AIOS Creator

Local-first AI Creator Commerce OS by P2P Labs

Research → Create → Animate → Publish → Engage → Measure → Monetize → Learn.

The moat is not the model. The moat is Product + Character + Creative DNA + Distribution + Attribution.

github.com/joefebrian/AIOSCreator

---

What it is

AIOS Creator is a workstation OS for UGC and affiliate operators. You lock a character, import a SKU, generate on-model stills, animate, and push toward official publish APIs — mostly on your GPU, with optional paid cloud when a job needs it.

It is not a generic image playground. Characters, products, jobs, spend, and lineage stay in one loop.

Built for a single-owner RTX box (reference: RTX 3060 12GB). ComfyUI is the local inference runtime. Next.js is the control plane.

Who it is for

• Affiliate / UGC operators who need the same face on many SKUs
• Studios that want local generation without a per-clip SaaS meter
• Operators who mix local (Qwen / Klein / MiniMax H3) with paid cloud (GPT Image, Seedream, Wan 3.0, Kling)

Loop

Step	Module	Role
Research	Intelligence	Image/video → prompt, trends, opportunities, winner hooks
Product	Commerce	Catalog SKUs (Amazon, Shopee, Tokopedia, Lazada, TikTok Shop)
Create	Characters / Studio	Identity lock, stills, on-model, faceless pack, clone
Animate	Motion / Character I2V	Still → clip (H3 local, Wan / Kling / Seedance cloud)
Publish	Distribute	Official APIs or export pack. No ToS bypass.
Grow	Analytics	Measure, then mutate one variable

---

Features (what is live)

Intelligence

• Research — drop a still or clip, or fetch TikTok / YouTube / Instagram / X via yt-dlp. VL extract → photoreal prompt. Mega checklist from the same pass (no second paid generate). Cookies in System → Settings.
• Trends / Opportunities / Winner hooks — ranked from extracts and operator URLs.

Commerce

• Products — import listing URL or add SKU by hand. Photos saved to disk so Comfy can use them. Search, category, price, Amazon associate tag, localize remote images.
• Affiliate video lives on the character workspace, not Faceless VO.
• Programs / Campaigns are stubs that point back at the catalog.

Create

• Characters — GPT Image 2.5 for first identity (when keyed), then local Qwen Image Edit for keep-face / keep-body. Generate, Edit, Clone, aspect ratios (9:16, 3:4, 1:1, 16:9, 4:5).
• AI Studio — product hub: on-model (character + SKU), faceless pack, mixed keroyok. Save/load workflows, Autoflow, camera presets.
• MotionControl — still → clip. Engine picker first.
• ShortDrama — Script → Board → Generate → Export (one shot at a time).
• Faceless VO — UGC Factory scripts; VoiceStudio TTS when the local server is up.

Character workspace (image + video)

Ready: Generate image, Edit image, Affiliate video, Image to video, Complete set.
Not ready: storyboard / talking / extend / subtitles (shown as soon).

System

• Settings — LLM, image, and motion keys (never commit these).
• Models — which engine GEN uses.
• ComfyUI — GPU runtime health (/api/comfy/status), not nvidia-smi.
• Usage — cloud spend; monthly cap.

---

Engines

Local stills (ComfyUI, T0 12GB)

Engine	Use
FLUX.2 Klein 4B FP8	Fast identity / vibe
Z-Image Turbo	Photoreal “new person” plates
Qwen Image Edit 2511	Keep-face + body, new pose/scene (default edit)
Klein vibe + Qwen lock	2-pass: Klein scene, Qwen faceswap

GPT Image / Seedream block sexual content. Swimwear and NSFW go to Qwen / Klein / Z-Image. Local Comfy does not use that filter.

Local motion

Engine	Use
MiniMax H3 I2V	Default I2V + audio. Native 640×1152 / 20 step. Turbo LoRA v4 = 6 step at 576×1024 on 3060
H3 R2V	Copy motion from a reference clip (448×800)
Wan 2.2 5B / Hunyuan	Alternate local I2V when weights are installed
LTX-2	Avoid on 12GB (kills Comfy)

Cloud (keyed in Settings)

Engine	Use
GPT Image 2 / 2.5 Sunburst	First identity plates
Seedream 4.5 / 5 Pro	Cloud stills
Wan 3.0 std / prime	DashScope Singapore — WAN key is separate from Qwen LLM key
Kling 2.6 / 3.0, DreamActor V2	Paid motion / identity-lock video
Seedance 2.0 / 2.5	Paid I2V

Spend cap applies to all paid routes.

---

Stack

Layer	Choice
Control plane	Next.js 16 + React 19 + TypeScript + Tailwind 4
Local inference	ComfyUI (127.0.0.1:8188, --lowvram --reserve-vram 3 on 12GB)
Encode	ffmpeg + NVENC when present
Data	JSON under data/db (gitignored). SQLite path reserved
Queue	In-process jobs. One GPU owner — switching menus does not cancel
Hybrid	Cloud APIs optional; local worker always owns the box

NVIDIA: Studio Driver, NVENC for export. RTX Video Super Resolution is watch-only, not an export label.

---

Quick start

Requirements

• Windows 10/11, Node 20+, pnpm 11
• Python + ComfyUI under %LOCALAPPDATA%\Programs\ComfyUI
• ffmpeg on PATH (or FFMPEG_PATH)
• Optional: Tailscale for LAN HTTPS

App

git clone https://github.com/joefebrian/AIOSCreator.git
cd AIOSCreator
copy .env.example apps\web\.env.local
cd apps\web
pnpm install
pnpm dev

Open http://127.0.0.1:3000 (listens on 0.0.0.0:3000 for LAN).

From repo root:

pnpm dev

ComfyUI

.\scripts\start-comfyui.ps1 -Detach

Full stack (app + Comfy, does not kill other remote-desktop tools):

.\scripts\start-stack.ps1

GPU status in the header is Comfy HTTP, not VGA. GPU Off means :8188 is down.

First operator loop

1. Commerce → Products — paste a listing URL or Add SKU manually and upload pack shots.
2. Create → Characters — lock identity (GPT 2.5 if keyed, else local).
3. Open the character workspace → Generate / Edit stills (pick aspect). For swimwear/NSFW use Qwen Image Edit.
4. Affiliate video or Image to video (H3 local, Wan cloud if keyed).
5. Keep paid jobs under System → Usage spend cap.

---

Environment

Copy .env.example → apps/web/.env.local. Never commit .env.local.

Variable	Purpose
LLM_API_KEY / LLM_BASE_URL / LLM_MODEL	Chat / revamp (any OpenAI-compatible API)
COMFYUI_URL	Default http://127.0.0.1:8188
FFMPEG_PATH	ffmpeg binary if not on PATH
CREATOROS_DATA_DIR	Override data root (default ../../data from apps/web)
AMAZON_ASSOCIATE_TAG	Optional; also save in Products UI
DASHSCOPE_API_KEY	Qwen LLM (Alibaba)
DASHSCOPE_WAN_API_KEY	Wan 3.0 video only — separate key, Singapore
OPENAI_API_KEY	GPT Image (Character_Generation)

Prefer System → Settings for live keys. Env is fallback.

.gitignore excludes .env, .env.* (except .env.example), data/db/, data/media/, data/tmp*.

---

Repository layout

apps/web/          Next.js control plane
apps/worker/       Reserved Python worker
data/db/           Jobs, characters, products (gitignored)
data/media/        Stills, motion, product shots (gitignored)
docs/              PRD notes
scripts/           Comfy, stack, downloads, smokes
workflows/         Comfy API graphs

---

Hardware tiers

Tier	VRAM	Intent
T0	8–12GB	Images + short quantized video (this repo’s 3060 path)
T1	16GB	Default production 720p 5–8s
T2	24GB	Longer MotionControl / overnight
T3	32GB+	Long queues
TCloud	none	Cloud/hybrid adapters only

H3 I2V turbo on 12GB: 576×1024, 6 steps. Native 20-step can stay 640×1152. Do not mix nude LoRA with turbo. Do not upgrade PyTorch to cu130 unless the driver is 580+.

---

Product rules (short)

1. Do not train a new foundation model.
2. ComfyUI infers. AIOS Creator owns objects, rights, jobs, routing.
3. Affiliate is provider-agnostic. Amazon first; Shopee / Tokopedia / TikTok Shop URLs import too.
4. Publish via official APIs only, else export a pack.
5. One GPU owner. A running job stays on the box if you change pages.
6. Commercial stills should use license-safe checkpoints (Klein 4B / Qwen / Wan Apache where applicable). Klein Base 9B is research / non-commercial.

---

License and company

© P2P Labs — p2plabs.asia

Internal operator software. Checkpoints keep their upstream licenses (Apache-2.0, OpenAI API ToS, DashScope, Flux non-commercial, etc.). Do not redistribute weights from this repo; they are not vendored.

---

Status

Active workstation build. Some nav items (Engagement, Revenue, Growth Engine, several video tools) are parked or stubbed on purpose. The production path that is meant to work today:

Product catalog → Character lock → On-model still → H3 / Wan I2V.