FLUX.2 [klein] 9B Base LoRA

The text description of what to generate.

Tip: If your LoRA used a trigger word, include it. Describe the scene around the subject normally.

Input images for editing. The edit LoRA decides how they are transformed.

What the model should avoid generating.

How strictly generation follows the prompt (classifier-free guidance).

In the atelier: How tightly you hold the painter to the brief. Too tight and the work gets stiff and oversaturated; too loose and he wanders.

Tip: Stay near the endpoint default. Adjust in steps of 0.5.

Random seed. Same seed plus same inputs gives a nearly identical image.

Tip: Fix the seed when comparing LoRA scales or parameters, so the only thing changing is the thing you are testing.

Number of denoising steps per image. More steps, more refinement, more latency.

Tip: Defaults are tuned per model. Cutting steps in half is the quickest way to trade quality for speed.

Output resolution, as a preset name or explicit width and height.

How many images to generate in one call.

Speed and cost preset for the generation backend.

Tip: 'regular' balances quality and latency. Use higher acceleration for drafts and iteration.

Returns media as a data URI instead of a hosted URL, and skips storing it.

Tip: Useful for privacy; awkward for big files. Most workflows leave it off.

Runs a safety filter on outputs.

File format of the result: jpeg, png or webp.

Enables Colored Noise Sampling, an experimental stochastic sampler. The cns_* knobs below only matter when this is on.

Tip: Leave off unless you are deliberately experimenting. Defaults are research-grade settings.

How much stochastic (colored) noise is injected after each step. 0 = CNS off, i.e. pure deterministic flow-matching (identical to normal klein); higher = more stochasticity. Turn this first; the other CNS parameters only matter when s_churn > 0.

Where the per-frequency progress signal gamma(f, t) comes from. 'approximation' is self-contained but uniform across bands, so the frequency shaping comes entirely from the alpha-tilt parameters. 'matrix' uses the precomputed per-band gamma for true spectral-bias-aware coloring (more noise to genuinely under-resolved bands).

Exponent on the residual (1 - gamma). Higher concentrates noise more aggressively on unresolved bands; lower is gentler and flatter. No effect under 'approximation', where gamma is uniform.

Divides gamma first, weakening the coloring (pushing it toward uniform white noise). Paper uses 1.73 for distilled/unguided and 25.0 for CFG/base, so raise this for the /base (CFG) endpoints.

Frequency tilt at the start of sampling, interpolated toward the end value over steps. Positive boosts high frequencies (fine texture/detail); negative boosts low frequencies (structure).

Frequency tilt at the end of sampling. Positive boosts high frequencies (fine texture/detail); negative boosts low frequencies (structure). The default favors low-frequency structure at the end.

Use a smooth, frequency-position-weighted tilt (exp(alpha * f)) instead of a flat (1 + alpha) multiplier.

Use an exponential (vs linear) curve for the start-to-end tilt transition across sampling steps.

How sharp the exponential tilt-interpolation curve is. Only used when exponential interpolation is enabled.

How many radial frequency bands the spectrum is split into. More bands = finer-grained coloring.

Global multiplier on the injected colored noise after variance normalisation. <1 slightly damps the kick, >1 amplifies it (paper: 0.98 unguided, 0.998 guided).

List of LoRA weights to load, each with a path (URL or HF repo) and a scale.

In the atelier: Which bracelets the painter wears for this painting, and how hard he leans on each.

Tip: Order does not matter; scales do. Start every LoRA at 1.0 and adjust one at a time.

Watch out: Stacking three strong LoRAs usually degrades all of them. Lower each scale when combining.