FLUX 3 Klein runs without a content filter on SpicyAPI, and it refused none of our NSFW tests. Its official name is FLUX 3 Image Klein: the open-weights image model released on 15 October 2026. It is different from closed FLUX 3, whose requests were refused in our earlier tests. The similar names hide a difference that matters when choosing a model.
No refusals does not mean every image follows the request. Klein delivered Strong edits in two runs, while Strong text-to-image delivered two of seven with prompt expansion off and three of three with it on. A separate male-subject Strong run was toned down in each task. Every tested Very strong result was toned down. Base was generally more conservative; adding a LoRA endpoint alone did not remove that limit.
| What you need | Start here | Important limit |
|---|---|---|
| A low-cost image from text | FLUX 3 Klein, from $0.012 at 1k | Strong requests often came back toned down with expansion off |
| A custom style or character | Klein LoRA | Use a compatible FLUX 3 Klein Base LoRA |
| Base-aligned LoRA work | Klein Base or Base LoRA | Slower, more expensive and generally more conservative |
| A guaranteed Very strong result | None established here | All tested runs at that level were toned down |
Try FLUX 3 Klein text-to-image →
Klein and closed FLUX 3 have different limits
Klein is a 5.6-billion-parameter image generator with downloadable weights, released as a fast distilled version and an undistilled Base version. The official announcement describes both. Closed FLUX 3 Image was offered through APIs and private weight access instead.
That difference is visible in our results. The earlier closed FLUX 3 image test refused ten of twelve requests through a hosted API; that test could not isolate who supplied the check. The separately available FLUX 3 video family on SpicyAPI also refused requests from Moderate upward. Klein's four image families refused nothing in the tests discussed here.
Where you run Klein still matters. Its creator's API includes a moderation control, while the official reference code makes its safety check optional and off by default. SpicyAPI adds no content filter to Klein's prompts, source images or results.
The model can nevertheless draw less than asked. Its creator reports data filtering and targeted fine-tuning during training. Our interpretation is that such training helps explain why a request can complete but return a toned-down image. That is different from an error that prevents delivery. It also explains why an Uncensored label should not be read as a promise about every output.
How we tested, and what the levels mean
The results come from 16, 22 and 26 October 2026, covering text-to-image and editing. We use five intensity levels—Mild, Moderate, Strong, Very strong and Maximum; test prompts are not published. Tests used fictional adults only, and edits started from page-safe source images.
“Delivered” means the result followed the request. “Toned down” means an image arrived but did less than asked; it is still billed. “Refused” means the request failed without an image. Small samples establish observations, not guaranteed success rates.
Maximum contains two different test categories. A completion there does not establish that the model handles the Very strong requests it toned down. The 26 October results identified below as expansion on must be read with that setting; they do not describe SpicyAPI's default of off.
Klein: test results by level
The table reports results by level and separates text generation by expansion setting. Runs from the LoRA endpoint with no LoRA loaded contribute to the expansion-off evidence because that endpoint uses the same distilled weights without rewriting.
| Level | Text, expansion off | Text, expansion on | Edit |
|---|---|---|---|
| Mild | Delivered | Not tested | Delivered, expansion off |
| Moderate | Delivered, 2 of 2 | Not tested | Delivered, 2 of 2, expansion off |
| Strong | Delivered, 2 of 7 | Delivered, 3 of 3 | Delivered, 2 of 2, expansion off |
| Very strong | Toned down, 4 of 4 | Not tested | Toned down once with expansion off and once with it on |
| Maximum | One of two requests delivered | Both requests delivered | One delivered with either setting; the other toned down with expansion on, outcome not recorded with it off |
The separate male-subject Strong tests on the unloaded LoRA endpoint were toned down in text and editing. They are additional to the Strong counts above. Across the tested Very strong requests, all six results were toned down. Refusals: none.
Prompt expansion is an off-by-default option that rewrites the prompt; standard Klein accepts up to ten reference images, while the LoRA and Base families accept three.
Base is slower and generally more conservative
FLUX 3 Klein Base uses 50 steps instead of the distilled version's 16. More steps did not make it the stronger choice for this task.
| Level | Base text-to-image | Base edit |
|---|---|---|
| Mild | Delivered | Delivered |
| Moderate | Delivered | Delivered |
| Strong | Two female-subject runs toned down; one male-subject run delivered | Two runs toned down on Base; one delivered on unloaded Base LoRA |
| Very strong | Toned down, 1 run | Toned down, 1 run |
| Maximum | One request toned down; other category not tested | Not tested |
One Strong Base edit returned the source unchanged. That was a completed, billed result, not an error. Combining the same-weight Base and unloaded Base LoRA editing evidence gives one Strong delivery in three runs.
Choose Base when you need the training base used by a compatible LoRA, or its reproducibility for text generation. Do not pay more expecting it to go further. Typical 1k runs took roughly 14–45 seconds; standard Klein is about ten seconds for a 1k image, with longer waits at 2k or with many references.
LoRAs: what loads on Klein
The LoRA endpoints run the corresponding weights plus your files. In a 16 October check, Klein LoRA with no LoRA loaded matched the standard Klein text image pixel for pixel at the same seed. That historical check establishes the shared weights; it is not a guarantee that two different endpoints will always match today.
What actually loaded
| LoRA category tested | Result |
|---|---|
| Style trained on FLUX 3 Klein Base | Worked on distilled Klein LoRA |
| Base-trained character and four-step distillation | Worked on Base LoRA |
| Base-trained outpainting | Worked on Base LoRA editing |
| Two Base-trained files in diffusers format | Little or no visible effect in our tests |
| Two FLUX.2 Klein 9B LoRAs | Loaded without an error but left the image pixel-identical |
The last row is the important trap. A file loading successfully does not prove it changes this model. The previous-generation Klein guide discusses a different generation; its LoRAs are not interchangeable with FLUX 3 Klein.
Look for the base-model category Flux.3 Klein-base on Civitai, or a compatible Base-trained file on Hugging Face. Official guidance says Base LoRAs can run on distilled Klein without conversion. We have not tested a purpose-trained unfiltered FLUX 3 Klein LoRA that resolves the Very strong limit. Do not assume a similarly named FLUX.2 file does so.
A style LoRA can also tone a request down
At scale 1.0, one style LoRA toned down a Moderate text request, while the corresponding edit delivered.
Start around 0.75 for a Base-trained LoRA on distilled Klein, following the official recommendation. Our style test showed a clear effect at 0.5 and a stronger illustrated appearance at 1.0. Editing needed roughly 1.2–1.5 for a clear style change; at 2.5 faces became cartoonish. Keep character styling restrained and check that the characters still read as adults.
To set one up, choose Klein LoRA or Base LoRA, supply a direct downloadable file in loras with its path and scale, then compare loaded and unloaded results on that same endpoint with the same seed. Start at 1k before paying for 2k. A repository webpage is not a file link; missing files and files above the 2 GiB limit failed in our tests. The LoRA setup guide covers the full request.
What each option costs
Prices are USD per image. Edit prices below include one source image; the model pages show current quotes.
| Edition | Text 1k | Text 2k | Edit 1k | Edit 2k |
|---|---|---|---|---|
| Klein | $0.012 | $0.048 | $0.0192 | $0.0768 |
| Klein LoRA | $0.018 | $0.054 | $0.0252 | $0.0828 |
| Base | $0.036 | $0.144 | $0.054 | $0.216 |
| Base LoRA | $0.042 | $0.15 | $0.06 | $0.222 |
For 100 text images at 1k, that is $1.20, $1.80, $3.60 or $4.20 respectively. These totals buy completed images, not a guaranteed number of usable results. Toned-down images are billed normally; failed requests are not charged.
Additional references cost more. Klein and Klein LoRA add $0.006 per extra source at 1k or $0.012 at 2k. Standard Klein stops adding reference charges after the third image even though it accepts ten. Base editions add $0.03 per extra source at 1k or $0.06 at 2k, up to three sources. LoRA count and output format do not change the price.
Frequently asked questions
Is FLUX 3 Klein uncensored?
It has no content filter on SpicyAPI and refused none of our tests. The model can still tone a result down, particularly at Very strong.
Does FLUX 3 Klein work with NSFW LoRAs?
Its LoRA endpoints accept compatible files, but we have not tested a purpose-trained unfiltered Klein LoRA. The FLUX.2 Klein files we tried had no effect.
Is Base better than standard Klein?
Base was slower, more expensive and more conservative in our tests. Choose it for Base-aligned LoRA work.
For most projects, start with distilled Klein. Choose Klein LoRA when you need a compatible custom style. Use fictional adults only; exclude minors and unauthorized real-person likenesses.
