Open Image Model Directory

How to read this table

Weights is the full download at BF16: the image model, the text encoder and the VAE combined. It is usually much larger than people expect, because the text encoder is a language model in its own right — T5-XXL is roughly 9.8GB before you have loaded any image weights at all.

Peak / Min VRAM brackets what the model actually costs to run. Peak is BF16 with every component resident, which is the fastest configuration. Min is NF4 with sequential offload, which is the smallest — and considerably slower, because weights stream over PCIe on every step. Most people land somewhere between: FP8 with the text encoder on CPU.

Commercial is the one that catches teams out. “Open weights” is not the same as a permissive licence, and two of the strongest models here cannot be used in a product without a separate agreement.

Running these locally

ComfyUI is where most of these end up, and it is unusually good at running a model your card cannot hold outright — it will offload the text encoder, or move components through VRAM one at a time, at a cost in speed. The ComfyUI VRAM page gives the peak memory for every model here at each precision and offload mode.

Common questions

Which open image model should I use?

For most people, SDXL. It runs at 1024x1024 on a 12GB card, has by far the largest ecosystem of fine-tunes and LoRAs, and its text encoders are small. Move to FLUX.1 [schnell] if you want current-generation quality and can spare the VRAM — it is Apache 2.0, so you can ship with it, and it needs only 1-4 steps.

Which ones can I use commercially?

FLUX.1 [schnell] and Qwen-Image are Apache 2.0, which permits commercial use outright. FLUX.1 [dev] is non-commercial. The Stable Diffusion models carry OpenRAIL or Stability Community terms, which permit commercial use but attach use restrictions that travel with the model.

Why do these need so much more VRAM than the file size suggests?

Because a diffusion pipeline is three models, not one: the image model, a text encoder and a VAE. For Flux and SD 3.5 the text encoder is T5-XXL, which is around 9.8GB on its own. Add activation memory, which scales with resolution and batch size, and peak usage runs well above the headline weight figure.

What is the difference between a UNet and a diffusion transformer?

Architecture. SD 1.5 and SDXL use UNets, the original latent-diffusion design. SD 3.5, Flux and Qwen-Image use diffusion transformers, which scale better with parameters and handle text rendering and prompt adherence noticeably better, at the cost of size.

Are these VRAM figures measured?

The weight sizes are exact, summed from the real safetensors files in each model's Hugging Face repository. Activation memory is an approximation, because it varies with the attention implementation, sampler and pipeline. Plan with headroom rather than to the last gigabyte.

Comparing the models, running them, and the checkpoint, LoRA and licence questions that follow.

• Best Local Image Generation Models in 2026: VRAM and Licences Compared15 open image models on memory and licence, with picks for 8, 12, 16 and 24GB.Read the guide
• How to Run Stable Diffusion Locally: Hardware, VRAM and SetupCurrent models from SD 1.5 to Qwen-Image, and the VRAM each needs.Read the guide
• MOSS-Image in ComfyUI: Native Nodes, Files, Memory and SettingsThe Apache 2.0 model, the nightly-only native nodes and the noise trap.Read the guide
• Stable Diffusion Checkpoints, LoRAs and VAEs ExplainedWhat each file is, how big it is, and where it goes.Read the guide
• Stable Diffusion LoRA: Base Models, Weights and StackingMatching the base model, setting strength, and combining adapters.Read the guide
• Flux LoRA: Which Base Model You Can Legally Train OnA LoRA is a Derivative; only klein 4B and schnell are Apache-2.0.Read the guide