From 52a01e0adca4a34c130d6eff61d2be88178bc81f Mon Sep 17 00:00:00 2001
From: Towsty <ianjmcarr@gmail.com>
Date: Sun, 20 Sep 2026 18:13:07 -0500
Subject: [PATCH] Enable Qwen 2.1 on Edit: wire picker, runner, and
 source-still validation.

Co-authored-by: Cursor <cursoragent@cursor.com>
---
 components/studio2/Bench.vue   | 16 +++++--
 docs/qwen21.md                 | 10 ++--
 server/utils/studio2/runner.ts | 83 +++++++++++++++++++++++++---------
 shared/studio2/contracts.mjs   |  6 +--
 shared/studio2/input.mjs       | 10 +++-
 tests/studio2-bench.test.mjs   |  9 +++-
 6 files changed, 98 insertions(+), 36 deletions(-)

diff --git a/components/studio2/Bench.vue b/components/studio2/Bench.vue
index 895e2aa..ab62884 100644
--- a/components/studio2/Bench.vue
+++ b/components/studio2/Bench.vue
@@ -89,7 +89,8 @@
        <label class="inline-flex items-center gap-2"><input v-model="form.lockOutfit" type="checkbox" class="accent-[var(--amber)]">Lock outfit</label>
        <p class="text-[11px]">hero encode + lock text — not a MiniMax face/clothes split on stills</p>
       </div>
-      <p v-else class="text-[var(--type-sm)] text-[var(--muted)]">Qwen 2.1 is T2I — no hero, locks, or LoRA stack in this build.</p>
+      <p v-else-if="form.mode==='edit'" class="text-[var(--type-sm)] text-[var(--muted)]">Qwen edit uses source as &lt;image1&gt;, not Klein hero-ref. No locks or LoRA stack.</p>
+      <p v-else class="text-[var(--type-sm)] text-[var(--muted)]">Qwen 2.1 — no hero, locks, or LoRA stack.</p>
 
       <label>
        <span class="field-label">Prompt</span>
@@ -354,7 +355,7 @@ const videoDuration=ref(0)
 const focusedJobId=ref('')
 const video=computed(()=>['video','extend'].includes(form.mode))
 const kreaSticky=computed(()=>['edit','compose','iterate'].includes(form.mode))
-const qwenOnlyGenerate=computed(()=>form.mode!=='generate')
+const qwenAllowed=computed(()=>['generate','edit'].includes(form.mode))
 const engineOptions=computed(()=>{
  if(video.value){
   return [
@@ -365,7 +366,7 @@ const engineOptions=computed(()=>{
  return [
   {value:'flux',label:'Klein',disabled:false,hint:''},
   {value:'krea',label:'Krea 2',disabled:kreaSticky.value,hint:kreaSticky.value?'sticky hero = Klein':''},
-  {value:'qwen21',label:'Qwen 2.1',disabled:qwenOnlyGenerate.value,hint:qwenOnlyGenerate.value?'Qwen 2.1 is T2I in this build':''}
+  {value:'qwen21',label:'Qwen 2.1',disabled:!qwenAllowed.value,hint:qwenAllowed.value?'':'Generate and Edit only'}
  ]
 })
 watch(engineOptions,opts=>{
@@ -381,6 +382,15 @@ watch(()=>form.engine,(engine,prev)=>{
   form.settings.loraStack=[]
  }
 })
+watch(()=>[form.mode,form.engine],([mode,engine])=>{
+ if(engine==='qwen21' && mode==='edit'){
+  form.settings.steps=25
+  form.settings.cfg=1
+  form.lockFace=false
+  form.lockOutfit=false
+  form.settings.loraStack=[]
+ }
+})
 const stills=computed(()=>library.value.stills.filter((s:any)=>s.folderId===form.folderId))
 const clips=computed(()=>library.value.clips.filter((s:any)=>s.folderId===form.folderId))
 const stillUrl=(id:string)=>`/api/library/stills/${id}`
diff --git a/docs/qwen21.md b/docs/qwen21.md
index 70220f5..551d28f 100644
--- a/docs/qwen21.md
+++ b/docs/qwen21.md
@@ -34,9 +34,11 @@ Restart Comfy **only when idle** (`COMFY_CONTROL_URL/status` → `gpu.busy=false
 ## App
 
 - Engine key: `qwen21` · UI label: **Qwen 2.1**
-- Generate (T2I) only in this build. Edit / Compose / Iterate / Video disable with: “Qwen 2.1 is T2I in this build”.
-- Sampler defaults: euler / simple / cfg **1** / steps **25** · `ModelSamplingAuraFlow` shift **3.1**
-- Default canvas **1024×1024** (aspect 16:9 / 9:16 / 1:1 → long-edge square via `TextEncodeQwenImage21` resolution; multiples of 32). Drop to 768 if VRAM errors.
+- **Generate** (T2I) and **Edit** (same checkpoint, second graph). Compose / Iterate / Video / Extend / Music stay disabled unless a graph exists.
+- Edit requires Start still → `images.image_1`; prompt slots use `<image1>` / `<image2>`. If the user omits `<image1>`, the runner prepends `Keep the subject in <image1>.`
+- Face / outfit locks stay Klein semantics — they do not drive Qwen slots.
+- Sampler defaults: euler / simple / cfg **1** / steps **25**. Edit uses `QwenImage21Cache` (device auto, dtype int8). T2I uses `ModelSamplingAuraFlow` shift **3.1**.
+- Default canvas **1024×1024** (aspect 16:9 / 9:16 / 1:1 → long-edge ~1024 via `TextEncodeQwenImage21` resolution; multiples of 32). Edit with aspect `auto` follows `image_1` aspect at ~1024.
 - No hero / locks / Klein LoRA stack on this engine.
 
-Graph: `server/assets/studio2_qwen21_t2i.json`.
+Graphs: `server/assets/studio2_qwen21_t2i.json`, `server/assets/studio2_qwen21_edit.json`.
diff --git a/server/utils/studio2/runner.ts b/server/utils/studio2/runner.ts
index e1d05fa..57587e4 100644
--- a/server/utils/studio2/runner.ts
+++ b/server/utils/studio2/runner.ts
@@ -8,7 +8,8 @@ import { queueSeeds } from '~/shared/studio2/seed.mjs';
 import { readFileSync, writeFileSync, mkdirSync, existsSync, unlinkSync } from 'node:fs';
 import { join } from 'node:path';
 import template from '../../assets/studio2_minimax_native.json';
-import qwen21Template from '../../assets/studio2_qwen21_t2i.json';
+import qwen21Template from '../../assets/studio2_qwen21_t2i.json'
+import qwen21EditTemplate from '../../assets/studio2_qwen21_edit.json';
 import { nativeVideoGraph, attachHeroReference, applyResolvedImageSize } from '~/shared/studio2/graphs.mjs';
 import { compilePrompt, scopedFile } from '~/shared/studio2/contracts.mjs';
 import { resolveQwen21Size } from '~/shared/studio2/input.mjs';
@@ -39,7 +40,7 @@ const NEVER_FIT = new Set(['hero', 'start']);
 async function upload(r: any, name: string, data: Buffer): Promise<string> {
     const prefix = String(useRuntimeConfig().comfyFilenamePrefix).replace(/\/$/, '') + `/studio2/${r.id}/${r.index}`;
     let fitted = false;
-    if (!NEVER_FIT.has(name) && needsFit(data, r.request.settings.width, r.request.settings.height)) {
+    if (!NEVER_FIT.has(name) && r.request?.engine !== 'qwen21' && needsFit(data, r.request.settings.width, r.request.settings.height)) {
         data = await fitStill(data, r.request.settings.width, r.request.settings.height);
         fitted = true;
     }
@@ -129,26 +130,64 @@ async function prepareGraph(r: any) {
         }
     }
     else if (q.engine === 'qwen21') {
-        if (q.mode !== 'generate')
-            throw new Error('Qwen 2.1 is T2I in this build.');
-        const size = resolveQwen21Size(s.aspect);
-        s.width = size.width;
-        s.height = size.height;
-        graph = structuredClone(qwen21Template);
-        graph['9'].inputs.prompt = q.compiledPrompt;
-        graph['9'].inputs.negative_prompt = stylePrompt(q.imageStyles, true);
-        // TextEncodeQwenImage21 builds the 64-ch empty latent from resolution (square T2I).
-        graph['9'].inputs.resolution = Math.max(size.width, size.height);
-        graph['15'].inputs.seed = s.seed;
-        graph['15'].inputs.steps = s.steps || 25;
-        graph['15'].inputs.cfg = s.cfg ?? 1;
-        graph['15'].inputs.sampler_name = 'euler';
-        graph['15'].inputs.scheduler = 'simple';
-        graph['21'].inputs.filename_prefix = prefix + '/image';
-        r.sampleLatent = 'qwen21 textencode latent';
-        r.sampleDenoise = null;
-        r.heroReferenceAttached = false;
-        r.graphId = 'studio2_qwen21_t2i.json';
+        if (!['generate', 'edit'].includes(q.mode))
+            throw new Error('Qwen 2.1 supports Generate and Edit.')
+        if (q.mode === 'edit') {
+            if (!a) throw new Error('Choose the source still.')
+            let prompt = String(q.compiledPrompt || '')
+            if (!/<image1>/i.test(prompt))
+                prompt = `Keep the subject in <image1>. ${prompt}`.trim()
+            graph = structuredClone(qwen21EditTemplate)
+            graph['10'].inputs.image = a
+            if (b) {
+                graph['11'] = {
+                    inputs: { image: b },
+                    class_type: 'LoadImage',
+                    _meta: { title: 'image_2 optional' }
+                }
+                graph['9'].inputs['images.image_2'] = ['11', 0]
+            }
+            graph['9'].inputs.prompt = prompt
+            graph['9'].inputs.negative_prompt = stylePrompt(q.imageStyles, true)
+            graph['9'].inputs.resolution = 1024
+            if (s.aspect && s.aspect !== 'auto') {
+                graph['16'] = {
+                    inputs: { width: s.width, height: s.height, batch_size: 1 },
+                    class_type: 'EmptyLatentImage',
+                    _meta: { title: 'Qwen edit canvas' }
+                }
+                graph['15'].inputs.latent_image = ['16', 0]
+            }
+            graph['15'].inputs.seed = s.seed
+            graph['15'].inputs.steps = s.steps || 25
+            graph['15'].inputs.cfg = s.cfg ?? 1
+            graph['15'].inputs.sampler_name = 'euler'
+            graph['15'].inputs.scheduler = 'simple'
+            graph['21'].inputs.filename_prefix = prefix + '/image'
+            r.sampleLatent = s.aspect && s.aspect !== 'auto' ? 'qwen21 empty latent' : 'qwen21 edit latent'
+            r.sampleDenoise = null
+            r.heroReferenceAttached = false
+            r.graphId = 'studio2_qwen21_edit.json'
+        } else {
+            const size = resolveQwen21Size(s.aspect)
+            s.width = size.width
+            s.height = size.height
+            graph = structuredClone(qwen21Template)
+            graph['9'].inputs.prompt = q.compiledPrompt
+            graph['9'].inputs.negative_prompt = stylePrompt(q.imageStyles, true)
+            // TextEncodeQwenImage21 builds the 64-ch empty latent from resolution (square T2I).
+            graph['9'].inputs.resolution = Math.max(size.width, size.height)
+            graph['15'].inputs.seed = s.seed
+            graph['15'].inputs.steps = s.steps || 25
+            graph['15'].inputs.cfg = s.cfg ?? 1
+            graph['15'].inputs.sampler_name = 'euler'
+            graph['15'].inputs.scheduler = 'simple'
+            graph['21'].inputs.filename_prefix = prefix + '/image'
+            r.sampleLatent = 'qwen21 textencode latent'
+            r.sampleDenoise = null
+            r.heroReferenceAttached = false
+            r.graphId = 'studio2_qwen21_t2i.json'
+        }
     }
     else {
         const mode = identityPlan?.mode ?? (q.mode === 'iterate' ? (a ? 'edit' : 'generate') : q.mode);
diff --git a/shared/studio2/contracts.mjs b/shared/studio2/contracts.mjs
index 40ffeb1..b8e9544 100644
--- a/shared/studio2/contracts.mjs
+++ b/shared/studio2/contracts.mjs
@@ -34,14 +34,14 @@ export function validateRequest(raw, xaigen = false) {
   if (!(video ? ['minimax','ltx'] : ['flux','krea','qwen21']).includes(engine)) fail('Engine does not support this task.')
   if (engine === 'ltx' && !xaigen) fail('LTX is available only in xAIGen.')
   if (engine === 'krea' && ['edit','compose','iterate'].includes(mode)) fail('Sticky hero edits require Klein. Krea remains available for Generate and masked Refine; no engine fallback is performed.')
-  if (engine === 'qwen21' && mode !== 'generate') fail('Qwen 2.1 is T2I in this build.')
+  if (engine === 'qwen21' && !['generate', 'edit'].includes(mode)) fail('Qwen 2.1 supports Generate and Edit.')
   if (Array.isArray(raw.batch) && raw.batch.length>30) fail('A batch supports up to 31 shots including the first prompt.')
   if (Array.isArray(raw.guides) && raw.guides.length>8) fail('Use up to eight guide stills.')
   if (!raw.folderId) fail('Choose a folder.')
-  if ((video || ['edit','compose','iterate'].includes(mode)) && (raw.lockFace !== false || raw.lockOutfit !== false) && !raw.identityStillId && (video || !raw.imageAId)) fail('Choose a sticky hero still.')
+  if (['edit','compose','refine'].includes(mode) && !raw.imageAId && !(mode==='edit' && engine==='flux' && raw.identityStillId)) fail('Choose the source still.')
+  if ((video || ['edit','compose','iterate'].includes(mode)) && engine !== 'qwen21' && (raw.lockFace !== false || raw.lockOutfit !== false) && !raw.identityStillId && (video || !raw.imageAId)) fail('Choose a sticky hero still.')
   if (mode === 'compose' && !raw.imageBId) fail('Compose requires a second still.')
   if (mode === 'refine' && !raw.maskId) fail('Refine requires a mask.')
-  if (['edit','compose','refine'].includes(mode) && !raw.imageAId && !(mode==='edit' && engine==='flux' && raw.identityStillId)) fail('Choose the source still.')
   if (mode === 'extend' && !raw.startClipId) fail('Select the clip in the family strip or Previous clip.')
   const allSections = sectionsFrom(raw.promptSections || {action:raw.prompt || ''})
   const promptSections = video ? allSections : {action:allSections.action}
diff --git a/shared/studio2/input.mjs b/shared/studio2/input.mjs
index e1321f6..60c3381 100644
--- a/shared/studio2/input.mjs
+++ b/shared/studio2/input.mjs
@@ -1,13 +1,19 @@
 import {restoreLoras,loraUse} from './lora-memory.mjs'
 import { restoreInput } from './contracts.mjs'
 export const aspectOptions = [{id:'auto',label:'Automatic',hint:'Match start or hero still'},{id:'16:9',label:'16:9',hint:'1344 × 768 · padded'},{id:'9:16',label:'9:16',hint:'768 × 1344 · padded'},{id:'1:1',label:'1:1',hint:'960 × 960 · padded'}]
-export function resolveQwen21Size(aspect) {
+export function resolveQwen21Size(aspect, source) {
   const presets = {'16:9':[1024,576],'9:16':[576,1024],'1:1':[1024,1024]}
   if (presets[aspect]) return {width:presets[aspect][0],height:presets[aspect][1]}
+  if (source?.width && source?.height) {
+    const ratio = source.width / source.height
+    const width = Math.max(32, Math.round(Math.sqrt(1024 * 1024 * ratio) / 32) * 32)
+    const height = Math.max(32, Math.round(Math.sqrt(1024 * 1024 / ratio) / 32) * 32)
+    return {width, height}
+  }
   return {width:1024,height:1024}
 }
 export function resolveSize(aspect, source, fallback = {width:1344,height:768}, engine = '') {
-  if (engine === 'qwen21') return resolveQwen21Size(aspect)
+  if (engine === 'qwen21') return resolveQwen21Size(aspect, source)
   const presets = {'16:9':[1344,768],'9:16':[768,1344],'1:1':[960,960]}
   if (presets[aspect]) return {width:presets[aspect][0],height:presets[aspect][1]}
   if (!source?.width || !source?.height) return {width:fallback.width || 1344,height:fallback.height || 768}
diff --git a/tests/studio2-bench.test.mjs b/tests/studio2-bench.test.mjs
index a99f9eb..098c4db 100644
--- a/tests/studio2-bench.test.mjs
+++ b/tests/studio2-bench.test.mjs
@@ -74,10 +74,15 @@ test('image chips compile positive and negative independently; video chips do no
  assert.doesNotMatch(compilePrompt(q.promptSections,false,q),/Picture 1/)
  assert.deepEqual(parseBatch('A wrapped\nsentence\n\nSecond shot\r\n\r\nThird'),['A wrapped\nsentence','Second shot','Third'])
 })
-test('qwen21 is generate-only T2I with 25/1 defaults and no locks',()=>{
+test('qwen21 supports Generate and Edit with 25/1 defaults and no locks',()=>{
  const q=validateRequest({mode:'generate',engine:'qwen21',folderId:'f',promptSections:{action:'a red bicycle'},settings:{aspect:'1:1'}})
  assert.equal(q.engine,'qwen21');assert.equal(q.settings.steps,25);assert.equal(q.settings.cfg,1)
  assert.equal(q.settings.width,1024);assert.equal(q.settings.height,1024)
  assert.equal(q.lockFace,false);assert.equal(q.lockOutfit,false);assert.deepEqual(q.settings.loraStack,[])
- assert.throws(()=>validateRequest({mode:'edit',engine:'qwen21',folderId:'f',imageAId:'s',promptSections:{action:'x'}}),/T2I in this build/)
+ const edit=validateRequest({mode:'edit',engine:'qwen21',folderId:'f',imageAId:'s',promptSections:{action:'change the sky'},settings:{aspect:'auto'}})
+ assert.equal(edit.mode,'edit');assert.equal(edit.settings.steps,25);assert.equal(edit.settings.cfg,1)
+ assert.equal(edit.lockFace,false);assert.deepEqual(edit.settings.loraStack,[])
+ assert.throws(()=>validateRequest({mode:'edit',engine:'qwen21',folderId:'f',promptSections:{action:'x'}}),/source still/)
+ assert.throws(()=>validateRequest({mode:'compose',engine:'qwen21',folderId:'f',imageAId:'s',imageBId:'b',promptSections:{action:'x'}}),/Generate and Edit/)
+ assert.throws(()=>validateRequest({mode:'iterate',engine:'qwen21',folderId:'f',imageAId:'s',promptSections:{action:'x'}}),/Generate and Edit/)
 })