Open-source image generation models compared: licence, VRAM and which GPU

The open image model I would start with is Z-Image Turbo if the licence has to be clean and the GPU small, FLUX.2 [klein] 4B if I also need editing, and Qwen-Image-2512 or Qwen-Image-Edit-2511 when quality and text rendering matter more than speed. All four are Apache-2.0. The differences I weigh first are licences and memory, because a quality ranking depends on your prompts. Several popular models carry terms people miss: FLUX.1 [dev], FLUX.2 [dev] and the 9B [klein] models are non-commercial, Qwen-Image-2.1 is for research only, Krea 2 and Stable Diffusion 3.5 are free only below USD 1 million in revenue, and Tencent's Hunyuan Image licence does not apply in the EU, the UK or South Korea. Every model in the table fits a 48 GB card in 8-bit form except FLUX.2 [dev] and HunyuanImage-3.0.

The open image models, compared#

The table lists the open models with real use, checked at their own repositories and model cards on 2026-11-12. The file totals are what ComfyUI loads, the diffusion model, text encoder and VAE, summed from the Hugging Face byte sizes (our calculation): "16-bit" is everything at full precision, and "8-bit" is the fp8 or INT8 files Comfy-Org publishes, with the template's text encoder.

A few of those figures need a word. HiDream-I1's 16-bit total uses the fp8 T5 and Llama encoders, because those are the only forms Comfy-Org publishes. HiDream-O1's ComfyUI template also holds an optional Gemma 4 E4B prompt enhancer, 9.06 GB in fp8, switched off by default and left out of its totals. FLUX 3 Image Klein's maker says to always run its prompt upsampler, so both its totals include one: the 16-bit upsampler in the first column, and the 4-bit one ComfyUI's template loads in the second. HunyuanImage-3.0 has no native ComfyUI support, and every other model in the table has. MiniMax-H3-Image, released on 10 November 2026, and Pruna's P-H3-Image, which carries the MiniMax H3 Community License, are not in the table. Both licences define their "Excluded Territories" as "the European Union, the United Kingdom, the Republic of Korea and the United States of America" and grant rights "solely within the Applicable Territory", and every QuantaCloud region is in the US. MiniMax-H3-Image's licence also asks for written authorization above USD 1 million in annualized revenue or for any hosted service, and the H3 licence above USD 20 million in yearly revenue.

What the makers and ComfyUI's docs say about memory is thinner than the table, and the makers' figures come from their own reference code. Tongyi says Z-Image Turbo fits a 16 GB card, Stability says SDXL works on 8 GB and SD 3.5 Medium needs 9.9 GB without its text encoders, and Tencent gives 24 GB for HunyuanImage-2.1 at 2,048 by 2,048. ComfyUI's own docs say HiDream-I1 needs more than 27 GB at full precision and more than 16 GB in fp8, and that Qwen-Image in fp8 used 86% of a 24 GB card. Black Forest Labs gives a 31 GB peak for FLUX 3 Image Klein in BF16 with the upsampler off, plus about 28 GB for the upsampler. ComfyUI streams weights between system RAM and the GPU, so it runs most of these in less memory than the totals, and more slowly when it has to.

What each licence allows#

The licence decides more than the benchmark, because an image that is fine for a hobby project can be off-limits for your company. The licence travels with the weights, whichever mirror you download them from.

Two clauses catch people. Krea 2's revenue limit applies to the images as well as the model, unlike the FLUX non-commercial licences, which let you use the images commercially. And a workflow's licence is the sum of its files: a LoRA trained on a non-commercial model, or a text encoder with its own licence, brings its terms into your output. Read the licence on each model card before you build a product on it, and treat this table as a summary, not legal advice.

What fits a 48 GB card, and what wants 96 GB#

The rule I follow is to compare the 8-bit and 16-bit totals with the memory budget of each card, in GiB: 41.39 GiB on a 48 GB card with ECC on, 73.60 GiB on an A100 80GB, 87.95 GiB on the RTX PRO 6000 Blackwell and 129.17 GiB on an H200 NVL, each 0.92 of what the card reports (our calculation).

Architecture changes which file to pick on the same memory. The Ada, Hopper and Blackwell cards compute fp8 natively, while the Ampere RTX A6000 and A100 do not. There ComfyUI expands the layers of an fp8mixed or fp8_scaled file to 16-bit as it computes, and it converts a plain fp8 file to its 16-bit size as it loads whenever that copy fits under about 88% of the card's memory. The 8-bit files of Qwen-Image-2512, HiDream-I1 and HunyuanImage-2.1 are plain fp8, so on an RTX A6000 their diffusion models can take twice their 8-bit size, and a 48 GB Ada card suits them better. INT8 files run natively on every NVIDIA GPU in ComfyUI's code, which makes them the 8-bit choice on an RTX A6000. A second model loaded beside the first is the best reason to rent more memory: Qwen-Image-2512 and Edit-2511 in fp8 share a text encoder and VAE and come to 47.1 GiB together, which a 48 GB card cannot hold at once and an 80 GB or 96 GB card can.

The H100 PCIe and H200 NVL run these workflows when the catalog lists them. HunyuanImage-3.0 needs a multi-GPU configuration when the catalog lists one, or a server we build to your spec: send a capacity brief and we quote it in writing. The RTX PRO 6000 Blackwell page lists its configurations with RAM and disk, and ComfyUI GPU requirements covers the same arithmetic with the video models added.

Which model for which job#

The honest answer depends on three questions, in this order: may you use the model for your business, does it do what you need, and does it fit the GPU.

Each family has its own guide with the files, folders and settings: FLUX.1, FLUX.2, Qwen-Image, Qwen-Image-Edit and Z-Image Turbo. For SDXL outside ComfyUI, Forge in Docker runs it in a browser UI, and training a FLUX LoRA covers training on a cloud GPU.

ComfyUI support and the template's version#

Every model in the table except HunyuanImage-3.0 runs natively in ComfyUI, with no custom nodes. The catch is the ComfyUI version: ComfyUI's template index gives each model's templates a minimum, from 0.0.1 for SDXL to 0.37.2 for Qwen-Image-2.1 and 0.41.0 for FLUX 3 Image Klein. ComfyUI itself was at v0.45 on 2026-11-12, and the version inside QuantaCloud's ComfyUI template is not published yet. Check it before you download the files:

curl -s http://127.0.0.1:8188/system_stats | python3 -c 'import json, sys; print(json.load(sys.stdin)["system"]["comfyui_version"])'

If it is older than the model, install a current ComfyUI yourself on the Bare Metal template, as running ComfyUI on a cloud GPU shows. GGUF files exist for most of these models, but on a 48 GB card the native fp8 and INT8 files do the same job without a custom node: GGUF models in ComfyUI explains when a quant earns its place.

Questions about open image models#

What is the best open-source image generation model?

There is no single answer that holds for every prompt, so I choose on licence and memory. For commercial work that means Z-Image Turbo, FLUX.2 [klein] 4B or Qwen-Image-2512, all Apache-2.0. FLUX.2 [dev] is the largest open FLUX model, but its licence is non-commercial.

Which open image models can I use commercially?

FLUX.1 [schnell], FLUX.2 [klein] 4B, Qwen-Image and its Edit models and Z-Image under Apache-2.0, HiDream under MIT, and SDXL under its RAIL++-M licence. Stable Diffusion 3.5 and Krea 2 are free below USD 1 million in annual revenue, and FLUX 3 Image Klein allows commercial images below USD 5 million. Hunyuan Image allows it outside the EU, the UK and South Korea.

What is a good alternative to FLUX?

For a commercial licence, Qwen-Image-2512 and Z-Image Turbo, both Apache-2.0. Qwen-Image is the stronger choice for text in images, and Z-Image Turbo the faster one, at 8 steps and about a third of the files at 16-bit, 20.69 GB against 57.70 GB.

Can I run these models locally on my own GPU?

Most of them, yes: SDXL, SD 3.5 Medium, Z-Image Turbo and FLUX.2 [klein] 4B are within reach of a 16 GB consumer card by their makers' own figures, and ComfyUI's offloading stretches further at the cost of speed. FLUX.2 [dev], Qwen-Image at 16-bit and HunyuanImage-3.0 are where a rented 80 GB, 96 GB or multi-GPU machine starts to make sense.

My decision rule: start with Z-Image Turbo or FLUX.2 [klein] 4B on a 48 GB RTX A6000, move to Qwen-Image-2512 and Edit-2511 when you need text or editing, and rent the RTX PRO 6000 Blackwell when two models must share the GPU or you run FLUX.2 [dev]. Read the licence before you pick anything with a revenue limit. Each instance starts with an empty disk and stopping it deletes the disk, so keep a list of model URLs and copy your images off before you stop. Open-source video generation models is the same comparison for video, and the ComfyUI page lists every GPU with the template and its live price.