Flux 3 Klein local tests split verdict: weak text-to-image, impressive editing

After Black Forest Labs open-sourced the Flux 3 Image Klein image model (including a 5.6B base and a distilled version at 2.25MP, 16 steps), Reddit quickly filled with local test results, and opinions split sharply: text-to-image output is widely seen as worse than the previous generation, while high-resolution consistent editing has drawn high praise.

Confirmed

• The model is not trained directly on natural language; it relies on a prompt upscaler that converts natural language into a specific format. @StrictMove5839 and @desktop4070 both noted that default ComfyUI/Diffusers templates lack this step, causing poor outputs; the latter also posted comparison images at 1.0–9.0 MP using the Default Template (euler+simple, 15 steps) and asked for optimal parameters.
• Editing is the biggest highlight: @herbertseabra (a photographer) called photo edits that preserve pose, camera angle, and consistency "near perfect," retaining realistic skin texture at 3000×4000—something FLUX.2 Klein and Qwen 2.1 cannot do; @ZootAllures9111 tested straight-through editing at 4096×4096, converting house photos into colored-pencil, sketch, and watercolor styles while keeping the subject's structure; @Ant6431 also finds the editing "promising."
• Distilled version efficiency: on an RTX A4500 (20GB), @Kandoo85 compared the bf16 base (50 steps, CFG 3.5) against the distilled version (15 steps, CFG 1) with 50 same-seed prompts, finding a 6x speedup with almost no quality loss.
• Generation reviews are mostly negative: @Few-Intention-1526 found editing results actually worse than F2K, contradicting the official benchmark claim that "F3K comprehensively surpasses F2K"; @Both-Rub5248, after systematic comparisons on an RTX 3090, considers F3K inferior to the previous generation, with the Base version "almost unusable"; @Capitan01R says both F3K and Qwen 2.1 Image feel "off"; @GlassLow5680 reports community sentiment is "pretty bad."
• In a six-version comparison of FLUX.2 Klein (4B/9B) versus F3K (5.6B), @ZootAllures9111 found anatomy issues less severe than rumored, but overall it's at best a sidegrade to F2K 4B, also flagging concerns about dataset quality.
• There are positive cases too: grass-lying portraits shared by @DifferentFix2217 and @Similar-Reserve-3581 reached photorealistic quality—the latter was even suspected of being from OpenAI, while complaining about overly strict safety filters; @Crazy-Repeat-2006 found that progressively stacked prompts can partially rescue the distilled version's quality.
• On head-to-heads: @Devajyoti1231 ran the same prompts against FLUX 3 Klein, FLUX.2 Klein 9B, Qwen Image 2.1, and Krea2, concluding FLUX 3 has the strongest prompt adherence but the worst anatomy, with Krea2 and Qwen being better. @Maleficent-Tell-2718's video claims F3K with BBox control is among the strongest open-source text-to-image/editing solutions today.
• On specific aspects: @Schattenmensch finds atmospheric lighting (blizzards, gothic night scenes) excellent, but creature details and proportions prone to breaking down.

Unconfirmed

• Whether the gap between official benchmarks and community tests stems from non-default workflows like skipping the prompt upscaler remains unsettled; @Capitan01R and others say they're holding out hope for the upcoming Krea3.

Why it matters

• F3K is one of the most-watched open-source image models right now, and its split verdict—mediocre generation, stunning editing—directly shapes user choices: those needing high-resolution consistent editing (photography, design workflows) clearly benefit, while text-to-image-focused users can stay on Flux 2 Klein or switch to Qwen/Krea for now.

2026-10-16 ~ 2026-10-18 · 18 related posts

Full story(5 episodes)→

• Episode 1: Black Forest Labs Open-Sources FLUX.3 Image Klein: 5.6B Model Runs 4K Generation on a Single Consumer GPU(2026-10-14, 35 posts)
• Episode 2: Black Forest Labs open-sources FLUX.3 Image Klein with day-one ecosystem support(2026-10-15, 7 posts)
• Episode 3: BFL Updates FLUX 3 Klein ComfyUI Workflows and Shares New Samples(2026-10-16, 4 posts)
• Episode 4: Flux 3 Klein local tests split verdict: weak text-to-image, impressive editing(2026-10-16, 18 posts)
• Episode 5: Quantization Benchmarks Show FLUX 3 Klein Running Over 2x Faster with Better Quality(2026-10-17, 3 posts)

• Flux3 Klein Feels Worse Than Its Predecessor, Complains Reddit User Awaiting Krea3 — Capitan01R- · 2026-10-16
• Flux 3 Klein underperforms Flux 2 Klein as edit model in Comfy despite better benchmarks — Few-Intention-1526 · 2026-10-16
• Early local test of distilled FLUX.3 Klein: mediocre t2i, editing hopes remain — Ant_6431 · 2026-10-16
• Flux 3 Klein's photorealistic portraits spark Reddit buzz — Different_Fix_2217 · 2026-10-16
• Flux 3 Klein tested: distilled is 6x faster than base with near-identical quality — Kandoo85 · 2026-10-16
• [source] Flux 3 Klein's Default Templates Are Broken Without Its Prompt Upscaler — StrictMove5839 · 2026-10-16
• Flux 3 Klein excels at lighting and atmosphere but fumbles living subjects — Schattenmensch · 2026-10-16
• Flux 3 Klein vs Flux 2 Klein hands-on: the new base model 'barely usable' for T2I — Both-Rub5248 · 2026-10-16
• BFL Flux's photorealistic 'woman on grass' output stuns Reddit, safety rules irk users — Similar-Reserve-3581 · 2026-10-16
• Flux 3 Klein's default ComfyUI template struggles at high resolutions — desktop4070 · 2026-10-16
• FLUX 3 vs Qwen vs Krea2 community test: best prompt following, worst body anatomy — Devajyoti1231 · 2026-10-16
• [source] FLUX.3 Klein shines at consistent high-res photo editing on RTX 5090: 3000×4000 in ~230s — herbertseabra · 2026-10-16
• [source] FLUX.3 Klein looks like a sidegrade: comparison reveals AI-slop training data contamination — ZootAllures9111 · 2026-10-17
• Flux 3 Image Klein with BBox: hands-on with top open-source text-to-image and editing — Maleficent-Tell-2718 · 2026-10-17
• Flux.3 Klein demonstrated doing direct straight-through edits at 4096x4096 — ZootAllures9111 · 2026-10-17
• Testing whether better prompting can save distilled Flux 3 Klein at 2.25MP — Crazy-Repeat-2006 · 2026-10-17
• Flux Klein 3 and MOSS Image land within days, early verdict says Klein 3 disappoints — GlassLow5680 · 2026-10-17
• Flux 3 Klein first tests in ComfyUI: Qwen-Image wins character consistency — JuniorEnsign · 2026-10-18