Flux models for local use
Flux is a family of open-weight image models from Black Forest Labs, the second family covered on the image models index and one of the main alternatives to the older Stable Diffusion lineage for generating images locally. The team behind it came out of the group that built the original latent diffusion work, so the family arrived with a mature architecture rather than as a first attempt.
What it is known for is prompt adherence. Flux tends to follow long, compositional prompts more literally than earlier families, and it handles rendered text inside images considerably better, which used to be a reliable weak point for local models. The cost of that is a heavier model with higher memory expectations.
Variants are released under different terms depending on which one you pick, and those terms matter if the output is going anywhere commercial. Check the licence attached to the specific weights you download rather than assuming the family shares one.
Sizes and what they fit
The family ships in several variants that trade quality against cost. They differ in weight size and in how many sampling steps they need, and both affect what running one feels like.
- The full-quality variants are the memory-hungry end. Base weights plus the text encoders are a large working set, and on a discrete card this is what decides whether you can run it at all. These are aimed at cards with generous VRAM.
- The distilled and reduced-step variants are designed to produce an image in far fewer steps. They are the practical choice when throughput matters or when you are iterating on prompts rather than producing finals.
- Quantized weights are widely distributed for this family, and they are how most people fit it onto consumer hardware. A lower-precision copy shrinks the working set substantially with a modest quality cost. See picking a quantization for how to choose a level.
A 24 GB card such as the RTX 4090 runs the family comfortably, including full-precision variants at ordinary resolutions, with headroom for adapters. Below that, quantized weights and offloading some components to system memory are the normal approach, and both work, with offloading costing speed. The GPU pages cover where individual cards sit.
Apple silicon machines like the M3 Max benefit here. A large unified memory pool can hold the full working set where a discrete card of similar price cannot, so capacity stops being the limit and generation time becomes the thing you plan around. Expect a slower step rate than a desktop GPU.
Runtimes
- ComfyUI was where local support for this family landed first and it remains the most complete. The node graph makes it straightforward to swap encoders, apply quantized weights and manage offloading explicitly, all of which you are more likely to need here than with lighter families.
- SD WebUI supports the family, though the interface was shaped around an older pipeline and some family-specific options are less directly exposed.
- InvokeAI supports it within its canvas workflow, which suits editing and regional work on generated images.
Whichever you pick, the family needs its text encoders present as separate files alongside the main weights. A download that looks incomplete usually is: see reading a model card.
Ecosystem
The community body of work around Flux is younger than the Stable Diffusion one but growing steadily, and it leans in a different direction.
- LoRA adapters dominate. Because the base weights are large, distributing a small adapter is far more practical than distributing full fine-tuned checkpoints, so style and subject work here is mostly adapter-shaped.
- Quantized redistributions are a category in their own right. Much of the practical local usage runs on weights someone else converted and published.
- Control and conditioning modules exist for the family, though the selection is narrower than what has accumulated around older architectures.
Adapters are family-specific and often variant-specific. A file built for one family will not load against another, and one built for a distilled variant may behave oddly on a full one.