FLUX Klein 9B Inference¶

Practical reference for FLUX.2 Klein 9B image generation. Covers optimal sampler settings, multi-pass upscaling, LoRA stacking, anatomy fixes, and prompting patterns.

Distilled 9B (4-Step Model)¶

Recommended Settings¶

CFG above 2.0 produces "deep-fried" artifacts - the distilled model has guidance baked in. Using 50+ steps wastes compute and actively degrades quality.

Anatomy Fix: Use 8 Steps¶

At 4 steps, complex poses (seated, two-person) often produce extra limbs. 8 steps shows dramatic improvement for anatomy without excessive slowdown. Beyond 8 steps on distilled: diminishing returns.

Base 9B (Non-Distilled)¶

Recommended Settings¶

Why Base Over Distilled¶

• Higher output diversity
• Configurable quality/speed tradeoff
• Better for LoRA training and custom pipelines
• Better input for downstream upscalers

Sampler Deep-Dive¶

Key insight: res_2s at 4 steps equals euler at 8 steps in compute. Fixes anatomy without increasing step count explicitly.

Multi-Pass Upscaling¶

Base (Stage 1) + Distilled (Stage 2)¶

[Prompt] -> Klein Base 9B, 1024x1024, 20 steps, CFG 5.0, euler/simple
         -> Upscale Latent 2x -> 2048x2048
         -> Klein Distilled 9B, 4-8 steps, CFG 1.0, euler/simple, denoise 0.4-0.6
         -> VAE Decode -> 2048x2048 output

Fast Pipeline (Distilled Only)¶

[Prompt] -> Klein Distilled 9B, 1024x1024, 8 steps, CFG 1.0, euler/simple
         -> Image Resize -> 2048 longest side
         -> Klein Distilled upscale, 4 steps, euler_ancestral_cfg++,
            sgm_uniform, denoise 0.8, s_noise 1.2
         -> 2048x2048 output

Denoise Values for Second Pass¶

4K+ Output¶

Add tiled upscale stage: 4x4 grid, 256px tiles with 128px overlap, seam_fix_denoise=1.0, seam_fix_width=32-128px. Florence2 auto-caption before upscale reduces hallucinations at tile borders.

Seam Artifacts¶

Causes: tile boundary context loss, latent space discontinuity, high denoise in second pass.

Fixes: 128px overlap padding, mask blur 12-16px, half-tile reprocessing at boundaries, band-pass filtering (retain high-freq detail, blend low-freq).

LoRA Stacking¶

Capacity and Strength¶

Up to 3 LoRAs simultaneously, each with individual weight (0-4).

Multi-LoRA Rules¶

• Stack strongest-influence LoRA first (processed sequentially)
• Total combined strength should stay under 1.5-2.0
• If artifacts appear, reduce weakest LoRA first
• Watch for style conflicts (two different color grading LoRAs)

LoRA Training Parameters (from 50+ runs)¶

Klein vs Dev LoRA Behavior¶

Same LoRA settings produce fundamentally different results: - Klein: heavier grain structure (16mm film look) - Dev: cleaner, more like 35mm film - FP8 Klein: maintains desirable grain; non-FP8 is cleaner

Anatomy Problem Hierarchy¶

From most to least effective:

1. Increase steps (4 -> 8): strongest lever for complex poses
2. Simplify pose: standing > seated, single > multi-person, front > twisted
3. Adjust CFG carefully: 1.0 baseline, 1.2 can fix fused fingers, >1.5 risks new problems
4. Use res_2s sampler: doubles compute per step, fixes anatomy implicitly
5. Negative prompting: "distorted features, unnatural proportions, extra limbs"
6. Use Base model: 20 steps rarely has anatomy issues (much slower)

Face Blur Fix¶

• Increase steps (most effective)
• Higher resolution (1536x1920 for portraits)
• Flux2Klein-Enhancer node to boost text conditioning magnitude

Flux2Klein-Enhancer Node¶

Custom node for stronger prompt adherence:

Dampening 1.20-1.30 recommended for precise preservation.

Prompting Guide¶

Structure¶

Subject -> Setting -> Details -> Lighting -> Atmosphere

Write as flowing prose, not keyword lists. Front-load important elements.

Lighting (Highest Impact)¶

Specify: source type, quality, direction, temperature, surface interaction. "Soft, diffused natural light filtering through sheer curtains" >> "good lighting"

Prompt Length¶

Style Annotations (End of Prompt)¶

• "Style: Country chic meets luxury lifestyle editorial"
• "Shot on 35mm film with shallow depth of field"
• "Mood: Serene, romantic, grounded"

Native Resolutions¶

Both 4B and 9B support 11 aspect ratios up to 4MP (2048x2048 square). Range from 1:1 to 21:9.

Gotchas¶

• 4B LoRA incompatible with 9B: different text encoder sizes (Qwen 3-4B vs Qwen 3-8B). LoRAs are not interchangeable between Klein 4B and Klein 9B.
• FP8 specifically benefits from 8 steps: the FP8 quantized distilled model needs more steps than bf16 to match quality. Budget 8 steps minimum for FP8 inference.
• Qwen-Image VAE artifacts: the VAE decoder can introduce washed-out details and checkerboard noise. A dedicated fix LoRA exists (strength 1.0, trigger: "Remove compression artifacts. Restore the fine details of the photo."). Only works on Qwen-VAE artifacts - will degrade images from other VAEs.

Detail Enhancement LoRAs¶

Community-trained LoRAs specifically for increasing detail/realism in Klein 9B output.

Tier 1: General Detail¶

Tier 2: Skin-Specific¶

Stacking Strategy¶

For maximum detail: base realism LoRA (0.5-0.65) + detail slider (+2 to +5). For portraits: skin-specific (0.5-0.7) + general detail (0.4-0.5). Keep total combined influence moderate - artifacts appear when cumulative weight exceeds ~1.5.

Latent Space Detail Manipulation¶

Detail Daemon (ComfyUI-Detail-Daemon)¶

Modifies sigma schedule during sampling - keeps noise injection the same but lowers noise removal, effectively adding detail.

Large features form early, fine details late. Set start=0.3, end=0.8 to affect detail without breaking composition.

ComfyUI-Latent-Modifiers (Fooocus Sharpness)¶

Single mega-node with chained techniques. Most useful combo for Klein detail: Sharpness + Spectral Modulation.

• Sharpness: Fooocus-based, sharpens noise mid-diffusion. Strength 2.0-4.0.
• Spectral Modulation: converts latent to frequencies, clamps high frequencies while boosting low ones. Fixes oversaturation from high CFG.
• Divisive Norm: pooling to reduce artifacts from sharpness. Add after Sharpness.
• Tonemapping: 6 methods (Reinhard, Arctan, Quantile, Gated, CFG-Mimic, Spatial-Norm) for clamping oversaturation.

CFG Tuning for Detail¶

Distilled model: CFG must be 1.0 - higher values break generation entirely.

Tiled 4K+ Upscale Pipelines¶

Klein + SeedVR2 (Proven Two-Stage)¶

Stage 1: Klein 9B img2img to 2048x2048
  Sampler: euler_ancestral_cfg++
  Scheduler: sgm_uniform (BasicScheduler)
  Denoise: 0.8 (range 0.75-0.85)
  eta: 1.0, s_noise: 1.0-1.2
  Prompt: add "8K, intricate details"

Stage 2: SeedVR2 tile upscale to 4096x4096
  Grid: 4x4 = 16 tiles (~256x256 each)
  Default SeedVR2 parameters
  Variants: 3B (realism), 7B Sharp (stylized)

ControlNet Tile Upscale¶

For FLUX Klein, use Flux.1-dev-Controlnet-Upscaler as tile controlnet. ControlNet strength must stay below 0.7 - higher produces rough edge artifacts. Denoise 0.3-0.4. Tile size 1024x1024 (768x768 for low VRAM).

Florence2 Smart Tiling¶

Auto-captions image before upscaling. Caption guides diffusion to respect context during tile processing. Reduces hallucinations at tile borders. Can reach 4K-8K (45MP).

Prompt Engineering for Detail¶

Structure (BFL Official)¶

Subject -> Setting -> Details -> Lighting -> Atmosphere. Natural language prose, NOT tag lists. Front-load important elements.

Lighting descriptions have the single greatest impact on output quality. Be specific: "soft, diffused natural light filtering through sheer curtains" >> "good lighting."

Detail-Boosting Phrases¶

• "8K" + "intricate details" together: consistently produces strong results in upscale passes
• If too aggressive: remove "intricate details", keep only "8K"
• NEVER use the word "enhance" - model associates it with heavy AI upscaling artifacts
• Describe specific material properties: "silk with visible thread count", "leather with visible grain and patina"

Prompt Length¶

Model Variants (Full Catalog)¶

KV variant (9B-kv): caches key-value pairs for reference images during multi-image editing. 2.5× faster than vanilla 9B for reference-conditioned generation. Use when doing face swap, head swap, or reference-guided editing.

Official BFL LoRAs¶

Move LoRA: Visual Bounding Box Conditioning¶

Move objects by drawing red (source) and green (destination) rectangles directly on the image.

How it works: 1. Detection: Qwen3-VL-8B reads burned-in colored rectangles 2. Red rectangle = source location 3. Green rectangle = destination location 4. Klein 9B executes the move

# Workflow in ComfyUI:
# 1. Load image
# 2. Draw red box around object to move (paint tool or overlay)
# 3. Draw green box at destination
# 4. Load Move LoRA at scale 1.25 (fast mode) or 1.0 (quality mode)
# 5. Generate

# Prompt template:
prompt = f"Move the {object_description} to {destination_description}"

Two modes: - Fast mode: 8 steps, LoRA scale 1.25 - for quick iteration - Quality mode: 30 steps, LoRA scale 1.0 - for final output

Limitations: complex scenes with overlapping objects. When multiple objects overlap, Qwen3-VL-8B sometimes misidentifies which object is in the red box. Add explicit object description to prompt.

FLUX.1-Dev Adapter Incompatibility¶

ALL FLUX.1-dev adapters fail on Klein 9B due to dimension mismatch:

Klein 9B dimensions: hidden_dim=4096, joint_attention_dim=12288. A Klein-native adapter would require training a SigLIP→MLP projector (~85M params) on 50K-200K style pairs from scratch.

Gotchas¶

• 4B LoRA incompatible with 9B: different text encoder sizes (Qwen 3-4B vs Qwen 3-8B). LoRAs are not interchangeable between Klein 4B and Klein 9B.
• FP8 specifically benefits from 8 steps: the FP8 quantized distilled model needs more steps than bf16 to match quality. Budget 8 steps minimum for FP8 inference.
• Qwen-Image VAE artifacts: the VAE decoder can introduce washed-out details and checkerboard noise. A dedicated fix LoRA exists (strength 1.0, trigger: "Remove compression artifacts. Restore the fine details of the photo."). Only works on Qwen-VAE artifacts - will degrade images from other VAEs.
• Detail Daemon needs higher values on FLUX: SDXL uses detail_amount <0.25, FLUX needs 0.3-1.0+. Using SDXL values produces no visible change.
• "Enhance" in prompts causes artifacts: Klein associates this word with aggressive AI upscaling look. Use "8K" and "intricate details" instead.
• Steps above 50 degrade quality: Klein Base 9B is designed for 25-50 steps. Going higher produces overcooked results. For more detail, use multi-pass or LoRAs, not more steps.
• All FLUX.1-dev adapters incompatible with Klein 9B: hidden_dim mismatch (3072 vs 4096). No IP-Adapter, Redux, or USO works on Klein. Check adapter's target architecture before downloading.

See Also¶

• diffusion inference acceleration - Spectrum, Nunchaku quantization
• flow matching - underlying generation framework
• MMDiT - transformer architecture used by FLUX
• tiled inference - high-res output strategies
• diffusion lora training - training LoRAs for Klein
• lora fine tuning for editing models - training custom edit LoRAs