Local Image-Generation (Diffusion) Model Database

Local large language models get the attention, but a genuinely private AI stack can also see and create — and image generation follows the same hard rule as the rest of the sovereign stack: the model you can actually ship is decided by its licence before you ever judge its quality. Text-to-image is a licensing minefield, because the two best-looking open models — FLUX.1 [dev] and Stable Cascade — are non-commercial, and so is NVIDIA’s ultra-fast Sana. This database scores 15 self-hostable open-weight diffusion model families on licence commercial-safety first, then parameter count, architecture, minimum VRAM (at full precision and quantized), native resolution and speed — so you can find the model you can both afford to run and legally deploy.

The licence decides what you can ship

Only a handful of these models are cleanly commercial-safe. FLUX.1 [schnell] is Apache-2.0 — personal, scientific and commercial use are all permitted — which is exactly why it, and not its sibling FLUX.1 [dev], is the one you build a product on. Qwen-Image and Lumina-Image 2.0 are Apache-2.0 too, and HiDream-I1‘s transformer is MIT. The Stable Diffusion family sits in the middle: SD 1.5, SDXL and PixArt-Sigma ship under CreativeML OpenRAIL, which permits commercial use but attaches enforceable use-based restrictions, while SD 3.5 uses Stability’s Community License — free to use commercially only while your organisation is under USM in annual revenue. And then the dead ends: FLUX.1 [dev] (non-commercial model licence), Stable Cascade (an explicit nc-community licence) and Sana (NVIDIA’s research-only NSCL) look fantastic but cannot go into a commercial product. FLUX.1 [pro] is API-only closed weights, so it is not in this self-hostable table at all.

How much VRAM do you actually need?

Image models range from trivially cheap to data-centre-hungry. Stable Diffusion 1.5 runs on a 4 GB card and SDXL on 8 GB, which is why the SD ecosystem still dominates hobby rigs. The 12B FLUX models are a different story: they want roughly 24 GB at full bf16 precision — but a GGUF Q4 quant plus offloading the T5-XXL text encoder brings that down to about 8 GB, so they run on a mainstream gaming GPU with patience. The 2025-era flagships (20B Qwen-Image, 17B HiDream-I1) are genuinely large and lean on quantization to fit consumer hardware. Every VRAM figure in the table is a practical whole-pipeline estimate, not a vendor benchmark, so treat it as a starting point and confirm against your own setup. To size the GPU itself, cross-reference the AI & local-inference GPU database.

Where this fits in the local-AI stack

Image generation is one modality of a sovereign, self-hosted AI stack that never sends your data to a US cloud. For the models that generate text, see the local LLM model database; for the ones that hear and speak, the local voice-AI database; and for turning documents into searchable vectors, the local embedding-model database. To pick the runtime that serves any of them, the local AI runtime comparison and the quantization references, and to shop the hardware, the GPU database. Everything ties back to the Local LLM Canada hub and the wider Open Mining & AI Data catalogue, where this dataset is downloadable as CSV and JSON under a CC BY 4.0 licence.