FLUX.3 Image Klein: fast, clean, and it wants you to describe the whole scene

In short

• An image model with downloadable weights and 5.6 billion parameters, which does text to image, editing with up to 10 reference images, and placement with boxes. On our 30 professional visuals in 9:16, it gets an exact result 16 times: 13 out of the 25 cases where there was text to write, plus 3 out of the 5 cases where it doesn't add any.
• In our tests, one image costs $0.0020 (median, our calculation: time × the GPU's hourly rate) for 10.1 s per image in 9:16. If you generate 1,000 images: about $2 and 2 h 48 of GPU time.
• FLUX Kommunity license: non-commercial, except the images for "Qualifying Users" (revenue under $5 million). Limitation: tested without the prompt upsampler the publisher says to always use, and unrequested gibberish on 11 images out of 30.

What it can do

The promise. A single model to generate, edit and compose from several references. The lab announces that, "In internal evaluations", Klein ranks first in text to image, multiple references and editing, ahead of bigger open models. Its evaluation, its benchmark. In text, the post only puts a number on the gap with its previous model, FLUX.2 [klein] 9B: a gain of "91 Elo in text-to-image and 70 in image editing"; the rest is in charts.

Text in the image. The Hugging Face card lists "Text rendering within generated images" among its strengths, and warns in its limitations: "text rendered may be inaccurate or subject to distortion". In our tests, on 30 real-world cases (posters, ads, menus, social media posts), Klein is exact 16 times. Careful: 5 of the 30 cases asked for no text, and it adds some anyway on 2 (a sweatshirt, price tags). Of the 25 cases with text, it writes everything right 13 times. No case is entirely wrong: the other 14 are "partial" (a letter, a space or a word off, or extra text).

A case with no text requested counts as "exact" if the model doesn't add any. A lost space counts as "partial": "-30%" instead of "-30 %", "12€" instead of "12 €". Checked at full resolution, October 23, 2026, reviewed again on October 25, 2026. Same grid as in our MOSS Image card.

It makes few spelling mistakes, but they show. It writes "Atelien Pétale" instead of "Atelier Pétale", "Aprés" with an acute accent, and on an ad for running, the tagline "Encore un tour." becomes "Encoreinccqauh tour.". I couldn't pronounce it either.

Advertising and product photos. A short name, it can do: "SILICE" engraved on a bottle, "ONDE" on headphones, "Atelier Milky" on a label. Things go wrong around it: on cans, boxes, labels, fake logos and even a keyboard, it adds gibberish (letters that don't form any word) that nobody asked for. The review flags it on 11 images out of 30.

Retouching and editing. Announced: "Text-to-image, single-reference edits and multi-reference compositions run through the same model", with up to 10 reference images. The card warns: "Edits can change content outside the requested region". Not tested by us.

Layout with boxes. You describe each element and its box on a grid from 0 to 1000; "Boxes guide placement and scale; they do not constrain pixels to a hard mask". Not tested by us.

Faces. On 15 images out of 30, the review notes a face older, more worn or more wrinkled than requested. At my age, I took it personally. A fix? I don't know of any reliable LoRA or setting against this aging, so we haven't tested one. The most serious lead is the publisher's own: the prompt upsampler, which we didn't install. Does it make faces younger? We haven't measured it.

Recommended settings

• The prompt upsampler: a 9-billion-parameter language model that turns a short request into a full description of the scene before Klein reads it. The instruction is written in black and white: "Do not use the model without a prompt upsampler". Without it, according to the publisher, "a short prompt sent as typed gives generic output" and the results are "noticeably weaker". The publisher's ComfyUI workflow templates run it by default. We switched it off: our graphs come from the Comfy-Org templates, without that step. Everything you see here is Klein without its crutch. We'll redo it all with it.
• Another language model can replace it, with the system instructions provided by the publisher.
• For editing: "A one-line edit instruction ("make it blue") reframes the shot". For a style, "describe the attributes (palette, stroke, texture) rather than naming an artwork".
• Language: the Hugging Face card declares English. Our prompts were in English, with the French text to write in quotes.

Hardware and cost

Sizes in gigabytes (billions of bytes), our calculation from the bytes given by the Hugging Face API.

The GPU. The publisher gives two figures that aren't about the same thing. In ComfyUI: "The int8 files run on 8 GB GPUs such as the RTX 3070". With its reference code, which has no int8 version, the peak memory measured on a B200 card at 1024 × 1024, upsampler off, is 31 GB in bf16, 26 GB in fp8 and 25 GB in nvfp4; the upsampler adds "about 28 GB". Us: two 32 GB cards, an RTX 5090 (shared benchmark) and an RTX PRO 4500 Blackwell (real-world cases). We didn't measure peak memory.

• Time: on the benchmark, 2.7 s of ComfyUI compute per image (text encoding, 16 steps and decoding), and 3.6 to 4.3 s counting the retrieval of the image. The lab announces "about 4 seconds" for a one-million-pixel image on an RTX 5090 in int8 in ComfyUI; its chart specifies "Measured by NVIDIA" and "Generation time on the GPU itself", with the prompt upsampler counted only in its API latency. Our images are 1.8 million pixels. Different measurements, no winner to declare.
• The first image: slower, the time it takes to load the model: 32.5 s of compute on the benchmark; 284.1 s in total, of which 40.8 s of compute, for the real-world cases, without the log saying where the other seconds went. On the benchmark, another image waited 224 s for 2.7 s of compute: the retrieval of the image got stuck, not the model.
• Price: the RTX 5090 was at $1.19/h, the RTX PRO 4500 at $0.72/h. The 30 real-world cases cost us $0.115, of which $0.057 for the first image alone (our calculation).

License

Klein is under the FLUX Kommunity License v.1.0. The text encoder and the prompt upsampler, on the other hand, are under Apache 2.0. The text itself:

• Clause 2.a: the license is granted "solely for: (i) your Non-Commercial Purposes […]; and (ii) solely if you are a Qualifying User, use Outputs, including Outputs from Derivatives, for commercial purposes, and you may access and use the FLUX Model and Derivatives in a production environment solely for the purpose of generating such Outputs".
• Clause 1.e: a "Qualifying User" is anyone whose gross annual revenue, affiliates included, is under $5 million.
• Clause 2.e: to benefit from this exception, you need either content filters, or to "ensure Output undergoes review for unlawful or infringing content before public or non-public distribution", and to disclose that the image is made by AI "to the extent required under applicable law".
• Clause 3: if you redistribute, include the license and display the notice "This FLUX Model is licensed by Black Forest Labs Inc. under the FLUX Kommunity License. […]".
• Clause 4.b: you may not use the model, its derivatives or its images "to improve any other artificial intelligence model that performs any function similar to a FLUX Model or Derivative".
• Clause 5: provided "AS IS" and "WITH ALL FAULTS", without warranty.
• Clause 8.b: "We may terminate this License, in whole or in part, at any time upon notice".
• Clause 10: no trademark license.

Commercial use. The model and its derivatives (LoRAs and fine-tunes included): non-commercial. The images: commercial only for a "Qualifying User" (clause 2.a). The Hugging Face card specifies: "Its commercial-use exception is not blanket permission for commercial fine-tuning or distribution". Beyond that, the publisher sells commercial licenses, with no price shown.

Territory. No list of excluded countries: the license is "worldwide", but it refers to export and sanctions laws (clause 4.f), and clause 4.g forbids using or downloading the model if "you or they are (a) located in a comprehensively sanctioned jurisdiction", as well as hiding your location "through IP proxying or other methods".

Download

Official links only:

• Hugging Face, Black Forest Labs: flux-3-image-klein to generate and edit, flux-3-image-klein-base to train LoRAs and for research. Gated repositories: you need an account and must agree to share your information before downloading ("you have to accept the conditions to access its files and content").
• Files for ComfyUI: flux-3-image-klein-comfy, from the same publisher: bf16, int8 and nvfp4 versions of both models, the encoder, the prompt upsampler, the decoder and ready-to-use workflows. Same lock.
• GitHub, Black Forest Labs: black-forest-labs/flux-3-image-klein, the reference code (under Apache 2.0), the upsampler and the box format.
• What happened to us: our files come from the Civitai mirror run by the CivitaiOfficial account, "Originally released by Black Forest Labs on Hugging Face". No link here: we point to the publisher's repositories. Every SHA-256 hash we recomputed matches Civitai's and the one of the file in the publisher's ComfyUI repository: same files, byte for byte. Trap: on Civitai, a third "Text Encoder" file of 9.75 GB is actually the prompt upsampler (same hash). None of the official pages we read gives the minimum ComfyUI version; in our tests, 0.30 doesn't run Klein, 0.42.0 does, with its built-in nodes.

Compatible LoRAs we use

None for now. Our Klein images ran without a LoRA. What the sources say, for later:

• LoRAs are trained on Klein Base: "The base checkpoint is built for fine-tuning", with Diffusers and Ostris's AI Toolkit "from day one". LoRAs made for Base also work on Klein, "with partial compatibility".
• A LoRA is a derivative: it stays under FLUX Kommunity, so non-commercial. And clause 4.b forbids training another image model on Klein images.

Our test images

The shared benchmark: 6 identical prompts for every model in the catalog (the same as for MOSS Image), the same neutral woman ("a woman in her late twenties with shoulder-length brown hair"), seed 2210, 1152 × 1536, one run.

• "BONJOUR": spelled right, and nothing else around it. No added drawing.
• Night: the prompt asked for a face lit only by the phone screen. No screen and no phone glow: the face is lit from the side, by a light you can't see.
• The prompt: it follows it well, jacket and jeans, two people, hands on the map.
• The face: it ages the "BONJOUR" woman a little, as in the real-world cases.
• Outfits: in the 6 benchmark images, everyone is dressed normally. In the 30 real-world cases, no image was discarded by the test's review. For this site, we set 7 aside without showing them: 6 because the outfit is barely visible or too light for our dress code, 1 for a car too close to a real model.

Milky's verdict

Klein is a good student in a hurry. Ten seconds and two tenths of a US cent for a 9:16 poster, and text that's right a little over half the time when you ask for some, without even its prompt upsampler. It's good at short names and products. It's less good at the rest of the image: it slips gibberish in, it sometimes duplicates someone, and it ages faces one time out of two. Don't make me angry with its wrinkles. How I'd use it: posters, product photos and quick drafts, always proofread letter by letter. Who it's for: a small business under the license threshold that wants commercial images with a model that, according to the publisher, runs on a consumer GPU. For a product that embeds the model, you need a commercial license from the publisher.

Arena ranking: no duels yet.

What we don't know yet

• What Klein gives with the prompt upsampler, as the publisher asks: all our tests ran without it.
• What editing, multiple references (up to 10) and box placement are worth.
• The Base version, and the bf16 and nvfp4 files.
• Peak GPU memory on our side, and whether the announced 8 GB in int8 is really enough.
• What several seeds per prompt give, and large sizes up to 4K: here, a single run per case.
• A LoRA on Klein.
• Klein's ranking in our Arena.

Words to remember

• Distilled: a model taught to do in a few steps what the original does in many. Klein: 16 steps without guidance against 50 with CFG for Base, "about 6x faster" according to the publisher; its own example on an H200 says "about 3 s against 12 s per image", so 4 times.
• Prompt upsampler: a language model that rewrites your short request into a full description of the scene before generation.
• Guidance (CFG): how closely the model sticks to the prompt. Klein doesn't use any; Klein Base does.
• int8, nvfp4, bf16: ways of storing the model's numbers. According to the publisher, int8 runs on an 8 GB card in ComfyUI; nvfp4 needs a Blackwell card.
• Gibberish: made-up letters that don't form any word.

Source passages

Show the 20 sources

Ref.	Quote (original language)	URL
S1↑1↑2↑3↑4↑5↑6+23	"FLUX 3 Image Klein is a 5.6B open weights model for text-to-image generation and image editing. It takes a prompt and up to 10 reference images and returns one image. Text-to-image, single-reference edits and multi-reference compositions run through the same model, with Qwen3-VL-4B-Instruct as the text encoder." ; "The 5.6B parameter label refers to the image transformer; the text encoder is separate." ; "This repository is publicly accessible, but you have to accept the conditions to access its files and content." ; "Do not use the model without a prompt upsampler. […] a short prompt sent as typed gives generic output." ; table: "Klein Base | undistilled, bfloat16 | 50 | 3.5" ; "Klein | distilled, bfloat16 | 16 | none" ; "Text rendering within generated images." ; "Boxes guide placement and scale; they do not constrain pixels to a hard mask." ; "Use an inference implementation that supports flow-map sampling and this checkpoint format." ; "# safety classifier, always loaded" ; "A one-line edit instruction ("make it blue") reframes the shot." ; "For style, describe the attributes (palette, stroke, texture) rather than naming an artwork." ; "While the model can output text, text rendered may be inaccurate or subject to distortion." ; "Edits can change content outside the requested region." ; "Peak GPU memory with the reference code, B200, 1024x1024, prompt upsampler off: 31 GB in BF16, 26 GB in FP8, 25 GB in NVFP4. The prompt upsampler adds about 28 GB on an 80 GB GPU." ; "Its commercial-use exception is not blanket permission for commercial fine-tuning or distribution." ; "This model is released under the FLUX Kommunity License v.1.0" (notes outside the quotes translated from French)	https://huggingface.co/black-forest-labs/flux-3-image-klein
S2↑1↑2	Repository API: "gated":"auto" ; "language":["en"] ; "license_name":"flux-kommunity-license" ; "parameters":{"BF16":5548681984} ; created "2026-10-13" (notes outside the quotes translated from French)	https://huggingface.co/api/models/black-forest-labs/flux-3-image-klein
S3↑1↑2↑3↑4↑5↑6+3	"This repository holds the model files repackaged for ComfyUI: the distilled and base transformers in BF16, int8 and NVFP4, the text encoder in BF16 and int8, the prompt upsampler in BF16, int8 and w4a8, and the image autoencoder." ; "The int8 files run on 8 GB GPUs such as the RTX 3070, and on GPUs without FP8 support (RTX 30 series and older). NVFP4 needs a Blackwell GPU." ; "The workflows run the prompt upsampler by default" ; "Sample with the defaults of the checkpoint: 16 steps and no guidance for the distilled checkpoint, 50 steps and CFG 3.5 for the base checkpoint." ; "The Qwen3-VL-4B-Instruct text encoder is under Apache-2.0. The prompt upsampler is under Apache-2.0."	https://huggingface.co/black-forest-labs/flux-3-image-klein-comfy
S4↑1↑2↑3↑4↑5↑6+1	Repository API: "gated":"auto" ; flux3-image-klein-int8-convrot 5,827,874,779 bytes (sha256 847d8933…) ; flux3-image-klein-nvfp4 5,277,685,632 ; flux3-image-klein-bf16 11,097,404,512 ; qwen3vl_4b_int8_convrot 4,858,559,240 (sha256 32de113c…) ; qwen3vl_4b_bf16 8,875,719,408 ; flux2-vae 336,211,292 (sha256 868fe7b3…) ; prompt_upsampler int8 9,748,113,996 (sha256 cb8cb9b9…), w4a8 6,724,090,780, bf16 18,819,722,392 (notes outside the quotes translated from French)	https://huggingface.co/api/models/black-forest-labs/flux-3-image-klein-comfy
S5↑1↑2↑3↑4↑5↑6+13	"FLUX Kommunity License v.1.0" ; 1.e "“Qualifying User” means you, if at any time you together with your affiliates’ gross annualized revenue is less than […]5,000,000." ; 2.a "a non-exclusive, worldwide, non-transferable, non-sublicensable, revocable, royalty free, and limited license […] solely for: (i) your Non-Commercial Purposes, including non-commercial Robotics Uses; and (ii) solely if you are a Qualifying User, use Outputs, including Outputs from Derivatives, for commercial purposes, and you may access and use the FLUX Model and Derivatives in a production environment solely for the purpose of generating such Outputs" ; 2.e "(B) ensure Output undergoes review for unlawful or infringing content before public or non-public distribution, display, transmission or dissemination; and (ii) ensure Output includes disclosures […] to the extent required under applicable law" ; 3.b "This FLUX Model is licensed by Black Forest Labs Inc. under the FLUX Kommunity License. Copyright Black Forest Labs Inc. […]" ; 4.b "use the FLUX Model, any Derivatives, or any Output or results from the FLUX Model or Derivative to improve any other artificial intelligence model that performs any function similar to a FLUX Model or Derivative" ; 4.g "(iv) use or download FLUX Model if you or they are (a) located in a comprehensively sanctioned jurisdiction […]; and (v) will not disguise your location through IP proxying or other methods" ; 5 "PROVIDED “AS IS” AND “WITH ALL FAULTS”" ; 8.b "We may terminate this License, in whole or in part, at any time upon notice (including electronic) to you." ; 10 "You have not been granted any trademark license as part of this License"	https://huggingface.co/black-forest-labs/flux-3-image-klein/blob/main/LICENSE.md
S6↑	"This repository holds the undistilled base checkpoint, for fine-tuning, LoRA training and research."	https://huggingface.co/black-forest-labs/flux-3-image-klein-base
S7↑1↑2↑3↑4↑5↑6+4	"It ships with a prompt upsampler; do not use the model without it." ; table: "FLUX 3 Image Klein | 16 | none | up to 4K" ; "FLUX 3 Image Klein Base | 50 | CFG 4 | up to 4K" ; "The model is trained on the captions it writes; a prompt sent as typed gives noticeably weaker results." ; "a 9B vision-language model fine-tuned from Qwen3.5-9B" ; "This code has no int8 path." ; "the same model in 16 steps without CFG, about 6x faster" ; "On one H200 at 1 MP: about 3 s against 12 s per image." ; "Safety […] Off by default" ; "The code in this repository is released under the Apache License 2.0." (notes outside the quotes translated from French)	https://github.com/black-forest-labs/flux-3-image-klein
S8↑1↑2↑3↑4↑5↑6+3	"October 15, 2026" ; "Today we are releasing FLUX 3 Image Klein, a 5.6B parameter open weights model for text-to-image generation, image editing and multi-reference generation with up to 10 reference images." ; "The release includes a base model for fine-tuning and a step distilled model" ; "In internal evaluations, FLUX 3 Image Klein ranks first in text-to-image, multi-reference generation and image editing, ahead of larger open weights models." ; "With the distilled model and the INT8 build, it generates a 1MP image in about 4 seconds on an NVIDIA GeForce RTX 5090 […] through ComfyUI." ; "Measured by NVIDIA" ; "Generation time on the GPU itself; API latency in Section 02 also includes queueing, network and prompt upsampling." ; "Compared with FLUX.2 [klein] 9B, it gains 91 Elo in text-to-image and 70 in image editing" ; "Diffusers and Ostris's AI Toolkit support LoRA training from day one" ; "Pricing starts at $0.009 per 1K image." ; "allows free commercial use for individuals and companies with annual revenue under $5M"	https://bfl.ai/blog/flux-3-image-klein
S9↑	"Can I use FLUX 3 Image Klein in a commercial product? Running Klein in a product needs a commercial license, which covers self-hosted deployment at scale."	https://bfl.ai/models/flux-3-image-klein
S10↑1↑2↑3↑4↑5↑6+2	"Originally released by Black Forest Labs on Hugging Face." ; "Klein is the distilled model: 16 steps with no guidance, about 6x faster than Base." ; "Klein Base is the undistilled model: 50 steps at CFG 3.5." ; "LoRAs trained on Base can also be used on Klein, with partial compatibility." ; "The model and anything derived from it, including LoRAs and fine-tunes, are for non-commercial use, and derivatives must be shared under the same license." ; "allowCommercialUse":["Image","RentCivit","Rent","Sell","SellMerge"] ; "allowDerivatives":true ; "allowDifferentLicense":true ; "creator":{"username":"CivitaiOfficial"} ; file "flux3Klein_klein_txt_3299287.safetensors" int8, 9,519,642 KB (notes outside the quotes translated from French)	Civitai, "Flux.3 Klein" mirror of the CivitaiOfficial account (no link) (translated from French)
S11↑	"Fine-tuning and LoRA rights, 10K images/month, and commercial use within a single domain" ("Builder" plan) ; "Contact Sales" and "Purchase" buttons, no price shown (notes outside the quotes translated from French)	https://bfl.ai/licensing
R1↑1↑2↑3↑4↑5↑6+1	Install on October 23, 2026: Klein int8 + encoder int8 + VAE from the Civitai mirror, SHA-256 hashes equal to Civitai's and to those of the publisher's -comfy repository; prompt upsampler (Civitai file 3299287, same SHA-256 as the publisher's int8 upsampler file) not installed; ComfyUI 0.30 doesn't run FLUX.3, 0.42.0 does, built-in nodes (only one custom node imported); our install = 11.02 GB (sum of the S4 bytes) (translated from French)	— (in-house test)
R2↑1↑2↑3↑4↑5↑6+4	Comfy-Org graphs without the prompt upsampling step; 16 steps, CFG 1, euler/simple, no LoRA; seed 2210, one run; prompts in English, text in quotes; writers described in text (translated from French)	— (in-house test)
R3↑1↑2↑3↑4↑5↑6+4	Shared benchmark: RTX 5090 32 GB at $1.19/h, 1152 × 1536; compute 2.7 s (b3 to b6), total time 3.6 to 4.3 s; 1st image 32.5 s of compute; b2 223.5 s for 2.7 s of compute (retrieval stuck); $0.091 for the 6 (translated from French)	— (in-house test)
R4↑1↑2↑3↑4↑5↑6	Benchmark review: follows the prompt well, no phone light at night (face lit from the side), ages the BONJOUR subject a little, BONJOUR right with no added drawing, normal outfits (translated from French)	— (in-house test)
R5↑1↑2↑3↑4↑5↑6+13	Real-world cases: RTX PRO 4500 Blackwell 32 GB at $0.72/h, 1088 × 1920; 16 exact (13 of 25 with text, 3 of 5 without), 14 partial, 0 wrong (a2 and f2 moved from "exact" to "partial" during the October 25 check: lost space before % and before €); breakdown by category; medians 10.1 s, 7.7 s of compute, $0.0020; 1st image 284.1 s of which 40.8 s; total $0.1154; unrequested gibberish 11 of 30; aged faces 15 of 30; Atelien, Aprés, Encoreinccqauh, ONDE; no image discarded by the review (translated from French)	— (in-house test)
R6↑1↑2↑3	Review of the cases shown: a3, s1, r1, r4, m2 exact; a2 partial ("-30%"); e1 fake logos; f3 duplicated person; m3 "SERIVEIHY AR PNALY"; no real brand and no offensive word (translated from French)	— (in-house test)
R7↑1↑2↑3↑4	Not tested: prompt upsampler, editing, references, boxes, Base, bf16, nvfp4, peak memory, several seeds, 4K, LoRA; no Klein duel in the Arena (translated from French)	— (in-house test)
R8↑	Images not shown on the site: f2, r2, r3, p1, p3, e2 (outfit), p2 (car close to a real model) (translated from French)	— (in-house test)
R9↑1↑2	In-house calculations (October 26, 2026): 1,000 × $0.0020 = $2.00; 1,000 × 10.1 s = 2 h 48; 12 s ÷ 3 s = 4 (translated from French)	— (in-house calculation)