Self-hosted image generation
Local image generation with ComfyUI can pair SDXL or compatible quantized model workflows with your own hardware. Verify model and workflow requirements before committing; disk budgeting, custom-node security, and an honest comparison with hosted services are equally important.
Updated July 11, 2026
The models
Compare model documentation, license terms, supported hardware, and workflow requirements before choosing a model. Quantized variants can be worth evaluating where they are supported.
| Model | VRAM | Speed | Notes |
|---|---|---|---|
| SDXL 1.0 | Check the model documentation | Depends on hardware and workflow | A widely used baseline with a broad LoRA ecosystem. |
| FLUX.1-dev | Check the model and quantization documentation | Depends on hardware and workflow | A model family to evaluate for prompt adherence and photorealism. |
| Qwen-Image | Check the model and quantization documentation | Depends on hardware and workflow | Evaluate it when text rendering inside images matters. |
| SD 3.5 | Depends on the variant | Depends on hardware and workflow | Stability's current open-weight line. |
Interfaces to evaluate
Choose an interface based on the workflow you want to run: node-based composition, a guided application experience, or generation integrated into a creative tool. Confirm current model compatibility and maintenance before adopting a workflow.
A node-based runtime for composing and running local image-generation workflows. Review the workflow, required nodes, and supported hardware before adopting it.
The app-like alternative: model management, a gentler learning curve, and release notes full of stability work rather than new node types. Pick it if you want to make images rather than build graphs.
A web interface to evaluate when its workflow and model support fit your needs. Check its current documentation and maintenance activity before building around it.
Generation inside a real painting app: inpainting, live canvas, and layers, backed by a ComfyUI server underneath. The strongest option when images are a step in an illustration workflow rather than the end product.
A Python-library option to consider when image generation belongs inside your own code rather than behind a UI.
Custom nodes are a supply chain, treat them like one
ComfyUI's superpower is its node ecosystem, and the ecosystem is community Python that runs with your user's permissions. Treat custom nodes and their dependencies as untrusted code until you have reviewed their source, maintenance, and installation behavior.
The defenses are unglamorous. Keep the instance off the public internet or behind authentication, keep the runtime and dependencies updated, and read what a node does before installing it, especially anything that installs dependencies at runtime. Review imported workflows and their required nodes before running them.
Budget disk before VRAM
Model files, quantized variants, VAEs, text encoders, and LoRAs can consume substantial storage. Plan for the models and workflows you intend to keep, and use storage that keeps model loading practical for your workflow.
The hosted math, honestly
Compare current hosted pricing with your hardware, volume, and operating time. Self-hosting can make sense for iteration-heavy work (inpainting rounds, LoRA training, batch pipelines), existing hardware, and images you would rather not upload. If that GPU also serves a language model, note that an image model wants the card to itself while generating; the two share hardware, not VRAM. The self-hosted ChatGPT stack covers that side of the machine.
Questions people actually ask
Can I run FLUX on an 8 GB card?
Quantization and CPU offload can reduce memory requirements, but performance depends on the model, workflow, and hardware. Test a small workflow on your card before committing. A lighter model family may be more practical when memory is limited.
Is a used RTX 3060 12 GB enough?
It can be a workable starting point for local image generation. Confirm the memory requirements of your chosen model, quantization, resolution, and workflow, and leave room for the rest of the system.
Which UI should a beginner start with?
Start with the interface that matches how you work. Choose an app-like UI for a guided experience, or ComfyUI when you want to compose and reuse node-based workflows. Check project maintenance and model support before investing in a workflow.
Is self-hosting cheaper than Midjourney?
It depends on your current hosted plan, hardware, volume, and time spent operating the stack. Self-hosting can make sense for iteration-heavy workflows, existing hardware, and images you do not want to upload. Control is usually the more durable reason than price.
Sources
Every project above is one row in the Awesome Open Source AI registry, which resyncs twice a day from the curated GitHub list. The Generative Media Tools category carries the neighbors: FLUX itself, video generation, and the audio side.
