The all-in-one local AI studio for your desktop: chat, image and video generation and a coding agent in one free, open source app. Windows and Linux. No Docker, no terminal, no cloud required.
TypeScript agentai-chatai-studioall-in-onecoding-agentdesktop-appfluximage-generationlinuxllmlocal-ailocal-llmopen-sourceprivacyself-hostedstable-diffusionuncensoreduncensored-aivideo-generationwindows
// Static SEED of the hosted render catalog — the offline/never-fetched
// fallback only. The live truth is GET /api/jobs/catalog (fetched into
// cloudCatalogStore on every account probe), so a shipped build picks up
// fleet/pricing changes without an app update. Keep the ids in sync with
// the web app when touching this file.
import type { RenderKind, RenderOp } from './cloud-jobs'
import type { ModelTier, ModelWeights, Schema, StudioModel } from './studio-contract'
/** How a model that is metered by the length of its input is priced
* (`credits.length` of /api/jobs/catalog). */
export interface CloudLengthPrice {
/** What is measured. The server lists only the first three; 'extension'
* (a fixed length the model adds) is priced per run and has no entry, it
* is named here because the shared order code asks about it. */
basis: 'audio' | 'video' | 'chars' | 'extension'
/** Units per billed block: seconds, or characters. */
block: number
/** Credits per started block. */
per_block: number
/** The fewest units the provider bills. */
min: number
/** The most the server takes. */
max: number
/** The length `credits.base` is the price of. */
quote: number
}
export interface CloudModel {
id: string
/** The name the app shows. From the live catalog this is the server's
* `native_label` where it sends one, else its `label` (api/cloud/catalog.ts,
* desktopCatalog). */
label: string
kind: RenderKind
/** P3: present only on a Studio-capable catalog entry (server field
* `api_schema`), the provider's own field schema for the endpoint, so the
* Studio controls can be built without a second copy of it. Undefined on
* every classic model and on any catalog payload from a server that does
* not know Studio yet. */
api_schema?: Schema
/** P3: the provider's pricing formula for a Studio model (server field
* `pricing`, mirrors `StudioModel['price']` in studio-contract.ts). DISPLAY
* only, the preview shown before a run starts: a run always books the
* number `POST /api/jobs/studio-quote` confirms, never this formula. */
pricing?: StudioModel['price']
/** P3: marks a Studio model (server field `quote_required`). This is the
* ONLY signal "can the server do Studio" ever reads: absence, not a
* version number, is what an older catalog payload looks like, and the
* presence of this field on at least one entry is what unlocks the preset
* shelf. A model carrying it must never be priced from a client formula. */
quote_required?: boolean
/** 3.1.0: the server lists this not yet measured studio model only because
* its switch for the measurement is on (studio-contract.ts, unmeasuredHidden). */
unmeasured?: boolean
/** Supports the masked img2img 'edit' op (flux-dev only today). */
edit?: boolean
/** R5-66: an instruction-based edit endpoint that takes prompt + image and
* needs no mask at all (Web parity: the web app,
* qwen-image-edit). Only meaningful on an `ops: ['edit']` model; classic
* `edit: true` models (flux-dev) are always mask-required. */
maskless?: boolean
/** Video: renders text-to-video (the "Video" intent). Absent = yes; set false
* on an i2v-only model to keep it out of the Video picker. */
t2v?: boolean
/** Video: renders image-to-video (the "Animate Image" intent). Absent = yes;
* set false on a t2v-only model to keep it out of the Animate picker. */
i2v?: boolean
/** Animate (B38, W6-API 30.10.2026): the model reads an end image
* (`params.last_image_path`). Absent on a server from before W6, which
* took an end image and dropped it, so the app offers none there. */
end_image?: boolean
/** W6-API: the model reads a picture list (`params.image_paths`), at most
* this many, the start image counted. Studio reference models only. */
reference_images?: number
/** W6-API: a Studio model that belongs in the Animate picker. */
animate?: boolean
/** C2: the provider ships this endpoint with its own filter off, so it
* produces adult output when the account's content policy allows it (the
* server decides that). Drives the
* "No Filter" picker mark (ModelChip.tsx) and grants nothing on its own:
* the server re-checks the account's policy on every job, this field is
* display only. Optional and defaults to falsy: the live catalog
* (GET /api/jobs/catalog) only started emitting it once the web-side fix
* landed, and an older/offline payload without the field must read as "not
* adult", never crash or silently mismark a model. */
adult?: boolean
/** 2.5.8 op-specialized models (trainers, lipsync, voice, music, extend,
* motion, LoRA-gen): exactly the ops this model serves. Absent on classic
* models — every classic picker filters on `!m.ops`. */
ops?: RenderOp[]
/** Character-Studio generation endpoint: accepts `params.loras`. */
lora?: boolean
/** Lipsync base input: still portrait ('image') or existing clip ('video'). */
lipsync_source?: 'image' | 'video'
/** Music: the endpoint has a real lyrics input (server truth, 2026-08-08:
* only ace-step-1.5). The lyrics box is offered only when set; the other
* music endpoints write their own lyrics from the prompt. */
lyrics?: boolean
/** Whether the hosted endpoint honours guidance_scale (CFG). */
cfg?: boolean
/** Whether the hosted endpoint honours negative_prompt. */
negative_prompt?: boolean
/** Video: clip lengths the model books (5s short / 8s long).
*
* `durations` (P3, server field `clip.durations`, from dd29f359): every
* length the model actually books, not just the two named buttons; the
* booking truth is `useCloudCreate`'s call to `bookedVideoSeconds()` in
* `video-duration.ts`, which reads this SAME list first (via
* `effectiveVideoDurations()`, also what the picker shows) and only falls
* back to `video-durations.json` when this field is absent (review-
* studio-A.md B2, fixed Runde 2 20.09.2026). Optional: an older catalog
* payload omits it and the short/long pair above still holds. */
clip?: { short: number; long?: number; durations?: number[] }
/** Per-run credit cost (base = image or 5s clip, long = 8s clip; music
* models additionally quote per_s for the duration slider).
*
* `lora` is what one run costs when it carries a trained character: the
* worker then books the provider's `-lora` twin, a different endpoint at a
* different price. It REPLACES the base rate, it does not add to it, and it
* is set only where the twin really costs more than our base rate covers.
* /api/jobs/catalog emits it on exactly the same terms, so the client never
* needs a second price table.
*
* `by_duration` (P3, server field `credits.by_duration`, from dd29f359):
* the exact price per bookable length, keyed by the length in seconds as
* a string (matches `clip.durations`). Lets a client price a length other
* than the short/long pair precisely instead of rounding onto one of
* them; `runCredits()` in cloudCatalogStore.ts prefers this when present.
*
* `length` (server field `credits.length`, 09.10.2026): the provider meters
* this model by the length of what goes in, seconds of the audio or video
* or characters of the text. `base` is then only the price of `quote`
* units and the most a start without `max_credits` can book. The real
* price is `per_block` per started `block`, never below `min` units, and
* the server refuses more than `max`. Absent on a model priced per run and
* on an older server; `lengthPrice()` and `lengthCredits()` in
* cloudCatalogStore.ts read it. */
credits?: {
base: number; long?: number; per_s?: number; lora?: number; by_duration?: Record<string, number>
length?: CloudLengthPrice
}
/** Stufe in den Waehlern (server field `tier`, 02.10.2026): 'best' steht oben,
* 'older' gesammelt unten, nichts wird versteckt. Optional: ein aelterer
* Server und der Notvorrat eines frueheren Stands liefern es nicht, dann
* gilt das Modell als 'standard' (model-tier.ts, tierOf). */
tier?: ModelTier
/** Herkunft (server field `weights`): 'open' = genau diese Version hat offene
* Gewichte, 'open-family' = die Familie ist offen, diese Version noch nicht,
* 'closed' = nie offen. Optional aus demselben Grund wie `tier`; ohne das
* Feld traegt das Modell keine Marke (model-tier.ts, weightsOf). */
weights?: ModelWeights
}
/** Die Standardmodelle fuer eine NEUE Auswahl (seit 02.10.2026, wie im Web):
* offen, guenstig, auf dem Stand der Technik. Sie gelten nur, wo ein Kunde noch
* nichts gewaehlt hat UND der Server das Modell kennt. Kennt er es nicht (ein
* aelterer Server), faellt jede Stelle auf die erste Zeile ihres Waehlers
* zurueck, wie vor diesem Stand. */
export const DEFAULT_MODEL_IDS = {
image: 'z-image-turbo',
edit: 'qwen-image-2.1-edit',
video: 'minimax-h3-t2v',
animate: 'minimax-h3',
} as const
const CLIP = { short: 5, long: 8 }
// A model the web took out of its pickers (retired-models.ts) is not in this
// list either: the server no longer lists it, so the offline list must not
// offer it. Its name stays known for old runs (preset-models.ts, modelLabel).
export const CLOUD_MODEL_SEED: CloudModel[] = [
{ id: 'flux-schnell', label: 'Flux Schnell (fast)', kind: 'image', cfg: true, tier: 'older', weights: 'open' },
{ id: 'flux-dev', label: 'Flux Dev (quality)', kind: 'image', edit: true, cfg: true, tier: 'older', weights: 'open' },
{ id: 'flux-2-dev', label: 'Flux 2 Dev', kind: 'image', tier: 'standard', weights: 'open' },
{ id: 'qwen-image', label: 'Qwen Image', kind: 'image', tier: 'older', weights: 'open' },
{ id: 'z-image-turbo', label: 'Z-Image Turbo (fast)', kind: 'image', tier: 'best', weights: 'open' },
{ id: 'chroma', label: 'Chroma Uncensored', kind: 'image', adult: true, tier: 'standard', weights: 'open' },
{ id: 'prefect-pony', label: 'Prefect Pony XL (illustration) Uncensored', kind: 'image', adult: true, tier: 'standard', weights: 'open' },
{ id: 'neta-lumina', label: 'Neta Lumina (anime) Uncensored', kind: 'image', adult: true, tier: 'standard', weights: 'open' },
// Every hosted clip model does both t2v + i2v, so both flags are true. They're
// the enforced contract (Video/Animate pickers + submit filter on them), not a
// note — a future t2v-only or i2v-only model MUST set the flag it lacks to
// false. Server truth is /api/jobs/catalog; keep in sync with the web app.
{ id: 'wan-2.2-720p', label: 'Wan 2.2 720p', kind: 'video', t2v: true, i2v: true, negative_prompt: true, clip: CLIP, tier: 'older', weights: 'open' },
// One clip length: its endpoints take no duration field (video-durations.json).
{ id: 'wan-2.2-fast', label: 'Wan 2.2 Fast', kind: 'video', t2v: true, i2v: true, negative_prompt: true, clip: { short: 5 }, tier: 'older', weights: 'open' },
{ id: 'ltx-2', label: 'LTX-2 (with audio)', kind: 'video', t2v: true, i2v: true, clip: CLIP, credits: { base: 8000, lora: 10000 }, tier: 'older', weights: 'open' },
// Image-to-video endpoints the provider ships with its own filter off
// (mirroring the web app). None of them has a
// text-to-video twin at the provider, hence t2v: false: the path is an image
// first, then one of these. CLIP does not fit here, the provider quotes only
// a 5 s rate for them, so there is no 8s button to offer.
// 2026-09-10:
{ id: 'wan-2.2-spicy', label: 'Wan 2.2 Uncensored', kind: 'video', t2v: false, i2v: true, adult: true, clip: { short: 5 }, credits: { base: 15000, lora: 20000 }, tier: 'older', weights: 'open' },
{ id: 'ltx-2.3-spicy', label: 'LTX 2.3 Uncensored', kind: 'video', t2v: false, i2v: true, adult: true, clip: { short: 5 }, credits: { base: 10000, lora: 15000 }, tier: 'older', weights: 'open' },
{ id: 'wan-2.6-spicy', label: 'Wan 2.6 Uncensored', kind: 'video', t2v: false, i2v: true, adult: true, clip: { short: 5 }, credits: { base: 50000 }, tier: 'older', weights: 'open-family' },
{ id: 'wan-2.7-spicy', label: 'Wan 2.7 Uncensored', kind: 'video', t2v: false, i2v: true, adult: true, clip: { short: 5 }, credits: { base: 50000 }, tier: 'standard', weights: 'open-family' },
{ id: 'minimax-h3-spicy', label: 'MiniMax H3 Uncensored', kind: 'video', t2v: false, i2v: true, adult: true, clip: { short: 5 }, credits: { base: 20000 }, tier: 'best', weights: 'open' },
{ id: 'seedance-1.5-pro-spicy', label: 'Seedance 1.5 Pro Uncensored', kind: 'video', t2v: false, i2v: true, adult: true, clip: { short: 5 }, credits: { base: 26000 }, tier: 'older', weights: 'closed' },
// 2026-09-13:
{ id: 'seedance-2.5-spicy', label: 'Seedance 2.5 Uncensored', kind: 'video', t2v: false, i2v: true, adult: true, clip: { short: 5 }, credits: { base: 90000 }, tier: 'best', weights: 'closed' },
{ id: 'seedance-2.0-spicy', label: 'Seedance 2.0 Uncensored', kind: 'video', t2v: false, i2v: true, adult: true, clip: { short: 5 }, credits: { base: 60000 }, tier: 'older', weights: 'closed' },
{ id: 'seedance-2.0-fast-spicy', label: 'Seedance 2.0 Fast Uncensored', kind: 'video', t2v: false, i2v: true, adult: true, clip: { short: 5 }, credits: { base: 50000 }, tier: 'older', weights: 'closed' },
// 3.1.0 (web catalogue 3b5d5980):
{ id: 'seedance-2.0-mini-spicy', label: 'Seedance 2.0 Mini Uncensored', kind: 'video', t2v: false, i2v: true, adult: true, clip: { short: 5 }, credits: { base: 30000 }, tier: 'older', weights: 'closed' },
{ id: 'vidu-q3-spicy', label: 'Vidu Q3 Uncensored', kind: 'video', t2v: false, i2v: true, adult: true, clip: { short: 5 }, credits: { base: 35000 }, tier: 'standard', weights: 'closed' },
// ── 2.5.8 op-specialized fleet (Character-Studio / lipsync / voice / music /
// extend / motion). `ops` keeps them out of every classic picker; the live
// catalog (?v=2) is the pricing truth. Face-swap is banned from this list. ──
{ id: 'flux-lora-trainer', label: 'Flux Character Training', kind: 'image', ops: ['lora-train'], tier: 'older', weights: 'open' },
{ id: 'z-image-lora-trainer', label: 'Z-Image Character Training', kind: 'image', ops: ['lora-train'], tier: 'best', weights: 'open' },
{ id: 'qwen-image-lora-trainer', label: 'Qwen Character Training', kind: 'image', ops: ['lora-train'], tier: 'older', weights: 'open' },
// the web app (15.09.2026): der LTX-Video-Trainer verlangt Videos als
// Trainingsmaterial, diese Oberflaeche laedt nur Fotos hoch. Bewusst NICHT
// mitgefuehrt, kein Rueckstand: "do not advertise an input contract this
// client cannot satisfy" (Web-Kommentar an derselben Stelle).
{ id: 'flux-schnell-lora', label: 'Flux Schnell + Character', kind: 'image', ops: ['generate'], lora: true, cfg: true, tier: 'older', weights: 'open' },
{ id: 'flux-dev-lora-ultra-fast', label: 'Flux Dev Fast + Character', kind: 'image', ops: ['generate'], lora: true, cfg: true, tier: 'older', weights: 'open' },
{ id: 'z-image-turbo-lora', label: 'Z-Image Turbo + Character', kind: 'image', ops: ['generate'], lora: true, tier: 'best', weights: 'open' },
{ id: 'z-image-base-lora', label: 'Z-Image + Character', kind: 'image', ops: ['generate'], lora: true, tier: 'best', weights: 'open' },
// R5-57/58: missing from this seed entirely, so the edit picker never
// offered it and the model was unreachable until the live catalog fetch
// landed. Label and maskless flag copied verbatim from the web app.
{ id: 'qwen-image-lora', label: 'Qwen Image + Character', kind: 'image', ops: ['generate'], lora: true, tier: 'older', weights: 'open' },
{ id: 'qwen-image-edit', label: 'Qwen Image Edit (no mask needed)', kind: 'image', ops: ['edit'], maskless: true, tier: 'older', weights: 'open' },
{ id: 'infinitetalk-fast', label: 'InfiniteTalk (photo avatar)', kind: 'video', ops: ['lipsync'], lipsync_source: 'image', t2v: false, i2v: false, tier: 'standard', weights: 'open' },
{ id: 'p-video-avatar', label: 'P-Video Avatar (photo, fast)', kind: 'video', ops: ['lipsync'], lipsync_source: 'image', t2v: false, i2v: false, tier: 'standard', weights: 'closed' },
{ id: 'latentsync', label: 'LatentSync (resync a clip)', kind: 'video', ops: ['lipsync'], lipsync_source: 'video', t2v: false, i2v: false, tier: 'standard', weights: 'open' },
{ id: 'lipsync-2', label: 'Lipsync-2 (resync a clip)', kind: 'video', ops: ['lipsync'], lipsync_source: 'video', t2v: false, i2v: false, tier: 'older', weights: 'closed' },
{ id: 'qwen3-tts', label: 'Qwen3 TTS (voices)', kind: 'audio', ops: ['tts'], tier: 'standard', weights: 'open' },
{ id: 'qwen3-tts-clone', label: 'Qwen3 TTS Voice Clone', kind: 'audio', ops: ['tts'], tier: 'standard', weights: 'open' },
{ id: 'qwen3-tts-design', label: 'Qwen3 TTS Voice Design', kind: 'audio', ops: ['tts'], tier: 'standard', weights: 'open' },
{ id: 'ace-step-1.5', label: 'ACE-Step 1.5', kind: 'audio', ops: ['music'], lyrics: true, credits: { base: 1800, per_s: 30 }, tier: 'standard', weights: 'open' },
{ id: 'sonilo-music', label: 'Sonilo Music', kind: 'audio', ops: ['music'], credits: { base: 15000, per_s: 250 }, tier: 'standard', weights: 'closed' },
{ id: 'wan-2.2-spicy-extend', label: 'Wan 2.2 Uncensored Extend', kind: 'video', ops: ['extend'], t2v: false, i2v: false, adult: true, credits: { base: 15000, lora: 20000 }, tier: 'standard', weights: 'open' },
{ id: 'steady-dancer', label: 'SteadyDancer', kind: 'video', ops: ['motion'], t2v: false, i2v: false, tier: 'standard', weights: 'open' },
{ id: 'p-video-animate', label: 'P-Video Animate (fast)', kind: 'video', ops: ['motion'], t2v: false, i2v: false, tier: 'standard', weights: 'closed' },
{ id: 'dreamactor-v2', label: 'DreamActor v2', kind: 'video', ops: ['motion'], t2v: false, i2v: false, tier: 'standard', weights: 'closed' },
]