 @@ covered by the `styles.css` manifest.
   named by `comboTitle` — first LoRA or main model · two telling tags — and renamed by `renameCombo`;
   its prompts as a saved prompt, reading what was kept) without DOM. A new combo's toast is
   `ui_kept_toast.js` (the image's thumbnail, the name in a box to rename it on the spot, Open).
+  An image becomes a node-structure combo: `sampler_inputs.js` (no DOM) picks the nodes upstream
+  of its samplers' model / positive / negative / conditioning inputs (not through images,
+  latents or masks), and `image_keep_nodes.js` saves them with `captureStructure`, every value
+  as the image had it, models renamed to where they are on this computer; a workflow with no
+  sampler gives the older main model + LoRAs + prompt combo.
 - `ui_combos.js` renders the Combos page's list (搭配: a main model, LoRAs and a prompt, formerly Prompt Notes): search,
   ? (what a combo is), New, and a card per combo with its model's cover (one kept from an output image shows that image, the model's cover as a badge) and Put on canvas; a card opens the combo's editor, and dragged onto the
   canvas (`material_drag.js`) becomes a new group of nodes where it is dropped. Combos keep the notes' files.

CHANGELOG.md

@@ -22,6 +22,11 @@
 ### ★ See only the kept pictures (图库只看收藏)
 - **★ Kept** beside the gallery's search lists only the pictures you kept as a workflow, combo or prompt (the ★ on their cards). Works with the search and with Pictures / Videos. Pointing at a card's ★ says what it was kept as.
 
+### 📌 A picture kept as a combo keeps its nodes as they were (图片存搭配，节点原样记住)
+- **Keep → Its combo** now keeps the nodes that made the picture's model and prompts: the model and CLIP loaders, the whole LoRA chain (Power Lora Loader and other packs' nodes too), CLIP skip and model patches, the positive **and** negative prompt nodes, with every value as the picture had it. The sampler, size and decoding are left out, so Put on canvas gives a group you wire to your own sampler, like a combo saved from the canvas.
+- It used to keep only the main model, the LoRAs' names and the positive prompt: LoRA strengths came back as 1.0, and the negative prompt and CLIP skip were lost. **A combo only remembers; it never changes a value** (now one of the plugin's rules). Combos kept before stay as they are.
+- Models are named by where they are on this computer; one this computer lacks is kept by the picture's name and said after keeping. A workflow without a sampler to read from still gives the simple kind.
+
 ### ⭐ Combos kept from a picture are easy to tell apart (图片存的搭配一眼认得出)
 - **The card shows the picture you kept it from**, with the main model's cover as a small round badge in its corner, instead of the model's cover alone (several combos of one model used to look the same).
 - **A name that says which one it is**: the first LoRA (or the main model) and the prompt's first two telling tags, like *Ahri · fox ears, kimono*, skipping words nearly every picture has (*1girl*, *solo*, *looking at viewer*…).

docs/architecture/material-library.md

@@ -116,8 +116,11 @@ override, which travels into the saved record but never writes back to a Recipe.
 
 The gallery's star and the image workbench's Keep (`ui_keep_menu.js`) keep an output PNG
 as one of three things, each recording the image as its source: the whole workflow as a
-Workflow Recipe (`recipe_save.js`, `source_image`), its main model, LoRAs and positive
-prompt as a combo (`image_keep.js`; models are looked up among this computer's files with
+Workflow Recipe (`recipe_save.js`, `source_image`), the nodes that make its sampler's model
+and prompts as a node-structure combo (`image_keep_nodes.js`, `sampler_inputs.js`: loaders,
+LoRA chains, CLIP skip, model patches and both prompt nodes, every value as it was; a combo
+never changes a value) — or, for a workflow with no sampler, its main model, LoRAs and
+positive prompt as a combo (`image_keep.js`; models are looked up among this computer's files with
 `/anomalous/resolve_paths_to_previews`, missing ones are left out; the notebook file keeps
 `source_image`, which the Combos card and the editor's title show; the name is the first
 LoRA's (else the main model's) · the prompt's first two tags that are neither quality words

tests/sampler_inputs.mjs

@@ -0,0 +1,43 @@
+// What a combo kept from an output image holds (web/modules/sampler_inputs.js).
+import assert from 'node:assert/strict';
+import { isSampler, samplerSideNodes } from '../web/modules/sampler_inputs.js';
+
+// A small graph: links [id, from, to, type]; inputs name the link feeding them.
+function graph(nodes, links) {
+    const byId = new Map(links.map(([id, from, , type]) => [id, { id, origin_id: from, type }]));
+    return {
+        _nodes: nodes.map(([id, type, inputs = {}]) => ({ id, type, inputs: Object.entries(inputs).map(([name, link]) => ({ name, link })) })),
+        getLink: id => byId.get(id) ?? null,
+    };
+}
+
+const g = graph([
+    [1, 'CheckpointLoaderSimple'],
+    [2, 'LoraLoader', { model: 1, clip: 2 }],
+    [3, 'CLIPSetLastLayer', { clip: 3 }],
+    [4, 'CLIPTextEncode', { clip: 4 }],
+    [5, 'CLIPTextEncode', { clip: 5 }],
+    [6, 'EmptyLatentImage'],
+    [7, 'KSampler', { model: 6, positive: 7, negative: 8, latent_image: 9 }],
+    [8, 'VAEDecode', { samples: 10, vae: 11 }],
+    [9, 'LoadImage'],
+    [10, 'ControlNetApply', { conditioning: 12, image: 13 }],
+    [11, 'Note'],
+], [
+    [1, 1, 2, 'MODEL'], [2, 1, 2, 'CLIP'], [3, 2, 3, 'CLIP'], [4, 3, 4, 'CLIP'], [5, 3, 5, 'CLIP'],
+    [6, 2, 7, 'MODEL'], [7, 10, 7, 'CONDITIONING'], [8, 5, 7, 'CONDITIONING'], [9, 6, 7, 'LATENT'],
+    [10, 7, 8, 'LATENT'], [11, 1, 8, 'VAE'], [12, 4, 10, 'CONDITIONING'], [13, 9, 10, 'IMAGE'],
+]);
+
+assert.equal(isSampler(g._nodes[6]), true);
+assert.equal(isSampler(g._nodes[1]), false);
+// Loaders, the LoRA, CLIP skip, both prompts and what sits between them and the sampler;
+// not the sampler, the latent, the decode, the image loader or a note.
+assert.deepEqual(samplerSideNodes(g).map(node => node.id), [1, 2, 3, 4, 5, 10]);
+// A workflow without a sampler: nothing to keep.
+assert.deepEqual(samplerSideNodes(graph([[1, 'CheckpointLoaderSimple']], [])), []);
+// A guider (Flux): model and conditioning.
+const flux = graph([[1, 'UNETLoader'], [2, 'CLIPTextEncode'], [3, 'BasicGuider', { model: 1, conditioning: 2 }]],
+    [[1, 1, 3, 'MODEL'], [2, 2, 3, 'CONDITIONING']]);
+assert.deepEqual(samplerSideNodes(flux).map(node => node.id), [1, 2]);
+console.log('sampler inputs: ok');

web/modules/image_keep.js

@@ -1,8 +1,10 @@
 /**
  * Keeping an output image (the gallery's star, the image workbench) as one of three things:
- * its whole workflow as a Workflow Recipe (recipe_save.js), its main model, LoRAs and
- * positive prompt as a combo (搭配, a file of GET /anomalous/notebooks), or its prompts as a
- * saved prompt (POST /anomalous/save_prompt_plan). The image's workflow is laid on a
+ * its whole workflow as a Workflow Recipe (recipe_save.js), the nodes that make its sampler's
+ * model and prompts as a combo (搭配, a file of GET /anomalous/notebooks; image_keep_nodes.js,
+ * every value as the image had it; a workflow without a sampler to read from gives its main
+ * model, LoRAs and positive prompt), or its prompts as a saved prompt
+ * (POST /anomalous/save_prompt_plan). The image's workflow is laid on a
  * canvas of its own and read like a recipe saved from it (recipe_parser.js); its models are
  * looked up among this computer's files. Each thing kept remembers its image, so
  * GET /anomalous/kept_images can say what an image was kept as. No DOM.
@@ -40,9 +42,10 @@ const fileStem = name => String(name || '').split(/[\\/]/).pop().replace(/\.[^.]
 /** What the image's workflow uses (extractRecipeMetadata) and a name: its prompt's first tags, else the file's. */
 async function readImage(sourceImage) {
     const info = await jsonResponse(await post('/anomalous/inspect_image_material', { source_image: sourceImage }), 'read image workflow');
-    const metadata = extractRecipeMetadata(detachedGraph(info.workflow));
+    const graph = detachedGraph(info.workflow);
+    const metadata = extractRecipeMetadata(graph);
     const name = promptTitle(metadata.promptPositive[0]) || fileStem(sourceImage?.filename);
-    return { name, metadata };
+    return { name, metadata, graph };
 }
 
 /** A model's name as its card shows it: the name the user gave it, else the file's. */
@@ -85,16 +88,41 @@ export function freeComboName(name, combos) {
     }
 }
 
+/** Saves combo `data` under a free name from `wanted`. Returns { filename, name }. */
+async function saveCombo(data, wanted, combos) {
+    const comboName = freeComboName(wanted, combos);
+    const filename = `${comboName}.json`;
+    const response = await post('/anomalous/save_notebook', { filename, name: comboName, data: { ...data, name: comboName } });
+    if ((await jsonResponse(response, 'save combo')).status !== 'success') throw new Error('combo save failed');
+    return { filename, name: comboName };
+}
+
 /**
- * Keeps the image's main model, LoRAs and positive prompt as a combo; an image kept before
- * gives that combo back (`existed`). Models this computer does not have are left out and
- * named in `missing`. Returns { filename, name, existed, missing }.
+ * Keeps the image as a combo of the nodes that make its sampler's model and prompts, every
+ * value as it was; an image kept before gives that combo back (`existed`). Models this
+ * computer does not have are named in `missing` (kept by the workflow's name, `missingKept`). A workflow
+ * without a sampler to read from gives a combo of its main model, LoRAs and positive prompt.
+ * Returns { filename, name, existed, missing }.
  */
 export async function keepImageAsCombo(sourceImage, { name = '' } = {}) {
     const combos = await listCombos();
     const kept = combos.find(note => sameImage(note.data?.source_image, sourceImage));
     if (kept) return { filename: kept.filename, name: kept.name, existed: true, missing: [] };
     const image = await readImage(sourceImage);
+    const prompt = image.metadata.promptPositive.join('\n');
+    // Loaded only here: it reaches the canvas modules (no DOM in this file, and tests import it).
+    const { keepGraphNodes } = await import('./image_keep_nodes.js');
+    const nodes = await keepGraphNodes(image.graph, localModels);
+    if (nodes) {
+        const title = comboTitle({ prompt, mainModel: nodes.mainModel, loras: nodes.loras, fallback: image.name });
+        const saved = await saveCombo({ ...nodes.data, source_image: sourceImage }, name || title, combos);
+        return { ...saved, existed: false, missing: nodes.missing, missingKept: true };
+    }
+    return keepSimpleCombo(sourceImage, image, prompt, name, combos);
+}
+
+/** The older kind of combo: the image's main model, LoRAs and positive prompt. */
+async function keepSimpleCombo(sourceImage, image, prompt, name, combos) {
     const main = image.metadata.baseModel || '';
     const loraPaths = [...new Set((image.metadata.loras || []).map(lora => lora.name).filter(Boolean))];
     const [mains, loras] = await Promise.all([
@@ -103,22 +131,16 @@ export async function keepImageAsCombo(sourceImage, { name = '' } = {}) {
     ]);
     const mainModel = mains[main] || null;
     const localLoras = loraPaths.filter(path => loras[path]).map(path => loras[path]);
-    const prompt = image.metadata.promptPositive.join('\n');
-    const comboName = freeComboName(name || comboTitle({ prompt, mainModel, loras: localLoras, fallback: image.name }), combos);
-    const data = {
-        name: comboName,
+    const saved = await saveCombo({
         baseModel: mainModel ? String(mainModel.metadata?.baseModel || '').trim() || UNLABELED_BASE_MODEL : '',
         mainModel,
         loras: localLoras,
         promptEn: prompt,
         promptZh: '',
         source_image: sourceImage,
-    };
-    const filename = `${comboName}.json`;
-    const response = await post('/anomalous/save_notebook', { filename, name: comboName, data });
-    if ((await jsonResponse(response, 'save combo')).status !== 'success') throw new Error('combo save failed');
+    }, name || comboTitle({ prompt, mainModel, loras: localLoras, fallback: image.name }), combos);
     const missing = [...(main && !mainModel ? [main] : []), ...loraPaths.filter(path => !loras[path])];
-    return { filename, name: comboName, existed: false, missing };
+    return { ...saved, existed: false, missing };
 }
 
 /** Gives the combo kept in `filename` the name `name` (its file keeps its name). Returns the name. */

web/modules/image_keep_nodes.js

@@ -0,0 +1,55 @@
+/**
+ * An output image kept as a combo of its own nodes (image_keep.js keepImageAsCombo): the part
+ * of its workflow that makes the sampler's model and prompts (sampler_inputs.js), laid on a
+ * graph of its own, saved like nodes picked on the canvas (combo_structure.js) with every value
+ * as the image had it: LoRA strengths, CLIP skip, the negative prompt. A combo never changes a
+ * value (AGENTS.md, 功能原则). Models are looked up among this computer's files: one found
+ * there is named by where it is here; one not found keeps the workflow's name and is reported.
+ */
+
+import { captureStructure } from './combo_structure.js';
+import { NONE, STRUCTURED } from './combo_slots.js';
+import { slotLabel } from './ui_combo_structure.js';
+import { samplerSideNodes } from './sampler_inputs.js';
+
+const MAIN_FOLDERS = new Set(['checkpoints', 'unet', 'diffusion_models']);
+const relOf = model => [String(model.subfolder || '').replace(/^\/+|\/+$/g, ''), model.filename].filter(Boolean).join('/');
+
+/**
+ * The combo data of `graph`'s sampler side, or null when the workflow has no sampler to read
+ * from. `localModels(paths, folders)` gives { path → this computer's model }. Returns { data:
+ * { kind, structure, values, covers }, mainModel, loras, missing }.
+ */
+export async function keepGraphNodes(graph, localModels) {
+    const nodes = samplerSideNodes(graph);
+    if (!nodes.length) return null;
+    const { structure, values } = await captureStructure(nodes, { labelFor: slotLabel });
+    if (!structure) return null;
+    // An empty row of a stack node is no slot (as when saving from the canvas, unticked).
+    const slots = structure.slots.filter(slot => !(slot.optional && NONE.test(String(values[slot.id] ?? ''))));
+    const kept = Object.fromEntries(slots.map(slot => [slot.id, values[slot.id]]));
+    const covers = {};
+    const missing = [];
+    let mainModel = null;
+    const loras = [];
+    const models = slots.filter(slot => slot.kind === 'model' && !NONE.test(String(kept[slot.id] ?? '')));
+    const folders = [...new Set(models.map(slot => slot.folder).filter(Boolean))];
+    const found = await localModels([...new Set(models.map(slot => kept[slot.id]))], folders);
+    for (const slot of models) {
+        const model = found[kept[slot.id]];
+        if (!model) {
+            missing.push(kept[slot.id]);
+            continue;
+        }
+        kept[slot.id] = relOf(model);
+        if (model.preview_url) covers[slot.id] = model.preview_url;
+        if (!mainModel && MAIN_FOLDERS.has(slot.folder)) mainModel = model;
+        if (slot.folder === 'loras') loras.push(model);
+    }
+    return {
+        data: { kind: STRUCTURED, structure: { ...structure, slots }, values: kept, covers },
+        mainModel,
+        loras,
+        missing,
+    };
+}

web/modules/locales.js

@@ -820,7 +820,7 @@ export const i18n = {
         keepAs_recipe: "整个工作流",
         keepAs_recipeHint: "存成工作流配方，在“工作流”页，可以整个打开",
         keepAs_combo: "模型搭配",
-        keepAs_comboHint: "主模型 + LoRA + 正向提示词，在“搭配”页一键放到画布",
+        keepAs_comboHint: "出这张图的模型、LoRA 和正负提示词节点，数值原样记下，在“搭配”页一键放到画布",
         keepAs_prompt: "只要提示词",
         keepAs_promptHint: "正向和负向提示词，存进提示词库",
         keepKeptAs: "已存：{name}，点这里打开",
@@ -852,6 +852,7 @@ export const i18n = {
         keptAs_prompt: "已存下提示词：{name}",
         keepExisted: "之前就存过了：{name}",
         keepComboMissing: "有 {count} 个模型本机没有，搭配里先空着。",
+        keepComboMissingNamed: "有 {count} 个模型这台电脑没有，先按原图的名字记着；在搭配里点它换一个本机的模型就行。",
         keepOpen: "打开",
         keepRemove: "取消收藏",
         keepRemoveHint: "把存下的这一项移到系统回收站",
@@ -3632,7 +3633,7 @@ export const i18n = {
         keepAs_recipe: "The whole workflow",
         keepAs_recipeHint: "As a Workflow Recipe on the Workflows page, to open whole",
         keepAs_combo: "Its combo",
-        keepAs_comboHint: "Main model, LoRAs and positive prompt, on the Combos page",
+        keepAs_comboHint: "The nodes that made its model and prompts, every value as it was, on the Combos page",
         keepAs_prompt: "Just the prompts",
         keepAs_promptHint: "The positive and negative prompts, in your saved prompts",
         keepKeptAs: "Kept as {name}; open it",
@@ -3664,6 +3665,7 @@ export const i18n = {
         keptAs_prompt: "Kept the prompts: {name}",
         keepExisted: "Kept before: {name}",
         keepComboMissing: "{count} of its models are not on this computer, so the combo leaves them out.",
+        keepComboMissingNamed: "{count} of its models are not on this computer; the combo keeps them by the picture's names; click one in the combo to choose a model of yours.",
         keepOpen: "Open",
         keepRemove: "Remove",
         keepRemoveHint: "Move what was kept to the Recycle Bin",

web/modules/sampler_inputs.js

@@ -0,0 +1,46 @@
+/**
+ * The part of a workflow that makes a sampler's model and prompts (what a combo kept from an
+ * output image holds): every node upstream of a sampler's model / positive / negative /
+ * conditioning inputs — model and CLIP loaders, LoRA chains, CLIP skip and model patches, the
+ * prompt nodes — but not the sampler itself, nor anything reached only through an image,
+ * latent or mask (image loaders, sizes, VAE decoding). A sampler is a node with a model input
+ * and a positive or conditioning input (KSampler, KSampler Advanced, CFG / Basic guiders).
+ * No DOM; `graph` is a LiteGraph graph (or anything with nodes and getLink / links).
+ */
+
+const FEEDS = new Set(['model', 'positive', 'negative', 'conditioning']);
+const CUT = new Set(['IMAGE', 'LATENT', 'MASK']);
+
+const inputNames = node => (node.inputs || []).map(input => input?.name);
+
+/** Whether `node` is a sampler (or a guider) whose model and prompts a combo keeps. */
+export function isSampler(node) {
+    const names = inputNames(node);
+    return names.includes('model') && (names.includes('positive') || names.includes('conditioning'));
+}
+
+/** The nodes upstream of the samplers' model and prompt inputs, in the graph's order. */
+export function samplerSideNodes(graph) {
+    const nodes = graph?._nodes || graph?.nodes || [];
+    const byId = new Map(nodes.map(node => [node.id, node]));
+    const linkOf = id => graph.getLink?.(id) ?? graph.links?.[id] ?? null;
+    const samplers = new Set(nodes.filter(isSampler).map(node => node.id));
+    const kept = new Set();
+    const waiting = [];
+    const follow = (input) => {
+        if (input?.link == null) return;
+        const link = linkOf(input.link);
+        if (!link || CUT.has(String(link.type ?? input.type ?? '').toUpperCase())) return;
+        waiting.push(link.origin_id);
+    };
+    for (const id of samplers) {
+        for (const input of byId.get(id).inputs || []) if (FEEDS.has(input?.name)) follow(input);
+    }
+    while (waiting.length) {
+        const id = waiting.pop();
+        if (kept.has(id) || samplers.has(id) || !byId.has(id)) continue;
+        kept.add(id);
+        for (const input of byId.get(id).inputs || []) follow(input);
+    }
+    return nodes.filter(node => kept.has(node.id));
+}

web/modules/ui_combos.js

@@ -91,7 +91,7 @@ function comboCard(owner, note) {
     });
     use.title = t('comboUseHint');
     body.append(use);
-    card.append(structured ? cover({ preview_url: structured.cover }) : cover(data.mainModel, data.source_image), body);
+    card.append(cover(structured ? { preview_url: structured.cover } : data.mainModel, data.source_image), body);
     bindMaterialDrag(card, owner, {
         payload: () => ({ type: 'combo', filename: note.filename, dragHint: t('comboDragHint'), dragTargetHint: t('comboDragCanvas') }),
         accepts: () => false,

web/modules/ui_keep_menu.js

@@ -41,10 +41,13 @@ function place(menu, anchor) {
     menu.style.top = `${below + height > window.innerHeight - 8 ? Math.max(8, box.top - height - 6) : below}px`;
 }
 
+/** The models a new combo names that this computer lacks: kept by the picture's name (nodes), or left out. */
+const missingWords = result => t(result.missingKept ? 'keepComboMissingNamed' : 'keepComboMissing', { count: result.missing.length });
+
 function keptMessage(spec, result) {
     if (result.existed) return t('keepExisted', { name: result.name });
     const message = t(`keptAs_${spec.kind}`, { name: result.name });
-    return result.missing?.length ? `${message} ${t('keepComboMissing', { count: result.missing.length })}` : message;
+    return result.missing?.length ? `${message} ${missingWords(result)}` : message;
 }
 
 async function keep(owner, spec, sourceImage, kept, { name, onKept, beforeOpen }) {
@@ -62,7 +65,7 @@ async function keep(owner, spec, sourceImage, kept, { name, onKept, beforeOpen }
         };
         // A new combo's name is a guess (image_keep.js comboTitle): it can be renamed right here.
         if (spec.kind === 'combo' && !result.existed) {
-            const missing = result.missing?.length ? t('keepComboMissing', { count: result.missing.length }) : '';
+            const missing = result.missing?.length ? missingWords(result) : '';
             showKeptComboToast(sourceImage, kept.combo, {
                 message: [t('keptAs_comboShort'), missing].filter(Boolean).join(' '),
                 onRenamed: (name) => {

web/modules/ui_quick_search.js

@@ -37,7 +37,8 @@ function comboSearchItem(note) {
     const data = note.data || {};
     if (data.kind === STRUCTURED) {
         const meta = structuredComboMeta(data);
-        return comboItem(note, { detail: meta.meta, thumb: meta.cover, text: meta.prompt });
+        const thumb = data.source_image?.filename ? outputThumbnail(data.source_image) : meta.cover;
+        return comboItem(note, { detail: meta.meta, thumb, text: meta.prompt });
     }
     const loras = (data.loras || []).length;
     const detail = [modelName(data.mainModel), loras ? t('comboLoraCount', { count: loras }) : ''].filter(Boolean).join(' · ');