Local Image Generation in 5 Minutes: VRAM Decides, Flux Wants CFG 1.0

Local image generation starts in 5 minutes, but VRAM (graphics-card memory) decides what you can run, and Flux uses different settings than SDXL. There are two beginner answers: easiest is Fooocus (click only, SDXL, even 4GB), recommended is Forge (low-VRAM, Flux support), and most powerful with day-one new-model support is ComfyUI (2-3 hours to learn). For models, SDXL is a measured ~7-8GB, and FLUX.1 dev is 23GB at full precision but runs even on 8GB via GGUF quantization (Q5_K_S keeps ~95% quality). And download only .safetensors, not .ckpt - the old format can execute code on load.

In one line: the tool is the type of camera. Fooocus is a point-and-shoot (press and it appears), Forge is a hybrid (easy but adjustable), and ComfyUI is a manual DSLR (does everything but you must learn it). You pick the lens (model) to fit your bag (VRAM).

First, the terms. VRAM is GPU-dedicated memory where the model and intermediate computation live. CFG (guidance scale) is how strongly the image follows the prompt, and a step is one iteration of the repeated noise-removal that builds the image.

What can your VRAM do (tools and models)?#

8GB is comfortable with SDXL and Flux via quantization, 16GB runs almost everything, and 24GB runs it all unquantized. The key trap is grabbing the quantized T5 encoder for Flux - the fp16 T5 alone is 9GB and will not fit in 8GB. On low VRAM, enable --lowvram (a measured 20-30% slower) and tiled VAE decode (the VAE step spikes VRAM and can crash even after the model loaded fine). If your card is not fast, 4-step models (SDXL Lightning/Turbo, Flux schnell, Klein 4B) are the rescue.

Why does Flux use different settings than SDXL?#

Because Flux has guidance built in, so the old defaults are wrong. SDXL's standard is CFG 5-7, negative prompts, and DPM++ 2M Karras at 20-30 steps, but Flux needs CFG set to 1.0 (higher just oversaturates). Negative prompts are ignored, the sampler is euler + simple, and dev's safe default is 20-30 steps. Not knowing this one line is the most common beginner complaint ("Flux looks weird"). If speed is urgent, 4-step distilled models turn tens of seconds per image into a few on the same card.

Where do beginners get stuck?#

Three things: VRAM, the encoder, and the VAE spike.

• VRAM: even after the model loads, the VAE decode can spike and crash - use tiled VAE.
• Encoder: Flux needs the quantized (smaller) T5 to fit in 8GB (the fp16 T5 is 9GB).
• First run: the first image is slow due to loading, so measure speed from the second image, and for safety download only .safetensors.

How do you do it in 5 minutes?#

Step through the easiest path in order and you can pull your first image within five minutes.

• On 8GB, render a 1024x1024 image with Forge plus an SDXL checkpoint (or Fooocus).
• If you use Flux, start with CFG 1.0, empty negative, euler/simple - those are the defaults.
• Vary only steps and sampler on the same prompt to see the speed/quality curve, and drop to a 4-step model if it is slow.

Reference links#

• ComfyUI (node-based generator)
• Stable Diffusion WebUI Forge (beginner pick)
• Fooocus (easiest SDXL)
• ComfyUI-GGUF (low-VRAM quantization)
• FLUX (Black Forest Labs repo)

Note: VRAM, time-per-image, and quality figures are public 2026 measurements and vary by GPU, resolution, quantization, and sampler. Measure exact speed on your own device with the method above (the first image is slow due to loading). Download only .safetensors, and since models and tools update often, this is reviewed quarterly.