Krea 2 LoRA: How to Find, Load and Test LoRAs on Krea 2 Turbo

Krea 2 softens some requests by default, and our verdict on that is in Is Krea 2 uncensored?. That article tested Krea's hosted models. This guide is about the open-weights Krea 2 Turbo, which takes your own LoRA files, and about getting NSFW-capable and style LoRAs to load and actually change the picture.

Short answer: Krea recommends training a LoRA on the Raw checkpoint and running it on Turbo. On Krea 2 Turbo LoRA you pass up to three direct .safetensors links, each with a scale. The one trap that cost us the most time: a LoRA repository can ship two files, and only the ComfyUI-format one worked on this endpoint. The other loaded without an error, was charged, and left every pixel unchanged. A 1k image costs $0.0216, with or without LoRAs.

What a Krea 2 LoRA is: Raw to train, Turbo to run

Krea released two open-weights checkpoints in June 2026, a 12-billion-parameter model in two forms:

• Raw is the pre-trained model with no distillation (52 steps). Krea's README says: "We highly recommend using RAW for training LoRAs and applying them on Turbo for inference."
• Turbo is the 8-step distilled model, the one that is fast enough to serve cheaply. ComfyUI's Krea 2 page says LoRAs trained on Raw apply seamlessly to Turbo.

So a Krea 2 LoRA you find online should say it was trained for Krea 2 (Raw or Turbo). Both checkpoints sit under the Krea 2 Community License, and companies with $1 million or more in annual revenue need a paid license from Krea, so read it before you build a product on it.

Where to find Krea 2 LoRAs

Treat Hugging Face and Civitai as sources, not as a menu we endorse file by file.

• Krea's own nine style LoRAs on Hugging Face (repositories starting krea/Krea-2-LoRA-): soft watercolor, dark brush, sunset blur, dot matrix, retro anime, rainy window, neon drip, kids drawing and vintage tarot. They are not gated, and ComfyUI's page suggests a strength of 1.0.
• Community style collections. ilkerzgi's Krea 2 style LoRAs is an index of well over a thousand styles, each in its own repository. The author puts the trigger phrase at the end of the prompt, suggests 1.0 to 1.25 and says shorter prompts work better. These are style LoRAs.
• Civitai and Hugging Face search. Filter by base model "Krea 2" on Civitai, or search krea-2 on Hugging Face. Look at the stated base model before you download anything. Realism, character, speed-up and editing LoRAs are all mixed in with the styles.

Do not point the endpoint at a base-model repository such as krea/Krea-2-Turbo itself: it is gated, and a gated file fails with an HTTP 401 error (not charged).

Where Krea 2 LoRA training happens

This is the search term people type most, so here is the split:

• Open-source trainers. Krea's README lists Diffusers, Ostris AI Toolkit and Kohya's musubi tuner. Train on Raw, export a .safetensors file, and you have something that works on Turbo.
• Training inside krea.ai. Krea added LoRA training to its own site in May 2026, as a beta for Max and Business subscribers. It trains on Krea 2 Medium or Large, takes at least three captioned images, and keeps the result private by default. The page does not say you can export a file. We did not test it, and we would not assume those LoRAs can be loaded anywhere else.

How to load a Krea 2 LoRA by API

Send a standard createTask request with your key in SPICY_API_KEY. This example uses a style LoRA from the collection above, a short prompt and the trigger phrase at the end:

{
  "model": "krea/2-turbo-lora/text-to-image",
  "input": {
    "prompt": "A woman in a green wool coat walking through a rainy night market, warm lantern light, cinematic, black ink linocut style",
    "loras": [
      {
        "path": "https://huggingface.co/ilkerzgi/krea-2-black-ink-linocut-lora/resolve/main/black-ink-linocut-comfy.safetensors",
        "scale": 1.25
      }
    ],
    "aspect_ratio": "2:3",
    "resolution": "1k",
    "seed": 202
  }
}

POST it to https://api.spicyapi.ai/api/v1/jobs/createTask with Authorization: Bearer $SPICY_API_KEY and an Idempotency-Key header, then poll the task. The tasks guide covers polling and webhooks, and Run your own LoRA on a hosted API covers the general workflow.

• loras: up to three, each {"path", "scale"}. A fourth is rejected with a 400. Scale runs from 0 to 4; 0.5 to 1.2 is the useful range, and 4 breaks the image into colour blocks.
• path: a direct file link. On Hugging Face that is the /resolve/main/<file>.safetensors address. A repository page link fails with "Failed to load LoRA: You should provide the direct download link to the model". It is not charged.
• Trigger phrase: put it where the LoRA's author says. If the card is silent, put it first. Ours worked at the end because that is what the author specified.
• Speed: a LoRA file seen for the first time can take a minute or more to load. In our 22 runs on 24 October, requests finished in 20 to 60 seconds, and a file already loaded earlier runs in about 13 seconds at 1k.

The model page also has an image-to-image mode with the same loras slot. For the page, go to Krea 2 Turbo LoRA.

Public examples from the model page

These two were made on the endpoint with Krea's own style LoRAs at strength 1.

Did the LoRA actually load? The same-seed control

The endpoint does not tell you whether a LoRA changed anything. You have to look. The reliable control is to run the same seed with and without the file.

One catch we hit: a trigger phrase alone changes the style, because the base model understands words like "ink linocut". So the control is not "no LoRA and no trigger". It is "no LoRA, trigger only" against "LoRA plus trigger". We ran this on 24 October 2026, 1k, same seed throughout:

Two lessons. First, the style needs the trigger phrase. Second, the first file was loaded, charged and ignored.

How to tell which file you have

Some LoRA repositories ship two exports of the same LoRA. The collection above lists an "original" and a "comfy" file. On our endpoint:

• The ComfyUI file, whose tensor names start with diffusion_model., took effect. The file name often ends in -comfy.
• The original file, whose tensor names start with base_model.model. (the PEFT prefix from training tools), was silently ignored. Twice, with a long and a short prompt, it produced exactly the same pixels as no LoRA.

You can check the names without downloading the whole file: the first bytes of a .safetensors file are a JSON header, and Hugging Face's file viewer shows the tensor list for any file. If you only have one file and are unsure, the same-seed control above settles it. We did not test Diffusers files that use a transformer. prefix.

Krea2 Filter Bypass: what it changed in our tests

"Krea2 Filter Bypass" is the community name for a tiny file on Civitai, with mirrors on Hugging Face. We ran the "3vector" mirror on the endpoint. Three things to know:

• It is not a normal LoRA. The file is 160 bytes and holds one 12-number offset on the text-conditioning layer, with no style or character weights. It still goes into a loras slot and takes effect: the same seed gave a clearly different picture with it loaded.
• It changes how Krea 2 reads a request. SpicyAPI adds no content filter of its own. What the file shifts is the base model's own habit of softening certain prompts.
• The effect is not a clean unlock. We used one test prompt at each of two of our five intensity levels, Moderate and Strong (test prompts not published), with fictional adults only:

So the file turned 5 of 5 runs into delivered-as-written. But the Strong prompt was delivered without it too, while the Moderate one was toned down. The base model's softening depends on the wording of the prompt and does not rise steadily with intensity. We would not call this an unlock. Our earlier tests on 9 October, with a different prompt set, showed the base model softening from the second level up and community LoRAs delivering every level.

On strength, the author's page allows very high values and a third-party host suggests +1 to +2. In a Moderate run on one seed, 0.6, 1.0 and 2.0 were all delivered, with no visible quality drop at 2.0; a Civitai version at 3.0 also delivered. Start at 1.0, and use the lowest scale that works.

Civitai links also matter here. A direct Civitai download link returned a 401 when we fetched it anonymously, yet it loaded on the endpoint in both our tests. If one of yours fails, download the file with your own account, put it on your own Hugging Face repository or storage and paste that direct link.

LoRAs that silently do nothing

The pattern repeats: no error, the usual charge, and a picture that looks like the base model.

• Wrong base model. We loaded an illustration LoRA trained for FLUX.1-dev at strength 1 on Krea 2 Turbo. It was charged $0.0216, produced no illustration style and matched the no-LoRA picture closely. A Z-Image LoRA behaved the same way on 9 October. Check the stated base model: Krea 2, Krea-2-Turbo or Krea-2-Raw.
• Wrong file format. The PEFT-prefixed original file above.
• No trigger phrase. The style stays away.
• A krea.ai-trained LoRA. Trained on Medium or Large, private, and with no export documented.

Errors and fixes

Failed loads are not charged.

Price

At the time of writing, 24 October 2026, per image:

Zero, one, two or three LoRAs cost the same. The model page always shows the current price. For the Krea 2 models without LoRA slots, see Krea 2 Large and Krea 2 Medium Turbo.

FAQ

Is there a Krea 2 NSFW LoRA?

Yes. The best-known community file is the one covered in the filter section above, and we ran it on the endpoint. In our five-run test it delivered every request as written, though the base model also delivered our Strong prompt without it. SpicyAPI adds no content filter of its own. Fictional adults only, and no real person's likeness without consent.

Should I train a Krea 2 LoRA on Raw or Turbo?

Krea recommends Raw for training and Turbo for running the result. A LoRA trained on Raw applies to Turbo, which is how the endpoint uses it.

Can I use a LoRA trained on krea.ai?

We could not confirm it. Krea's own training runs on Medium or Large and the page does not say you can export a .safetensors file. The endpoint expects open-weights LoRAs for Krea 2 Turbo.

Why did my Krea 2 LoRA do nothing?

Usually the wrong file format, the wrong base model or a missing trigger phrase. Run a trigger-only control at the same seed. If the two pictures match, the LoRA did not load.

How many LoRAs can I stack?

Three per request, each with its own scale. Stacking adds no charge.

Does Civitai work?

In our two tests, a Civitai download link that returned 401 for an anonymous request still loaded on the endpoint. If yours fails, host the file yourself and link that.

Adults only. Use fictional adults in every prompt and upload, and you are responsible for what you generate.

Try Krea 2 Turbo LoRA →