AI Creator Studio
Private AI media generation studio: image generation + workflow editing + LLM-assisted prompting + photo library + object storage. GPU-mandatory for production workloads.
What this stack is
AI Creator Studio is the private alternative to Midjourney + DALL-E + Runway for self-hosted image and visual generation. One-click deploys a node-based workflow editor (ComfyUI), a simplified one-click image generator (Fooocus), an LLM for prompt assistance (Ollama), a photo library (Immich), and S3-compatible asset storage (MinIO).
The customer gets unlimited image generation at the cost of their VPS — no per-image credits, no content policies that block what you can create, no terms of service that claim rights over the outputs.
This stack is GPU-mandatory. Pricing is "Contact for pricing" pending the GPU partner agreement currently in negotiation.
What it's for
- Marketing agencies producing client imagery without per-asset SaaS bills
- Indie game studios generating concept art, sprite variations, environment textures
- E-commerce sellers generating product mockups, lifestyle shots, A/B test creatives
- Architecture and interior design firms iterating on visualization concepts before commissioning full renders
- Content creators producing thumbnails, social graphics, illustrated blog posts at scale
- AI artists running production-grade workflows (ControlNet, IP-Adapter, LoRA stacks) that cloud platforms cap or rate-limit
Who it's for
- Creative agencies with steady image-generation needs that exceed $50/mo Midjourney bills
- Bootstrapped product companies wanting unlimited creative iteration without burn rate impact
- AI hobbyists turned semi-pro who outgrew Replicate / RunPod / Vast.ai pay-per-minute pricing
- Privacy-conscious clients (legal cases, internal R&D, sensitive product designs) needing on-prem generation
- Resellers serving the creator economy — bundle this stack with the [[stacks-creator-stack|Creator Stack]] as a complete "private creator workspace"
Apps in this stack
| App | Role |
|---|---|
| ComfyUI | Node-based workflow editor — full power for advanced users (Stable Diffusion, FLUX, SDXL, ControlNet, IP-Adapter, custom LoRAs) |
| Fooocus | One-click image generation — simplified UI for users who don't want to learn workflows |
| Ollama | Local LLM for prompt assistance — helps users write better prompts, generates variations |
| Immich | Photo library — organizes generated outputs, face/object recognition, sharing |
| MinIO | S3-compatible storage — large model files, output archive, asset versioning |
Sizing & deployment
- Recommended VPS: 16 GB RAM, 8 vCPU, 250 GB NVMe storage
- GPU MANDATORY — minimum 12 GB VRAM for usable SDXL workflows; 16-24 GB VRAM for production (RTX 4090, RTX 6000 Ada, L40S, A100 class)
- Model storage: 50-100 GB just for checkpoints (SDXL Base, FLUX.1, custom fine-tunes, LoRAs)
- Deploy time: ~30-40 min one-click (slow due to model downloads on first install)
- Backup profile: nightly Immich library + MinIO bucket replication + ComfyUI workflow JSON exports
Why hosters sell this
The creator economy spends billions on image generation credits every year — Midjourney, DALL-E, Adobe Firefly, Runway, Pika. A single agency burns $100-500/mo per creator. This stack lets the hoster offer unlimited generation at flat-rate VPS+GPU pricing, capturing customers when they hit the credit-cost wall.
Common resale shape:
- Indie creator tier — single VPS, GPU spot or shared, hobbyist workflows
- Studio tier — dedicated VPS + dedicated GPU, production workflows, multi-user
- Agency tier — dedicated infra + GPU + multi-tenant for client isolation + white-label
Stack retail pricing depends entirely on the GPU partner agreement (Frame C) — currently in negotiation. Contact Bluix sales for a current cost-build model.
Why this is GPU-mandatory
ComfyUI + Fooocus generate images via diffusion models. On CPU, a single 1024×1024 image takes 3-10 minutes; on a GPU it takes 5-15 seconds. CPU-only deployment is functionally unusable for any commercial creator workflow. Unlike the LLM stacks (where CPU-only is a viable lower tier), image generation needs GPU as a hard floor.
Compliance profile
- Generated outputs stay on the VPS — no third-party API processes the customer's prompts or images
- No content policy restrictions beyond what the customer enables locally (the customer is the content owner, not Bluix or any cloud provider)
- Model provenance — the hoster ships only open-source / open-weight models (Stable Diffusion XL, FLUX.1 Schnell, etc.) with documented training data provenance
- Right-to-be-forgotten trivially supported (delete bucket + library entries = full removal)
- EU AI Act compliance for generators classified as "general-purpose AI" — model cards documented, deployment region picked at provisioning