Flux 1.1 Best Settings Guide (Steps, CFG, Resolution, Sampler)

Flux 1.1 Best Settings Guide (Steps, CFG, Resolution, Sampler)

Dora

Dec 4, 2025

Hi, I'm Dora. If you’re buried under client revisions, looming deadlines, and a never-ending content queue — breathe. By the time you finish this guide (yes, in one sitting), you’ll walk away with battle-tested Flux 1.1 settings that spit out lightning-fast drafts, jaw-dropping photoreal finals, and posters where the text is actually readable on the first try. Just copy-paste parameters, proven two-pass workflows, and ready-to-go presets for portraits, products, and typography that I use daily with paying clients. AI moves fast — these settings are locked in and verified as of December 2025.

Flux 1.1 Core Settings Explained

The right defaults do 80% of the work. Think of Flux's controls like a camera rig: guidance is your focus ring, steps are your exposure time, and the sampler is the shutter behavior. Counter-intuitively, I found that modest guidance with slightly longer steps yields the most realistic textures without melting details.

Core parameters to set first

  • Resolution: Start at 768–1024 on the short side for drafts: 1216–1536 for finals if your VRAM allows.

  • Aspect Ratio: Match the subject: 4:5 or 3:4 for portraits, 1:1 or 4:3 for products, 2:3 or 18:24 for posters.

  • Steps: 18–24 for drafts: 28–36 for finals. More than 40 rarely helps and may oversmooth.

  • Sampler: DPM++ SDE or Euler a for speed: UniPC or DPM++ 2M SDE for clean edges and typography.

  • Guidance (CFG): 3.0–4.0 for photorealism: 4.5–6.0 for text-heavy layouts to anchor lettering.

  • Seed: Lock for iteration: change when exploring variability.

  • Denoise strength (for refiner/upscale pass): 0.2–0.35 to preserve composition while sharpening.

  • Negative prompt: Use targeted terms (e.g., "extra fingers, mutated text, deformed logo, oversoften skin").

Suggested baseline (draft)


{

"resolution": "896x1152",

"sampler": "DPM++ SDE",

"steps": 20,

"guidance": 3.5,

"seed": 12345

}

Suggested baseline (final)


{

"resolution": "1344x1792",

"sampler": "DPM++ 2M SDE",

"steps": 32,

"guidance": 3.8,

"refiner_denoise": 0.25,

"seed": 12345

}

Why these work (evidence + reasoning)

  • Using the prompt: "natural light product shot, matte ceramic mug on linen, 50mm, f/2.8, soft rim light," Flux 1.1 at guidance 3.5 produced fabric weave and subtle rim highlights: increasing guidance to 6.5 flattened textures. This happens because higher guidance constrains the model toward tokens instead of microstructure.

  • For the prompt: "bold poster, ‘COASTAL CYCLE TOUR' in Futura Bold, teal and coral, grid layout," raising steps from 22 to 32 reduced character warping: the longer diffusion path helps stabilize letterforms.

For deeper reference, see the official Black Forest Labs documentation, the Flux technical report on arXiv (explains why low-to-mid guidance preserves microstructure), and my earlier breakdown Prompt Engineering for Photorealism on Z-Image.

Promise: you'll get useable concepts in minutes and studio-quality finals in one extra pass. Problem

  • Drafts are fast but noisy: finals are slow. The trick is a two-pass pipeline that holds composition while improving fidelity. Prerequisite
  • Prepare one locked seed, one base prompt, and a short negative prompt. Step-by-step solution
  • Create draft
  • Set Resolution to 896×1152 (or 1024 square).
  • Use Sampler: DPM++ SDE, Steps: 20, Guidance: 3.5.
  • Keep Seed fixed. Generate 4 images.
  • Select composition
  • Pick the strongest layout/pose. Note any text errors.
  • Final render pass
  • Increase Steps to 32–36.
  • Switch Sampler to DPM++ 2M SDE or UniPC.
  • Upscale to long side 1536–1792.
  • Apply Refiner / High-Res with Denoise 0.2–0.3: keep the same seed.
  • Add targeted corrections to the prompt (e.g., "clean serif letters, tight kerning"). Result
  • Expect sharper micro-textures, cleaner hair strands, and notably better letter shapes with minimal composition drift. Verification methodology
  • Re-run the same prompt with a fixed seed across step counts 20, 28, 32. Compare edges and texture detail at 200% zoom. Document differences. This standard A/B keeps claims grounded.

Flux 1.1 Best Portrait Presets (Skin, Lighting & Detail)

Goal: lifelike skin, controlled lighting, accurate features without plastic smoothing. Preset A, Natural daylight beauty

  • Use when you want editorial realism and gentle contrast.
  • Settings
  • Resolution: 1152×1536
  • Sampler: UniPC
  • Steps: 32
  • Guidance: 3.3–3.8
  • Refiner Denoise: 0.25
  • Prompt recipe

"portrait of a 28-year-old woman, natural window light, 85mm, shallow depth, subtle freckles, neutral makeup, soft catchlight, color-accurate skin"
  • Negative prompt: "waxy skin, airbrushed, over-sharpened pores, extra fingers."

  • Why it works: moderate guidance keeps skin tone transitions analog-smooth: UniPC plus longer steps preserves hair strands and eyelash detail.

Preset B, Cinematic tungsten + rim

  • Use for mood, richer color, and crisp edges.

  • Settings

  • Resolution: 1216×1600

  • Sampler: DPM++ 2M SDE

  • Steps: 34

  • Guidance: 3.8–4.2

  • Refiner Denoise: 0.22

  • Prompt recipe


"cinematic portrait, tungsten key with cool rim, 50mm Zeiss, high microcontrast, film grain 200, dramatic shadows"
  • Tip: add "subsurface scattering" to help ears and noses feel luminous rather than flat.

Input + output + rationale

  • Using Preset A with the phrase "subtle freckles," the model produced pore-level texture without plastic sheen. This is the detail that changes the outcome: keeping guidance under ~4 lets the diffusion path explore fine tonal transitions instead of clamping to token priors.

Where portraits can fail

  • If you need pixel-accurate likeness of a specific person, Flux 1.1 alone may not suffice. Consider a reference-controlled pipeline (e.g., Control-style adapters) or a licensed face dataset. For medical/dermatology use-cases, stick to expert-reviewed imagery.

Product Photo Presets

Goal: clean edges, believable materials, and consistent branding colors.

Preset C, Tabletop softbox

  • Settings

  • Resolution: 1152×1152

  • Sampler: DPM++ SDE

  • Steps: 28

  • Guidance: 3.6–4.0

  • Refiner Denoise: 0.25

  • Prompt


"studio product photo, matte ceramic mug on seamless white, softbox left, fill right, crisp shadow, color-accurate glaze, 70mm, f/8"
  • Notes: keep background minimal: add "true white BG" and "no banding" in negatives.

Preset D, Lifestyle on-location

  • Settings

  • Resolution: 1152×1536

  • Sampler: UniPC

  • Steps: 30–32

  • Guidance: 3.4–3.8

  • Prompt


"coffee mug on linen tablecloth, morning window light, soft rim, shallow depth, natural grain, muted palette"
  • Why it works: lower guidance allows natural variation in textiles and wood grain: steps ~30 keep edges believable.

Color and logo tips

  • For brand colors, explicitly name Pantone or hex in the prompt (e.g., "brand teal #1ABC9C"). Validate in post.

  • Logos: Flux can imply logos but isn't vector-precise. If you need print-ready marks, composite in Illustrator afterward.

Poster & Typography Presets

Text requires extra anchoring to reduce letter drift and garbled glyphs.

Preset E, Bold headline poster

  • Settings

  • Resolution: 1536×2048 (portrait) or 2048×1536 (landscape)

  • Sampler: DPM++ 2M SDE

  • Steps: 34–36

  • Guidance: 4.5–5.5

  • Refiner Denoise: 0.2

  • Prompt


"modern poster, headline: 'COASTAL CYCLE TOUR', Futura Bold, tight kerning, grid layout, teal and coral, high contrast, clean margins"
  • Negative prompt: "mutated letters, warped kerning, random symbols, double strokes."

  • Technique: add structure words like "grid," "baseline," "margin." They act like a layout scaffold during diffusion.

Preset F, Small text with subhead

  • Settings

  • Resolution: 1792 on long side

  • Sampler: UniPC

  • Steps: 36

  • Guidance: 5.0–6.0

  • Prompt


"Swiss-style poster, headline: 'URBAN RUN', subhead: '5K SAT 9AM', grotesk type, left-aligned, generous leading, white space"
  • Workflow add-on: Run a second pass with denoise 0.15 to tighten edges while preserving layout.

Evidence & rationale

  • At guidance ~5.2, character shapes become more stable, trading a touch of texture richness for legibility, a fair swap for posters. Validate by zooming to 300% and comparing stem symmetry.

Who this is NOT for

  • If you need vector-perfect typography or brand-approved type setting, generate the layout as a visual comp in Flux, then recreate final type in InDesign/Illustrator. Flux is excellent for concept exploration, not final print typography.

Low-VRAM Settings

Running on a laptop GPU or older card? Prioritize stability and composition over brute resolution.

Settings and workarounds

  • Resolution: 768×1024 or 832×1216.

  • Batch size: 1.

  • Precision: Enable half-precision/FP16 in your UI if available.

  • Sampler: Euler a or DPM++ SDE for speed.

  • Steps: 18–24 (draft) → 28 (final single image).

  • Tiled high-res: If your UI supports tile upscaling, enable it with small Tile size (e.g., 256) and Overlap (64) to avoid seams.

  • Refiner denoise: 0.2–0.25 to sharpen without re-composing the scene.

Practical routine

  • Draft at 832×1216 with steps 20.

  • Lock seed: save the best.

  • One final at 1216×1664, steps 28–32, sampler UniPC if memory allows.

Where it fails

  • Very small text and dense patterns (like herringbone at distance) can break. Generate larger, then downscale in post. When memory is truly limited, accept simpler compositions.

Reference: Civitai + Forge low-VRAM optimization bible for Flux (6–8 GB tested) – the exact settings I use on my laptop.

Use these as drop-in starting points and adjust guidance/steps within the ranges above.

Recipe 1, Natural portrait


Settings:
- 1216x1600, UniPC, 32 steps, guidance 3.6, refiner 0.25, seed locked

Prompt:

"editorial headshot, soft north-window light, 85mm, subtle freckles, lifelike skin tone, gentle catchlights"

Negatives:

"waxy skin, over-smooth, extra fingers, harsh sharpening"

Recipe 2, Cinematic product on white


Settings:
- 1152x1152, DPM++ SDE, 28 steps, guidance 3.8, refiner 0.25

Prompt:

"studio product photo, wireless earbuds on seamless white, softbox left, crisp shadow, realistic plastic texture, color-accurate"

Negatives:

"dirty background, banding, warped logo"

Recipe 3, Poster with headline + subhead


Settings:
- 1536x2048, DPM++ 2M SDE, 36 steps, guidance 5.2, refiner 0.2

Prompt:

"modernist poster, headline: 'SPRING NIGHT RUN', subhead: '10K APR 12', grotesk type, grid layout, tight kerning, balanced margins, teal on charcoal"

Negatives:

"mutated letters, off-axis kerning, random symbols"

Recipe 4, Low-VRAM lifestyle


Settings:
- 832x1216, Euler a, 20 steps, guidance 3.4: final 1216x1664 @ 28 steps

Prompt:

"sunlit kitchen scene, ceramic mug on linen, soft rim light, shallow depth, muted colors, natural grain"

Negatives:

"noisy fabric, blown highlights, banding"

How to validate your results

  • Keep the seed fixed and change one variable at a time (steps, guidance, sampler). Export a contact sheet. Record which change measurably improved skin texture, edge sharpness, or text legibility.

Ethical Considerations

  • Transparency: Label AI-assisted imagery in captions or credits. If a client piece blends photos and Flux renders, disclose the methodology in your project notes.

  • Bias mitigation: Use diverse prompt terms (skin tones, ages, body types) and review outputs for stereotypical patterns. If a batch trends toward biased defaults, adjust descriptors and rebalance your dataset or references.

  • Copyright/Ownership (2025 best practices): Confirm license terms of any training add-ons or reference images. Avoid mimicking living artists by name. For logos and trademarks, treat Flux renders as comps: recreate marks with licensed vector assets before delivery.

Where Flux 1.1 settings fall short

  • Exact brand typography and vector logos: use the model for concepting, then finish in Illustrator/InDesign.

  • Scientific accuracy: for medical, architectural, or safety-critical visuals, rely on expert-designed renders or photographs.

  • Real-time iteration at print resolutions: upscale externally and proof colors with a calibrated pipeline.