Self-Host Open Source AI Image Generators

Self-hosting an open source AI image generator used to mean owning a 24GB card and accepting that anything smaller was a toy. Hardware, licensing and inference-step counts all pushed in the same direction, which is why most people ended up paying an API instead.

That changed this year. Z-Image-Turbo generates in 8 steps inside 16GB of VRAM under Apache 2.0, FLUX.2 [klein] 4B runs in about 13GB under the same license, and Krea 2 Turbo fits a 16GB card as a 13.1GB FP8 file. The short answer: run Z-Image-Turbo if you ship commercial work on a 16GB card, and Krea 2 if your company earns under $1 million a year and you want the best-looking output.

Updated October 4, 2026. Added Krea 2 and Qwen-Image-2.1, and corrected three entries: FIBO's weights are non-commercial, the FLUX.2 dev Q4 file is 19.3GB rather than 7GB, and HunyuanImage 3.0 needs a multi-GPU server. Forge, untouched since July 2025, is no longer the recommended UI.

Why Self-Host?

A local AI image generator keeps prompts and outputs on your own machine, costs only electricity per image once the GPU is paid for, and opens the whole Hugging Face and Civitai ecosystem of fine-tunes, LoRAs and ControlNets instead of a vendor's curated menu. The price is setup time and a license you now have to read yourself. Four of the ten checkpoints below are non-commercial, and three more carry revenue or territory conditions.

Model Comparison

Licenses, file sizes and download figures below were checked against each model card and the Hugging Face API on October 4, 2026. Downloads are the trailing 30 days.

The Models

1. Z-Image-Turbo (Tongyi-MAI)

Z-Image-Turbo is the model that changed the hardware calculus, and still the most-liked open image model in this guide at 5,402 likes. It is a 6B parameter model distilled down to 8 function evaluations, which is where the speed comes from, and its card claims sub-second latency on an H800 while fitting comfortably within 16GB of consumer VRAM.

It is also Apache 2.0, unlike FLUX.2 dev. Unsloth's GGUF build pulled 675,000 downloads in the trailing 30 days, more than the official repo.

It is strongest on photorealism and bilingual text rendering, English and Chinese. The tradeoff is the one every distilled model makes: eight steps gives up some of the control a 30-step sampler offers. Tongyi-MAI released the undistilled base, Z-Image, in January 2026 for fine-tuning.

• License: Apache 2.0
• Parameters: 6B, 8 inference steps
• Model hub: Tongyi-MAI/Z-Image-Turbo

2. FLUX.2 (Black Forest Labs)

FLUX.2 is still the benchmark for output consistency in open image generation. FLUX.2 [dev] is a 32B open-weight model that generates and edits at up to 4MP and can reference up to 10 images at once for character, product or style consistency, per Black Forest Labs' launch post.

32B parameters is the catch. The BF16 weights are 64.4GB, and even city96's Q4_K_S GGUF is 19.3GB, so plan on a 24GB card. The smallest build, Q2_K, is still 12.9GB. ComfyUI and SwarmUI both support it natively.

Bash
Copy
# FLUX.2 dev Q4_K_S GGUF: 19.3GB download, plan on a 24GB card
hf download city96/FLUX.2-dev-gguf flux2-dev-Q4_K_S.gguf \
  --local-dir ./models/unet/

The hf CLI ships with pip install huggingface_hub. Add --dry-run to confirm the file and size before committing to a 19GB download.

• License: FLUX Non-Commercial (dev). Commercial use needs a license from Black Forest Labs
• Model hub: huggingface.co/black-forest-labs

3. FLUX.2 [klein] (Black Forest Labs)

FLUX.2 klein is Black Forest Labs' compact, distilled family for consumer hardware and sub-second generation, unifying text-to-image, editing and multi-reference in one architecture.

Read the license per variant, because they differ. The 4B is fully open under Apache 2.0, and its card says it fits in ~13GB of VRAM on an RTX 3090 or 4070 and above. The 9B is under the FLUX Non-Commercial License, the same terms as FLUX.2 dev, despite sharing the klein branding. If you picked klein to get out from under those terms, 4B is the variant that does that.

Adoption has tilted toward the open one: 395,000 downloads in 30 days for the 4B against 189,000 for the 9B.

• License: Apache 2.0 (4B) / FLUX Non-Commercial (9B)
• Parameters: 4B or 9B, sub-second inference
• Model hub: FLUX.2-klein-4B

4. Krea 2 (Krea AI)

Krea 2 is the newest serious contender, released in June 2026 as two checkpoints: Raw, an undistilled base for LoRA training, and Turbo, an 8-step distilled model for inference. Krea's advice is to train on Raw and run on Turbo, since LoRAs carry over. SwarmUI's model guide puts it bluntly: "Krea 2 is the champion of image generation."

The model is about 13B parameters, so the BF16 checkpoint is 26.3GB. Comfy-Org repackaged it as a 13.1GB FP8 file that fits a 16GB card, plus a 7.7GB NVFP4 build for tighter memory. It uses Qwen 3 VL 4B as its text encoder and has a built-in content filter.

The license is the part to read. The Krea 2 Community License allows commercial use only if your company's total annual revenue is under $1,000,000. Above that you need an enterprise license from Krea before any commercial use.

• License: Krea 2 Community (commercial use under $1M annual revenue)
• Parameters: ~13B, 8 steps (Turbo)
• Model hub: krea/Krea-2-Turbo

5. Qwen-Image-2512 and Qwen-Image-2.1 (Alibaba)

Qwen-Image-2512 is the December 2025 update to Alibaba's 20B Qwen-Image, and the specialist for photorealism and legible in-image text: signage, UI mockups and product labels in English and Chinese. It is Apache 2.0, and unsloth's Q4_K_M GGUF is 13.2GB, which puts it on a 16GB card.

Qwen-Image-2.1 arrived in September 2026, and it and its community GGUF builds now fill most of Hugging Face's trending text-to-image list. It is smaller (a 7B generation transformer), unifies generation and editing with up to 10 reference images, and generates native transparent RGBA images. The BF16 transformer is 14.2GB and the text encoder another 17.5GB, so a 16GB card needs the CPU offload the card documents.

Do not assume 2.1 inherits the Apache license. It ships under the Qwen Research License, which permits non-commercial use only; commercial use needs a separate license from Alibaba. For client work, stay on 2512.

• License: Apache 2.0 (2512) / Qwen Research, non-commercial (2.1)
• Model hub: Qwen/Qwen-Image-2512, Qwen/Qwen-Image-2.1

6. HunyuanImage 3.0 (Tencent)

HunyuanImage 3.0 is an 80B Mixture-of-Experts model with 13B parameters active per token. That size buys reasoning over long, layered prompts: complex spatial relationships and multi-element scenes that smaller models pattern-match badly.

It is not a workstation model. About 83B parameters is roughly 168GB of BF16 weights, and the card's setup defaults to four GPUs. It is also missing from ComfyUI's supported list, so you run Tencent's own inference code.

The license is the Tencent Hunyuan Community License, which does not apply in the European Union, the United Kingdom or South Korea and requires a separate license above 100 million monthly active users.

• License: Tencent Hunyuan Community (excludes EU, UK, South Korea)
• GitHub: github.com/Tencent-Hunyuan/HunyuanImage-3.0

7. Stable Diffusion 3.5 Large (Stability AI)

Stable Diffusion 3.5 Large is the oldest model here, from October 2024, and SwarmUI's model list now labels it outdated. It stays for the largest community ecosystem of any open model: fine-tuned checkpoints, LoRAs and ControlNet adapters for nearly everything. It is also the lightest; the 8B model's Q4_0 GGUF is 4.8GB, so an 8GB card runs it.

• License: Stability AI Community License (free commercial use under $1M annual revenue)
• Model page: huggingface.co/stabilityai/stable-diffusion-3.5-large

8. FIBO (Bria AI)

FIBO is an 8B model trained on long structured JSON captions, so lighting, composition, color and camera settings can be set programmatically and version-controlled like any other config. It is trained exclusively on licensed data, its pitch for IP-sensitive work. But the open weights are CC BY-NC 4.0, non-commercial; commercial use goes through Bria, so the clean-data story only pays off once you have that license.

• License: CC BY-NC 4.0 (weights); commercial license from Bria
• HuggingFace: huggingface.co/briaai/FIBO

Hardware Requirements

VRAM is the hard physical constraint. A practical rule of thumb for Q4 quantization is approximately 0.5 to 0.7GB of VRAM per billion parameters, and the GGUF files above bear it out: FLUX.2 dev's 32B Q4_K_S is 19.3GB, Qwen-Image-2512's 20B Q4_K_M is 13.2GB. The same memory math applies to text models.

For Apple Silicon users: M-series chips use unified memory, so a Mac Studio with 64GB can hand most of it to the model. Our GPU vs Mac comparison covers the throughput gap.

Recommended UI Frameworks

The models above run through self-hosted UI frameworks. Two are worth installing in October 2026, and one is not:

ComfyUI - node-based workflow editor, most capable, steepest learning curve, 136,000 GitHub stars as of October 4, 2026. It is first to support new releases, and its README lists FLUX.2, Z-Image, Qwen Image, Krea 2 and SD 3.5 among supported models. This is the default.

SwarmUI - a simpler tab-based interface that runs ComfyUI as its backend, inherits its model support, and spreads generation across multiple GPUs. Its model support doc is the best plain-English guide to which files to download.

Forge - was the easy-setup recommendation in this guide's first version. Its last commit was July 31, 2025, before FLUX.2, Z-Image or Krea 2 existed. Fine for SDXL-era checkpoints, not for this list.

Where to Find Models

Hugging Face holds base models and official releases; filter by Text-to-Image and sort by trending. Civitai is the largest collection of community fine-tunes and LoRAs.

On a card under 24GB, search the model name plus GGUF or FP8. City96, unsloth and Comfy-Org publish quantized builds for most models above, and they often out-download the official repos.

Which Open Source AI Image Generator Should You Use?

The best open source AI image generator depends on your license constraints first and your GPU second.

• Commercial work on a 16GB card: Z-Image-Turbo, or FLUX.2 klein 4B if you need editing and multi-reference. Both are Apache 2.0 with no revenue cap.
• Best-looking output at a company under $1M revenue: Krea 2 Turbo in FP8.
• Text in images and product mockups: Qwen-Image-2512. Use Qwen-Image-2.1 for research, or after you license it.
• An existing LoRA or ControlNet workflow, or an 8GB card: SD 3.5 Large.
• Research and personal projects: FLUX.2 dev for consistency, HunyuanImage 3.0 if you have the servers.

All of them keep prompts and outputs on hardware you control. If you want a self-hosted video generator as well, the same ComfyUI install runs open video models, covered in our AI video generation models guide.

Related DevToolLab Tools

• EXIF Viewer - ComfyUI and AUTOMATIC1111 embed the full prompt and workflow into generated PNGs. Check what is actually in a file before you publish it.
• EXIF Remover - Strip that metadata out, since sharing a generation otherwise shares your entire prompt chain and node graph.
• PNG to WebP - Generated PNGs at 4MP are large. WebP cuts the delivered size without a visible quality drop.
• Image Compressor - Batch down a folder of generations before they go anywhere near a web server.

Related Guides

• Local LLM Hardware: GPU vs Mac - the same hardware decision for text models
• Best GPU Cloud Providers - where to rent the multi-GPU server HunyuanImage 3.0 needs
• Best AI Image Generators - the hosted services these models compete with
• Best AI Video Generation Models - where this tooling goes next

Licenses and model availability change quickly. Verify the license on the model card before shipping commercial work.