Local Models • Oct 20, 2026

Local AI Image Generator: FLUX, ComfyUI, and Offline Routes

A local AI image generator loads open weights on hardware you control. Nano Banana 2.1 does not do that. This page separates the documented FLUX.1-dev route from the 8GB and Mac cases this site already covers.

People searching for a local AI image generator usually want one of three things: a model whose weights run on their own GPU, a ComfyUI workflow, or a way to generate images on a Mac or an 8GB card. Those are different constraints. This guide uses only the distinctions already published here and on the linked model cards.

Local weights are not a local API client

Nano Banana 2.1 is Google's hosted image model gemini-nano-banana-2.1. Google's image-generation documentation recommends it for new projects and describes 1K, 2K, and 4K output. Nano Banana 2 remains gemini-3.1-flash-image. In both cases the Python program can sit on your computer, including a Mac, while the model runs on Google's servers. That client is not an offline generator and there are no Nano Banana weights to install.

An offline route is a separate open-weight project. This site documents several, with different licenses and hardware notes: FLUX.1-dev, Qwen-Image-Edit-2511, HiDream-O1-Image, Z-Image, and HunyuanImage 3.0. FLUX 3 is not one of them. The quickstart lists FLUX 3 as early access without public weights or a local deployment path.

FLUX.1-dev in ComfyUI and Diffusers

The FLUX.1-dev model card describes a 12 billion parameter rectified flow transformer from Black Forest Labs. It generates images from text. The card says output quality is second only to FLUX.1 [pro], and that the model was trained with guidance distillation. The Hugging Face repository is gated: you accept the license before the files download.

Two local routes are named on that card:

  • ComfyUI. The card says FLUX.1-dev is available in ComfyUI for local inference with a node-based workflow. It does not publish a node graph, a Mac build, or a minimum VRAM figure for that workflow.
  • Diffusers. The documented example installs Diffusers, loads FluxPipeline in bfloat16, and calls enable_model_cpu_offload() to move weights to CPU when GPU memory is tight. The example uses a 1024×1024 output and 50 steps.
pip install -U diffusers
import torch
from diffusers import FluxPipeline

pipe = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-dev", torch_dtype=torch.bfloat16)
pipe.enable_model_cpu_offload()  # save some VRAM by offloading the model to CPU

image = pipe(
    "A cat holding a sign that says hello world",
    height=1024,
    width=1024,
    guidance_scale=3.5,
    num_inference_steps=50,
    max_sequence_length=512,
    generator=torch.Generator("cpu").manual_seed(0)
).images[0]
image.save("flux-dev.png")

Editing is not a feature of this checkpoint. The local model comparison treats FLUX.1-Fill-dev, FLUX.1-Redux-dev, FLUX.1-canny-dev, and FLUX.1-Depth-dev as separate FLUX.1 checkpoints, not as features built into FLUX.1-dev. Prompt recipes for background replacement and product shots live on the prompts page; which model can run them depends on the checkpoint you actually load.

The model card places FLUX.1-dev under the FLUX.1 [dev] Non-Commercial License. It also says generated outputs can be used for personal, scientific, and commercial purposes as described in that license. Read the license before a commercial deployment. The local model comparison puts this next to HunyuanImage and Qwen-Image-Edit-2511.

What 8GB of VRAM is documented to cover

The hardware guide gives an NVIDIA RTX 3060 with 8GB of VRAM and 16GB of system RAM as the floor for SDXL-class generation and basic inpainting. That figure is not a FLUX.1-dev minimum. The FLUX.1-dev card does not state one. CPU offload in the Diffusers example is a way to reduce GPU memory, not a promise that an 8GB card is enough.

Other published hardware notes on this site are wider than 8GB:

  • Qwen-Image-Edit-2511 is a 20B BF16 model. Its model card does not publish one universal VRAM minimum. The card's Diffusers example is the thing to measure.
  • HunyuanImage 3.0 is in a different class. The hardware guide, citing Tencent's model table, lists at least 3 × 80GB for the base checkpoint and at least 8 × 80GB for Instruct and Distil.

If a number is missing from the model card, this page leaves it out. The VRAM guide explains why parameter count, precision, and offload change the result.

Mac: what is documented, and what is not

A Mac can run the Nano Banana API client. The hardware guide's cloud-client verdict includes a MacBook Air, because that path needs no GPU. The image is still generated in Google's cloud.

For weights on the machine itself, the sources diverge:

  • The Qwen-Image-Edit-2511 card mentions switching the device to "mps" for Apple devices. That is a device string in the example, not a measured memory budget.
  • The HunyuanImage 3.0 repository documents CUDA 12.8 and does not describe a macOS path.
  • The FLUX.1-dev card documents ComfyUI and Diffusers. It does not document a Mac requirement or a Mac install.

Pick the next page

Sources