Add Image v2 Generate for Klein text-to-image.

New sibling graph and mode so T2I does not borrow Edit, Compose, or Refine. Stills are ignored; size is width x height only.

Co-authored-by: Cursor <cursoragent@cursor.com>

server/api/v2/generate.post.ts

@@ -13,6 +13,7 @@ import {
   IMAGE_V2_TURBO_STEPS,
   IMAGE_V2_DENOISE_DEFAULT,
   clampImageV2Denoise,
+  clampImageV2Size,
   clampImageV2Strength,
   parseImageV2Mode,
   parseImageV2Task,
@@ -109,9 +110,12 @@ export default defineEventHandler(async (event) => {
 
   const mode = parseImageV2Mode(fields.mode)
   if (!mode) {
-    throw createError({ statusCode: 400, statusMessage: 'mode must be edit, compose, or refine' })
+    throw createError({ statusCode: 400, statusMessage: 'mode must be edit, compose, refine, or generate' })
   }
-  const task = parseImageV2Task(fields.task, mode === 'refine' ? 'refine' : 'scene')
+  const task = parseImageV2Task(
+    fields.task,
+    mode === 'generate' ? 't2i' : mode === 'refine' ? 'refine' : 'scene'
+  )
   const prompt = String(fields.prompt || '').trim()
   if (!prompt) {
     throw createError({ statusCode: 400, statusMessage: 'A prompt is required' })
@@ -131,11 +135,11 @@ export default defineEventHandler(async (event) => {
   }
 
   const ownerKey = libraryOwnerKey(event)
-  const imageA = await resolveImageRef(ownerKey, fields.image_a, uploadedA)
-  const imageB = mode === 'refine' ? null : await resolveImageRef(ownerKey, fields.image_b, uploadedB)
+  const imageA = mode === 'generate' ? null : await resolveImageRef(ownerKey, fields.image_a, uploadedA)
+  const imageB = mode === 'refine' || mode === 'generate' ? null : await resolveImageRef(ownerKey, fields.image_b, uploadedB)
   const mask = mode === 'refine' ? await resolveImageRef(ownerKey, fields.mask, uploadedMask) : null
 
-  if (!imageA) {
+  if (mode !== 'generate' && !imageA) {
     throw createError({ statusCode: 400, statusMessage: 'image_a is required' })
   }
   if (mode === 'refine' && !mask) {
@@ -173,25 +177,34 @@ export default defineEventHandler(async (event) => {
   const turbo = parseBool(fields.turbo)
   const steps = turbo ? IMAGE_V2_TURBO_STEPS : clampImageSteps(fields.steps, IMAGE_V2_STEPS_DEFAULT)
   const cfg = turbo ? IMAGE_V2_TURBO_CFG : clampImageCfg(fields.cfg, IMAGE_V2_CFG_DEFAULT)
-  const megapixels = clampImageScaleMegapixels(fields.megapixels ?? fields.scaleMegapixels, 1)
+  const megapixels = mode === 'generate' ? 0 : clampImageScaleMegapixels(fields.megapixels ?? fields.scaleMegapixels, 1)
   const seed = fields.seed && String(fields.seed) !== 'random'
     ? Number(fields.seed)
     : Math.floor(Math.random() * 2_147_483_647)
-  const size = imageDimensions(imageA.data)
+  const generateSize = clampImageV2Size(fields.width, fields.height, fields.aspect)
+  const size = mode === 'generate'
+    ? { width: generateSize.width, height: generateSize.height }
+    : imageDimensions(imageA!.data)
   const clipName = String(fields.name || '').trim().slice(0, 80)
   const v2Mode = mode as ImageV2Mode
-  const v2Task = (mode === 'refine' ? 'refine' : mode === 'edit' ? 'scene' : task) as ImageV2Task
+  const v2Task = (
+    mode === 'generate' ? 't2i' : mode === 'refine' ? 'refine' : mode === 'edit' ? 'scene' : task
+  ) as ImageV2Task
   const refineStrength = mode === 'refine' ? clampImageV2Denoise(fields.strength, IMAGE_V2_DENOISE_DEFAULT) : undefined
+  const generateConsistencyModel = clampImageV2Strength(fields.consistency_model, mode === 'generate' ? 0 : IMAGE_V2_CONSISTENCY_MODEL)
+  const generateConsistencyClip = clampImageV2Strength(fields.consistency_clip, mode === 'generate' ? 0 : IMAGE_V2_CONSISTENCY_CLIP)
 
-  const still = await rememberInputStill({
-    ownerKey,
-    folderId,
-    filename: imageA.filename,
-    data: imageA.data,
-    width: size?.width || 0,
-    height: size?.height || 0,
-    hideInput
-  })
+  const still = imageA
+    ? await rememberInputStill({
+      ownerKey,
+      folderId,
+      filename: imageA.filename,
+      data: imageA.data,
+      width: size?.width || 0,
+      height: size?.height || 0,
+      hideInput
+    })
+    : null
   const savedRef = imageB
     ? await rememberInputStill({
       ownerKey,
@@ -219,7 +232,7 @@ export default defineEventHandler(async (event) => {
       prompt,
       name: clipName,
       folderId,
-      aspect: 'auto',
+      aspect: mode === 'generate' ? (generateSize.aspect || '1:1') : 'auto',
       width: size?.width || 0,
       height: size?.height || 0,
       steps,
@@ -244,15 +257,15 @@ export default defineEventHandler(async (event) => {
       negative: String(fields.negative || '').trim(),
       referenceStillId: savedRef?.id,
       referenceStillFilename: savedRef?.filename,
-      scaleToTotalPixels: true,
+      scaleToTotalPixels: mode !== 'generate',
       scaleMegapixels: megapixels,
       imagePipeline: 'v2',
       v2Mode,
       v2Task,
       snofsModel: clampImageV2Strength(fields.snofs_model, IMAGE_V2_SNOFS_MODEL),
       snofsClip: clampImageV2Strength(fields.snofs_clip, IMAGE_V2_SNOFS_CLIP),
-      consistencyModel: clampImageV2Strength(fields.consistency_model, IMAGE_V2_CONSISTENCY_MODEL),
-      consistencyClip: clampImageV2Strength(fields.consistency_clip, IMAGE_V2_CONSISTENCY_CLIP),
+      consistencyModel: generateConsistencyModel,
+      consistencyClip: generateConsistencyClip,
       maskStillId: savedMask?.id,
       maskStillFilename: savedMask?.filename,
       refineStrength
@@ -275,7 +288,11 @@ export default defineEventHandler(async (event) => {
     cfg,
     mode: v2Mode,
     task: v2Task,
-    workflow: v2Mode === 'refine' ? 'klein_v2_refine.json' : v2Mode === 'compose' ? 'klein_v2_compose.json' : 'klein_v2_edit.json',
+    workflow: v2Mode === 'generate'
+      ? 'klein_v2_generate.json'
+      : v2Mode === 'refine' ? 'klein_v2_refine.json' : v2Mode === 'compose' ? 'klein_v2_compose.json' : 'klein_v2_edit.json',
+    width: size?.width || 0,
+    height: size?.height || 0,
     strength: refineStrength,
     hideThumbnail,
     folderLocked