Add Image v2 Refine for on-device regional edits.

Mask painter and denoise strength sit beside Edit and Compose. The canned hand/chest prompt helper is gone so that text stays yours.

Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
Towsty
2026-08-28 22:03:20 -05:00
co-authored by Cursor
parent 21fe2412f9
commit a2895b8054
10 changed files with 825 additions and 56 deletions
+36 -8
View File
@@ -11,6 +11,8 @@ import {
IMAGE_V2_STEPS_DEFAULT,
IMAGE_V2_TURBO_CFG,
IMAGE_V2_TURBO_STEPS,
IMAGE_V2_DENOISE_DEFAULT,
clampImageV2Denoise,
clampImageV2Strength,
parseImageV2Mode,
parseImageV2Task,
@@ -37,7 +39,7 @@ async function fileFromUrl(url: string): Promise<ImageFile> {
}
const mime = String(res.headers.get('content-type') || 'image/png').split(';')[0]
if (!/^image\//i.test(mime)) {
throw createError({ statusCode: 400, statusMessage: 'image_a / image_b URL must be an image' })
throw createError({ statusCode: 400, statusMessage: 'image_a / image_b / mask URL must be an image' })
}
const data = Buffer.from(await res.arrayBuffer())
if (data.length > 40 * 1024 * 1024) {
@@ -72,16 +74,19 @@ function readMultipart(parts: Array<{ name?: string; filename?: string; type?: s
const fields: Record<string, string> = {}
let imageA: ImageFile | null = null
let imageB: ImageFile | null = null
let mask: ImageFile | null = null
for (const part of parts || []) {
if ((part.name === 'image_a' || part.name === 'image') && part.filename && part.data?.length) {
imageA = { filename: part.filename, data: part.data, type: part.type }
} else if ((part.name === 'image_b' || part.name === 'image2') && part.filename && part.data?.length) {
imageB = { filename: part.filename, data: part.data, type: part.type }
} else if (part.name === 'mask' && part.filename && part.data?.length) {
mask = { filename: part.filename, data: part.data, type: part.type }
} else if (part.name && part.data) {
fields[part.name] = part.data.toString('utf8')
}
}
return { fields, imageA, imageB }
return { fields, imageA, imageB, mask }
}
export default defineEventHandler(async (event) => {
@@ -89,6 +94,7 @@ export default defineEventHandler(async (event) => {
let fields: Record<string, unknown> = {}
let uploadedA: ImageFile | null = null
let uploadedB: ImageFile | null = null
let uploadedMask: ImageFile | null = null
if (contentType.includes('multipart/form-data')) {
const form = await readMultipartFormData(event).catch(() => null)
@@ -96,15 +102,16 @@ export default defineEventHandler(async (event) => {
fields = parsed.fields
uploadedA = parsed.imageA
uploadedB = parsed.imageB
uploadedMask = parsed.mask
} else {
fields = await readBody<Record<string, unknown>>(event).catch(() => ({}))
}
const mode = parseImageV2Mode(fields.mode)
if (!mode) {
throw createError({ statusCode: 400, statusMessage: 'mode must be edit or compose' })
throw createError({ statusCode: 400, statusMessage: 'mode must be edit, compose, or refine' })
}
const task = parseImageV2Task(fields.task, 'scene')
const task = parseImageV2Task(fields.task, mode === 'refine' ? 'refine' : 'scene')
const prompt = String(fields.prompt || '').trim()
if (!prompt) {
throw createError({ statusCode: 400, statusMessage: 'A prompt is required' })
@@ -125,11 +132,18 @@ export default defineEventHandler(async (event) => {
const ownerKey = libraryOwnerKey(event)
const imageA = await resolveImageRef(ownerKey, fields.image_a, uploadedA)
const imageB = await resolveImageRef(ownerKey, fields.image_b, uploadedB)
const imageB = mode === 'refine' ? null : await resolveImageRef(ownerKey, fields.image_b, uploadedB)
const mask = mode === 'refine' ? await resolveImageRef(ownerKey, fields.mask, uploadedMask) : null
if (!imageA) {
throw createError({ statusCode: 400, statusMessage: 'image_a is required' })
}
if (mode === 'refine' && !mask) {
throw createError({
statusCode: 400,
statusMessage: 'Refine requires a mask. Refusing to fall back to Edit.'
})
}
if (mode === 'edit' && imageB) {
throw createError({
statusCode: 400,
@@ -166,7 +180,8 @@ export default defineEventHandler(async (event) => {
const size = imageDimensions(imageA.data)
const clipName = String(fields.name || '').trim().slice(0, 80)
const v2Mode = mode as ImageV2Mode
const v2Task = (mode === 'edit' ? 'scene' : task) as ImageV2Task
const v2Task = (mode === 'refine' ? 'refine' : mode === 'edit' ? 'scene' : task) as ImageV2Task
const refineStrength = mode === 'refine' ? clampImageV2Denoise(fields.strength, IMAGE_V2_DENOISE_DEFAULT) : undefined
const still = await rememberInputStill({
ownerKey,
@@ -186,6 +201,15 @@ export default defineEventHandler(async (event) => {
hideInput
})
: null
const savedMask = mask
? await rememberInputStill({
ownerKey,
folderId,
filename: mask.filename || 'refine-mask.png',
data: mask.data,
hideInput
})
: null
const studio = await addStudioJob({
ownerKey,
@@ -228,7 +252,10 @@ export default defineEventHandler(async (event) => {
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)
consistencyClip: clampImageV2Strength(fields.consistency_clip, IMAGE_V2_CONSISTENCY_CLIP),
maskStillId: savedMask?.id,
maskStillFilename: savedMask?.filename,
refineStrength
}
})
await kickStudioQueue()
@@ -248,7 +275,8 @@ export default defineEventHandler(async (event) => {
cfg,
mode: v2Mode,
task: v2Task,
workflow: v2Mode === 'compose' ? 'klein_v2_compose.json' : 'klein_v2_edit.json',
workflow: v2Mode === 'refine' ? 'klein_v2_refine.json' : v2Mode === 'compose' ? 'klein_v2_compose.json' : 'klein_v2_edit.json',
strength: refineStrength,
hideThumbnail,
folderLocked
}