FLUX 3 Image Klein tested - 16GB Local AI setup

Summary

Luke’s Dev Lab evaluates FLUX.3 Image Klein, a 6-billion-parameter image-generation model from Black Forest Labs, using an expanded suite of 55 prompts. The video is a benchmark and review, not a setup tutorial.

Test setup

• Hardware: RTX 2000 Ada with 16 GB of VRAM and a 70 W power limit. No RTX 5090 was used.
• Settings: 16 steps, BF16, usually seed 42, and mostly 1280 × 1280 output.
• Test duration: About one hour. Most generations took roughly 52 seconds. Thumbnail and wallpaper tests used wider aspect ratios and were faster, while editing and multi-reference tasks generally took longer. One vehicle-generation test took an unusually long 452 seconds.

What was tested

The suite covered photographic and anime images, people and anatomy, motion, lighting, weather, art styles, objects and materials, reflections, perspective, game assets and interfaces, web and mobile designs, text, object counts, spatial layout, detailed scenes, seed variation, image editing, and censorship.

Findings

• Strong overall prompt following: The model handled many scenes, layouts, materials, and stylistic prompts well. Notable successes included rain scenes, oil and ink painting styles, glass and metal objects, mirror reflections, simple text, game HUD elements, and several designed compositions.
• Image editing was a standout strength: The model performed well when changing a hat, removing an object or background, changing clothing, outpainting a scene, and incorporating two or three reference images. The edits were generally convincing and consistent, though faces and lighting sometimes shifted.
• People and anatomy were mixed: Group scenes and ordinary hand anatomy were often good, but some eyes looked strange, and the anime yoga image had an obvious anatomy problem. A fox gained an extra leg; other misses included an awkward tennis-racket grip and a flawed storyboard door.
• Text worked reasonably well in simple cases: Single words, a short phrase, and digits were rendered well. Longer or incidental text in designs was less reliable and could become nonsense.
• Counting and complex details remained unreliable: A request for five apples produced six. A detailed classroom prompt included most requested elements, but the clock was wrong and one student was awkwardly positioned.
• Some images retained an AI-like finish, including overly shiny or glowing skin. The reviewer also observed that small eyes tended to lack detail.
• Censorship test: The model did not produce the requested uncensored image.

Overall assessment

FLUX.3 Image Klein performed well across a broad range of generation tasks and was particularly capable at image editing. Its recurring weaknesses were eyes and facial detail, occasional anatomy or counting errors, and visible AI-like rendering in some images.

Source

The tests and evaluation were narrated by Luke, host of Luke’s Dev Lab. The model is attributed to Black Forest Labs.