Skip to main contentThe Local Bench

Use case 11 of 13

A private image studio

Generate and edit illustrations, product shots and concept art with node-based workflows, without a subscription or a content filter you do not control.

Paper swatches, a sketch board and a brush with a blue handle next to a graphics card
Image: generated with Grok Imagine

The real example

ComfyUI with FLUX.1

ComfyUI is the open-source, node-based interface most local image work runs on. Its official FLUX examples cover FLUX.1 [dev] and the distilled four-step FLUX.1 [schnell], and suggest FP8 weights to roughly halve memory use when you are short. Licences differ: FLUX.1 [schnell] is Apache-2.0, while FLUX.1 [dev] and Kontext [dev] carry Black Forest Labs' non-commercial licence. On GenEval, the HiDream-I1 model card reports 0.83 for HiDream-I1 against 0.66 for FLUX.1 [dev], a vendor-reported comparison.

Sources: ComfyUI on GitHub · ComfyUI FLUX examples · HiDream-I1 model card

Why run it locally

Unlimited generations for the cost of electricity, your own LoRAs trained on your own style, and client work that stays on your disk.

Recommended stack

Three tiers built from our builds. Weights-in-memory is our arithmetic (parameters × bits), checked against each device's usable memory; it excludes KV cache and runtime overhead. Runtime compatibility is not validated.

Budget

Value 16 GB

$1,173Sample

16 GB on GeForce RTX 5060 Ti 16 GB

Weights ≈ 11.9 GB of 15.5 GB usableEstimate

Excludes KV cache and runtime overhead. Runtime compatibility is not validated.

FLUX in FP8 fits 16 GB. Load SDXL instead when you want its deep ecosystem of add-ons.

Runtime: ComfyUI

Mid

Used-3090 24 GB

$1,980Sample

24 GB on GeForce RTX 3090 24 GB (used)

Weights ≈ 11.9 GB of 23.5 GB usableEstimate

Excludes KV cache and runtime overhead. Runtime compatibility is not validated.

Room for FLUX plus ControlNets and LoRAs; load Kontext [dev] for editing. Mind the non-commercial licence.

Runtime: ComfyUI

Pro

Dual-GPU 48 GB

$4,044Sample

48 GB on 2× RTX 3090 (used) 48 GB

Weights ≈ 37.5 GB of 47.0 GB usableEstimate

Excludes KV cache and runtime overhead. Runtime compatibility is not validated.

Large 17–20B models; Qwen-Image is Apache-2.0 and good at text inside images.

Runtime: ComfyUI

Setup outline

  1. Install ComfyUI (portable build on Windows, or Python venv on Linux and macOS).
  2. Download the model, text encoders and VAE into the folders the FLUX example page lists.
  3. Drag an example image into ComfyUI to load its workflow, then generate.
  4. Switch to FP8 weights if you run out of memory.

Limits

  • Licences vary by model; commercial use needs checking for each one.
  • GenEval measures prompt following, not aesthetics or likeness.
  • Image models are compute-hungry. Bandwidth matters less here than raw GPU compute.

Next: Private translation for documents and speech

Build prices come from our build datasheets: GPUs and devices are sourced retail or sold prices; other parts are sample prices until sourced, and the flag on each total says which. Model links go to our datasheets, where each parameter count and licence was checked against the Hugging Face API.