towsty/aigen
1
0
Fork 0
Code Issues Pull Requests 1 Actions Packages Projects Releases Wiki Activity

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

Browse Source 
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>
This commit is contained in:
Towsty
2026-08-28 22:56:33 -05:00
co-authored by Cursor
parent a2895b8054
commit b35f9af10d
9 changed files with 511 additions and 112 deletions
Download Patch File Download Diff File Expand all files Collapse all files
server/utils/generationPresets.ts
+8 -5
View File
@@ -23,6 +23,7 @@ import {
IMAGE_V2_STEPS_DEFAULT,
IMAGE_V2_DENOISE_DEFAULT,
clampImageV2Denoise,
clampImageV2Size,
clampImageV2Strength,
parseImageV2Mode,
parseImageV2Task,
@@ -116,17 +117,19 @@ function sanitizeImageV2Settings(raw: unknown): ImageV2PresetSettings {
const turbo = rec.turbo === true
return {
mode,
task: mode === 'compose' ? parseImageV2Task(rec.task, 'scene') : mode === 'refine' ? 'refine' : 'scene',
task: mode === 'compose' ? parseImageV2Task(rec.task, 'scene') : mode === 'refine' ? 'refine' : mode === 'generate' ? 't2i' : 'scene',
negative: String(rec.negative || '').slice(0, 2000),
snofsModel: clampImageV2Strength(rec.snofsModel ?? rec.snofs_model, IMAGE_V2_SNOFS_MODEL),
snofsClip: clampImageV2Strength(rec.snofsClip ?? rec.snofs_clip, IMAGE_V2_SNOFS_CLIP),
consistencyModel: clampImageV2Strength(rec.consistencyModel ?? rec.consistency_model, IMAGE_V2_CONSISTENCY_MODEL),
consistencyClip: clampImageV2Strength(rec.consistencyClip ?? rec.consistency_clip, IMAGE_V2_CONSISTENCY_CLIP),
consistencyModel: clampImageV2Strength(rec.consistencyModel ?? rec.consistency_model, mode === 'generate' ? 0 : IMAGE_V2_CONSISTENCY_MODEL),
consistencyClip: clampImageV2Strength(rec.consistencyClip ?? rec.consistency_clip, mode === 'generate' ? 0 : IMAGE_V2_CONSISTENCY_CLIP),
steps: turbo ? 8 : clampImageSteps(rec.steps, IMAGE_V2_STEPS_DEFAULT),
cfg: turbo ? 1 : clampImageCfg(rec.cfg, IMAGE_V2_CFG_DEFAULT),
megapixels: clampImageScaleMegapixels(rec.megapixels ?? rec.scaleMegapixels, 1),
megapixels: mode === 'generate' ? undefined : clampImageScaleMegapixels(rec.megapixels ?? rec.scaleMegapixels, 1),
turbo,
strength: mode === 'refine' ? clampImageV2Denoise(rec.strength, IMAGE_V2_DENOISE_DEFAULT) : undefined
strength: mode === 'refine' ? clampImageV2Denoise(rec.strength, IMAGE_V2_DENOISE_DEFAULT) : undefined,
width: mode === 'generate' ? clampImageV2Size(rec.width, rec.height, rec.aspect).width : undefined,
height: mode === 'generate' ? clampImageV2Size(rec.width, rec.height, rec.aspect).height : undefined
}
}
server/utils/imageChainV2.ts
+35 -21
View File
@@ -13,7 +13,7 @@ import type { EditImageFile } from '~/server/utils/imageChain'
export type EditV2RunParams = {
mode: ImageV2Mode
task: ImageV2Task
image: EditImageFile
image?: EditImageFile | null
reference: EditImageFile | null
mask?: EditImageFile | null
prompt: string
@@ -27,6 +27,9 @@ export type EditV2RunParams = {
consistencyClip: number
megapixels: number
strength?: number
width?: number
height?: number
turbo?: boolean
}
export async function runEditV2(job: Job, params: EditV2RunParams) {
@@ -41,6 +44,9 @@ export async function runEditV2(job: Job, params: EditV2RunParams) {
if (params.mode === 'edit' && params.reference) {
throw new Error('Edit mode takes one image. Use Compose for two stills.')
}
if (params.mode !== 'generate' && !params.image) {
throw new Error('This v2 mode requires still A.')
}
try {
await ensureComfyReady((status) => {
@@ -55,15 +61,18 @@ export async function runEditV2(job: Job, params: EditV2RunParams) {
job.imageComfyHost = getComfyHost()
await withImageComfyHost(job.imageComfyHost, async () => {
const generate = params.mode === 'generate'
emitChainJob(job, {
type: 'status',
message: params.mode === 'refine'
? 'Uploading canvas and mask to Beast...'
: params.mode === 'compose' ? 'Uploading stills A and B to Beast...' : 'Uploading still A to Beast...',
progress: 8
message: generate
? 'Queueing Klein v2 generate on Beast...'
: params.mode === 'refine'
? 'Uploading canvas and mask to Beast...'
: params.mode === 'compose' ? 'Uploading stills A and B to Beast...' : 'Uploading still A to Beast...',
progress: generate ? 12 : 8
})
const uploaded = await uploadImage(params.image, job.id)
const uploadedRef = params.mode === 'refine'
const uploaded = generate || !params.image ? null : await uploadImage(params.image, job.id)
const uploadedRef = generate || params.mode === 'refine'
? null
: params.reference
? await uploadImage({
@@ -71,7 +80,7 @@ export async function runEditV2(job: Job, params: EditV2RunParams) {
filename: `ref_${params.reference.filename || 'image_b.png'}`
}, job.id)
: null
const uploadedMask = params.mode === 'refine' && params.mask
const uploadedMask = !generate && params.mode === 'refine' && params.mask
? await uploadImage({
...params.mask,
filename: `mask_${params.mask.filename || 'refine-mask.png'}`
@@ -79,23 +88,25 @@ export async function runEditV2(job: Job, params: EditV2RunParams) {
: null
if (job.status === 'cancelled') throw new Error('Job interrupted.')
emitChainJob(job, {
type: 'status',
message: params.mode === 'refine'
? 'Queueing Klein v2 refine on Beast...'
: params.mode === 'compose'
? `Queueing Klein v2 compose (${params.task}) on Beast...`
: 'Queueing Klein v2 edit on Beast...',
progress: 12
})
if (!generate) {
emitChainJob(job, {
type: 'status',
message: params.mode === 'refine'
? 'Queueing Klein v2 refine on Beast...'
: params.mode === 'compose'
? `Queueing Klein v2 compose (${params.task}) on Beast...`
: 'Queueing Klein v2 edit on Beast...',
progress: 12
})
}
await ensureComfyLoraNames('image')
await assertImageScaleToTotalPixelsNode()
if (!generate) await assertImageScaleToTotalPixelsNode()
const built = buildImageV2Workflow({
mode: params.mode,
task: params.task,
prompt: params.prompt,
negative: params.negative,
imageAName: uploaded.name,
imageAName: uploaded?.name,
imageBName: uploadedRef?.name,
maskName: uploadedMask?.name,
strength: params.strength,
@@ -107,6 +118,9 @@ export async function runEditV2(job: Job, params: EditV2RunParams) {
cfg: params.cfg,
seed: params.seed,
megapixels: params.megapixels,
width: params.width,
height: params.height,
turbo: params.turbo === true,
filenamePrefix: `aigen_v2_${job.id.slice(0, 8)}`
})
const queued = await queuePrompt(built.graph, job.clientId)
@@ -154,8 +168,8 @@ export async function runEditV2(job: Job, params: EditV2RunParams) {
job.stillId = still?.id
await purgeComfyArtifacts({
video: { filename: output.filename, subfolder: output.subfolder, type: output.type },
imageName: uploaded.name,
imageSubfolder: uploaded.subfolder,
imageName: uploaded?.name,
imageSubfolder: uploaded?.subfolder,
extraImageNames: [uploadedRef?.name, uploadedMask?.name].filter((name): name is string => Boolean(name)),
promptId: job.promptId
})
server/utils/imageWorkflowV2.ts
+73 -16
View File
@@ -1,6 +1,7 @@
import editTemplate from '../assets/klein_v2_edit.json'
import composeTemplate from '../assets/klein_v2_compose.json'
import refineTemplate from '../assets/klein_v2_refine.json'
import generateTemplate from '../assets/klein_v2_generate.json'
import { IMAGE_SCALE_TO_TOTAL_PIXELS } from '~/server/utils/comfy'
import { cachedComfyLoraNames } from '~/server/utils/loras'
import { resolveComfyLoraName, loraIdentityKey } from '~/utils/loras'
@@ -13,7 +14,10 @@ import {
IMAGE_V2_SNOFS_LORA,
IMAGE_V2_SNOFS_MODEL,
IMAGE_V2_DENOISE_DEFAULT,
IMAGE_V2_GENERATE_HEIGHT,
IMAGE_V2_GENERATE_WIDTH,
clampImageV2Denoise,
clampImageV2Size,
clampImageV2Strength,
composeImageV2Prompt,
type ImageV2Mode,
@@ -41,16 +45,20 @@ const CONSISTENCY = '8'
export const IMAGE_V2_EDIT_WORKFLOW = 'klein_v2_edit.json'
export const IMAGE_V2_COMPOSE_WORKFLOW = 'klein_v2_compose.json'
export const IMAGE_V2_REFINE_WORKFLOW = 'klein_v2_refine.json'
export const IMAGE_V2_GENERATE_WORKFLOW = 'klein_v2_generate.json'
export interface ImageV2BuildParams {
mode: ImageV2Mode
task: ImageV2Task
prompt: string
negative?: string
imageAName: string
imageAName?: string
imageBName?: string
maskName?: string
strength?: number
width?: number
height?: number
turbo?: boolean
snofsModel?: number
snofsClip?: number
consistencyModel?: number
@@ -102,8 +110,30 @@ function graphHasMaskInput(graph: WorkflowGraph) {
})
}
function bypassLoraNode(graph: WorkflowGraph, id: string, modelFrom: string, clipFrom: string) {
delete graph[id]
for (const node of Object.values(graph)) {
for (const [key, value] of Object.entries(node.inputs)) {
if (!Array.isArray(value) || value[0] !== id) continue
node.inputs[key] = value[1] === 1 ? [clipFrom, 1] : [modelFrom, 0]
}
}
}
export function assertImageV2Graph(graph: WorkflowGraph, mode: ImageV2Mode, imageBName?: string) {
const loaders = loadImageNames(graph)
if (mode === 'generate') {
if (loaders.length) {
throw createError({
statusCode: 500,
statusMessage: 'Generate graph has a required LoadImage. Refusing to run an edit fallback.'
})
}
const latent = Object.values(graph).find(node => node.class_type === 'EmptyFlux2LatentImage')
if (!latent) {
throw createError({ statusCode: 500, statusMessage: 'Generate graph is missing EmptyFlux2LatentImage.' })
}
}
if (mode === 'refine') {
const mask = graph[LOAD_MASK]
if (!mask || mask.class_type !== 'LoadImage' || !String(mask.inputs.image || '').trim()) {
@@ -140,7 +170,7 @@ export function assertImageV2Graph(graph: WorkflowGraph, mode: ImageV2Mode, imag
})
}
}
if (mode !== 'refine' && imageBName && loaders.length < 2) {
if (mode !== 'refine' && mode !== 'generate' && imageBName && loaders.length < 2) {
throw createError({
statusCode: 500,
statusMessage: 'image_b was sent but the executed graph has no second image input.'
@@ -161,43 +191,65 @@ export function assertImageV2Graph(graph: WorkflowGraph, mode: ImageV2Mode, imag
export function buildImageV2Workflow(params: ImageV2BuildParams) {
const compose = params.mode === 'compose'
const refine = params.mode === 'refine'
const generate = params.mode === 'generate'
if (refine && !String(params.maskName || '').trim()) {
throw createError({ statusCode: 400, statusMessage: 'Refine requires a mask. Refusing to fall back to Edit.' })
}
const graph = structuredClone(refine ? refineTemplate : compose ? composeTemplate : editTemplate) as WorkflowGraph
const graph = structuredClone(
generate ? generateTemplate : refine ? refineTemplate : compose ? composeTemplate : editTemplate
) as WorkflowGraph
const prompt = composeImageV2Prompt(params.mode, params.task, params.prompt)
const negative = String(params.negative || '')
const snofsModel = clampImageV2Strength(params.snofsModel, IMAGE_V2_SNOFS_MODEL)
const snofsClip = clampImageV2Strength(params.snofsClip, IMAGE_V2_SNOFS_CLIP)
const consistencyModel = clampImageV2Strength(params.consistencyModel, IMAGE_V2_CONSISTENCY_MODEL)
const consistencyClip = clampImageV2Strength(params.consistencyClip, IMAGE_V2_CONSISTENCY_CLIP)
const consistencyModel = clampImageV2Strength(
params.consistencyModel,
generate ? 0 : IMAGE_V2_CONSISTENCY_MODEL
)
const consistencyClip = clampImageV2Strength(
params.consistencyClip,
generate ? 0 : IMAGE_V2_CONSISTENCY_CLIP
)
const steps = clampImageSteps(params.steps, 24)
const cfg = clampImageCfg(params.cfg, 4)
const strength = refine ? clampImageV2Denoise(params.strength, IMAGE_V2_DENOISE_DEFAULT) : undefined
const workflowFile = refine ? IMAGE_V2_REFINE_WORKFLOW : compose ? IMAGE_V2_COMPOSE_WORKFLOW : IMAGE_V2_EDIT_WORKFLOW
const size = generate ? clampImageV2Size(params.width, params.height) : null
const workflowFile = generate
? IMAGE_V2_GENERATE_WORKFLOW
: refine ? IMAGE_V2_REFINE_WORKFLOW : compose ? IMAGE_V2_COMPOSE_WORKFLOW : IMAGE_V2_EDIT_WORKFLOW
setInput(graph, LOAD_A, 'image', params.imageAName)
if (!generate) setInput(graph, LOAD_A, 'image', params.imageAName || '')
if (compose) setInput(graph, LOAD_B, 'image', params.imageBName || '')
if (refine) {
setInput(graph, LOAD_MASK, 'image', params.maskName || '')
setInput(graph, SCHEDULER_DENOISE, 'denoise', strength)
}
if (generate && size) {
setInput(graph, '14', 'width', size.width)
setInput(graph, '14', 'height', size.height)
setInput(graph, SCHEDULER, 'width', size.width)
setInput(graph, SCHEDULER, 'height', size.height)
}
setInput(graph, PROMPT, 'text', prompt)
setInput(graph, NEGATIVE, 'text', negative)
setInput(graph, NOISE, 'noise_seed', params.seed)
setInput(graph, SCHEDULER, 'steps', steps)
setInput(graph, CFG, 'cfg', cfg)
setInput(graph, SAVE, 'filename_prefix', params.filenamePrefix || 'aigen-v2')
patchScaleMegapixels(graph, params.megapixels ?? 1)
setInput(graph, SAVE, 'filename_prefix', params.filenamePrefix || (generate ? 'aigen-v2-generate' : 'aigen-v2'))
if (!generate) patchScaleMegapixels(graph, params.megapixels ?? 1)
setInput(graph, SNOFS, 'lora_name', resolveRequiredLora(IMAGE_V2_SNOFS_LORA, 'SNOFS'))
setInput(graph, SNOFS, 'strength_model', snofsModel)
setInput(graph, SNOFS, 'strength_clip', snofsClip)
setInput(graph, CONSISTENCY, 'lora_name', resolveRequiredLora(IMAGE_V2_CONSISTENCY_LORA, 'Consistency'))
setInput(graph, CONSISTENCY, 'strength_model', consistencyModel)
setInput(graph, CONSISTENCY, 'strength_clip', consistencyClip)
if (generate && consistencyModel <= 0 && consistencyClip <= 0) {
bypassLoraNode(graph, CONSISTENCY, SNOFS, SNOFS)
} else {
setInput(graph, CONSISTENCY, 'lora_name', resolveRequiredLora(IMAGE_V2_CONSISTENCY_LORA, 'Consistency'))
setInput(graph, CONSISTENCY, 'strength_model', consistencyModel)
setInput(graph, CONSISTENCY, 'strength_clip', consistencyClip)
}
assertImageV2Graph(graph, params.mode, refine ? undefined : params.imageBName)
assertImageV2Graph(graph, params.mode, generate || refine ? undefined : params.imageBName)
const loaders = loadImageNames(graph)
console.log(JSON.stringify({
@@ -205,18 +257,22 @@ export function buildImageV2Workflow(params: ImageV2BuildParams) {
workflow: workflowFile,
mode: params.mode,
task: params.task,
canvas: { id: LOAD_A, file: graph[LOAD_A]?.inputs.image },
canvas: generate ? undefined : { id: LOAD_A, file: graph[LOAD_A]?.inputs.image },
mask: refine ? { id: LOAD_MASK, file: graph[LOAD_MASK]?.inputs.image } : undefined,
size: generate ? { width: size?.width ?? IMAGE_V2_GENERATE_WIDTH, height: size?.height ?? IMAGE_V2_GENERATE_HEIGHT } : undefined,
strength,
turbo: params.turbo === true,
loadImage: Object.fromEntries(loaders.map(item => [item.id, { title: item.title, file: item.image }])),
loras: {
snofs: { name: graph[SNOFS]?.inputs.lora_name, model: snofsModel, clip: snofsClip },
consistency: { name: graph[CONSISTENCY]?.inputs.lora_name, model: consistencyModel, clip: consistencyClip }
consistency: graph[CONSISTENCY]
? { name: graph[CONSISTENCY]?.inputs.lora_name, model: consistencyModel, clip: consistencyClip }
: { loaded: false, model: 0, clip: 0 }
},
steps,
cfg,
seed: params.seed,
megapixels: clampImageScaleMegapixels(params.megapixels ?? 1)
megapixels: generate ? undefined : clampImageScaleMegapixels(params.megapixels ?? 1)
}))
return { graph, workflowFile, loaders, prompt, strength }
@@ -238,6 +294,7 @@ export const IMAGE_V2_NODE_LABELS: Record<string, string> = {
'11': 'Encoding image A',
'24': 'Encoding image B',
'33': 'Applying mask',
'14': 'Building empty Klein latent',
'19': 'Sampling Klein v2',
'20': 'Decoding still',
'21': 'Saving still'
server/utils/studioQueue.ts
+26 -14
View File
@@ -52,8 +52,8 @@ export interface StudioJobPayload {
scaleToTotalPixels?: boolean
scaleMegapixels?: number
imagePipeline?: 'v1' | 'v2'
v2Mode?: 'edit' | 'compose' | 'refine'
v2Task?: 'scene' | 'identity' | 'outfit' | 'face_lock' | 'refine'
v2Mode?: 'edit' | 'compose' | 'refine' | 'generate'
v2Task?: 'scene' | 'identity' | 'outfit' | 'face_lock' | 'refine' | 't2i'
snofsModel?: number
snofsClip?: number
consistencyModel?: number
@@ -679,14 +679,17 @@ async function startStudioEditJob(item: StudioJob) {
const { runEditV2 } = await import('~/server/utils/imageChainV2')
const payload = item.payload
if (!payload.stillId || !existsSync(stillPath(item.ownerKey, payload.stillId))) {
const generate = payload.imagePipeline === 'v2' && payload.v2Mode === 'generate'
if (!generate && (!payload.stillId || !existsSync(stillPath(item.ownerKey, payload.stillId)))) {
throw new Error('The input still is missing from the library')
}
const image = {
filename: payload.stillFilename || 'still.png',
data: readFileSync(stillPath(item.ownerKey, payload.stillId)),
type: 'image/png'
}
const image = payload.stillId && existsSync(stillPath(item.ownerKey, payload.stillId))
? {
filename: payload.stillFilename || 'still.png',
data: readFileSync(stillPath(item.ownerKey, payload.stillId)),
type: 'image/png'
}
: null
const refId = payload.referenceStillId
const reference = refId && existsSync(stillPath(item.ownerKey, refId))
? {
@@ -697,7 +700,13 @@ async function startStudioEditJob(item: StudioJob) {
: null
if (payload.imagePipeline === 'v2') {
const mode = payload.v2Mode === 'compose' ? 'compose' : payload.v2Mode === 'refine' ? 'refine' : 'edit'
const mode = payload.v2Mode === 'compose'
? 'compose'
: payload.v2Mode === 'refine'
? 'refine'
: payload.v2Mode === 'generate'
? 'generate'
: 'edit'
const maskId = payload.maskStillId
const mask = mode === 'refine' && maskId && existsSync(stillPath(item.ownerKey, maskId))
? {
@@ -745,8 +754,8 @@ async function startStudioEditJob(item: StudioJob) {
void runEditV2(live, {
mode,
task: payload.v2Task || 'scene',
image,
reference: mode === 'refine' ? null : reference,
image: mode === 'generate' ? null : image,
reference: mode === 'refine' || mode === 'generate' ? null : reference,
mask,
prompt: payload.prompt,
negative: payload.negative || '',
@@ -755,10 +764,13 @@ async function startStudioEditJob(item: StudioJob) {
cfg: payload.cfg,
snofsModel: payload.snofsModel ?? 0.65,
snofsClip: payload.snofsClip ?? 0.35,
consistencyModel: payload.consistencyModel ?? 0.7,
consistencyClip: payload.consistencyClip ?? 0.7,
consistencyModel: payload.consistencyModel ?? (mode === 'generate' ? 0 : 0.7),
consistencyClip: payload.consistencyClip ?? (mode === 'generate' ? 0 : 0.7),
megapixels: payload.scaleMegapixels ?? 1,
strength: payload.refineStrength
strength: payload.refineStrength,
width: payload.width,
height: payload.height,
turbo: payload.turbo === true
}).catch((error) => {
const message = error instanceof Error ? error.message : String(error)
if (live && live.status !== 'error' && live.status !== 'cancelled' && live.status !== 'deferred') {