Data window: ZView Space production logs, 13 July – 19 October 2026. This describes our own ComfyUI setup; it is not Krea's official guidance.
Krea 2 Turbo is the image model behind most of what we generate. Krea released it in June 2026 as an 8-step distilled version of Krea 2, and we have run it for months in a ComfyUI workflow. This page collects the settings we actually use, how they differ from the official defaults, how fast a job really is on our hardware, and how much of our traffic uses a LoRA.
What the official page says
From Krea's repository:
- Krea 2 Raw is the undistilled base checkpoint, meant for fine-tuning, post-training and LoRA training. Recommended settings are 52 steps with CFG 3.5, trained for images up to about 1K.
- Krea 2 Turbo is the 8-step distilled checkpoint for fast text-to-image. The recommended inference settings are 8 steps with CFG 0.0 and `mu` 1.15, at resolutions from 1K up to 2048 x 2048.
- Krea's advice is to train LoRAs on Raw and apply them on Turbo for inference.
- The weights use Krea's community licence; the inference code is Apache-2.0.
Krea also states that Krea 2 ranks first among independent-lab text-to-image models on Artificial Analysis. That is Krea's own claim and we have not verified it.
Our production settings
| Setting | Our value | Official Turbo recommendation | | --- | --- | --- | | Steps | 8 | 8 | | CFG | 1 | 0.0 | | Sampler / scheduler | ER-SDE / beta | not specified on the page | | Checkpoint | int8-quantised Turbo | Turbo | | Batch size | 1 | 1 |
Two notes. First, our CFG is 1, not 0. In ComfyUI a CFG of 1 means "no classifier-free guidance", so it is close to the zero the official script uses; we have not run a controlled comparison between the two. Second, we run an int8-quantised checkpoint, which trades some precision for speed and memory, and we have not measured that trade separately.
Earlier versions of our workflow (v3.0) used a different sampler (dpmpp_sde with the beta57 scheduler) and a different checkpoint, so the setup changed during the three months covered by the logs; the table shows the current production workflow.
How fast is it, really
Our queue records the wall-clock time of every finished Krea2 job. In the 100 most recent samples the median was 22 seconds, the range 22 to 25 seconds, which is very consistent. That figure covers a whole job on our worker at our production sizes, which are mostly 9:16, and not only the sampler. It is not directly comparable with other people's numbers: a Gigazine hands-on test in ComfyUI on an RTX 5070 Ti reported roughly 9 to 15 seconds for a 1024 x 1024 image with reused prompts and 16 to 25 seconds when the bundled prompt enhancer rewrote them. Different hardware, size and prompt handling make a head-to-head comparison meaningless, but both point to Turbo being a seconds-per-image model, not minutes.
How much is LoRA traffic
Of the 21,111 Krea2 Turbo requests logged between 13 July and 18 October 2026, 64% (13,471) used a character LoRA, 2% (501) used our miniature-scene workflow, and the remaining 34% (7,139) used none. If your workflow resembles ours, expect LoRAs to be the main use of Turbo, not an extra. That matches Krea's own advice to treat Turbo as the place where LoRAs are applied.
The format split looks like the rest of our logs: 19,583 of the 21,111 Krea2 requests (93%) were 9:16, because we make vertical, phone-first content.
Practical advice
1. Start from 8 steps. More steps are not what Turbo is for; if you need more control, move to the Raw model instead. 2. Test CFG 0 and CFG 1 once in your own setup and keep whichever you cannot tell apart. 3. Write structured prompts. The median Krea2 prompt in our logs is about 1,100 characters (see our [prompt log report](/articles/what-25000-prompts-taught-us-generation-log-report/en)), organised as subject, setting, light and camera. 4. Train on Raw, apply on Turbo, as Krea recommends. 5. Measure your own worker. A median of 22 seconds is ours; your GPU, precision and size will change it.
Method note: step, CFG and sampler values are read from our production ComfyUI workflow files. Timing is the median and range of the queue's most recent 100 stored samples. LoRA shares come from a read-only count of the prompt log on 19 October 2026.