> ## Documentation Index
> Fetch the complete documentation index at: https://comfyui-mcp.artokun.io/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# How to Run Krea 2 Locally in ComfyUI (2026 Guide)

> Run Krea 2 Turbo — Krea.ai's 12B open-weight text-to-image model — locally in ComfyUI. 8-step turbo speed, Qwen3-VL prompting, JSON area prompts, install, and VRAM tips.

*June 24, 2026*

Krea 2 is the newest open-weight text-to-image model, and it lands with an
unusual pitch: turbo-class speed **and** strong prompt-following in the same
checkpoint. This guide covers what it is and how to run **Krea 2 Turbo** locally
in ComfyUI in one command.

## What is Krea 2?

Krea 2 is a **12-billion-parameter Diffusion Transformer** from
[Krea.ai, Inc.](https://www.krea.ai/), released June 22, 2026, with the weights
open-sourced the next day in two variants:

* **Krea 2 Raw** — the base checkpoint before extra post-training.
* **Krea 2 Turbo** — the post-trained, **distilled** checkpoint that generates in
  \~**8 steps at cfg 1**. This is the one this pack ships.

It runs on **native ComfyUI support** (a `krea2` text-encoder type built into
ComfyUI ≥ v0.26.0), pairing a **Qwen3-VL 4B** text encoder with the **Qwen image
VAE**. The Qwen3-VL encoder is what gives it sharp prompt adherence and the
ability to read **structured JSON prompts** (more on that below).

### The license actually lets you ship

Krea 2 is released under the
[Krea 2 Community License](https://www.krea.ai/krea-2-licensing): **free
commercial use** for individuals and businesses up to 50 seats, with an
enterprise agreement beyond that. For most solo creators and small studios,
that's a genuinely permissive "use it in production" license.

## Krea 2 vs Z-Image, Ideogram 4, and Flux

Krea positions Turbo as combining turbo-model speed with the quality and
prompt-control of much larger models. In practice the trade-offs look like this:

| Model            | Params | Speed                 | Standout                                    |
| ---------------- | ------ | --------------------- | ------------------------------------------- |
| **Krea 2 Turbo** | 12B    | \~8 steps (very fast) | speed **+** prompt-following + JSON prompts |
| Z-Image Turbo    | 6B     | \~8–14 steps          | smallest VRAM footprint                     |
| Ideogram 4       | 9.3B   | full steps            | best readable text + area prompting         |
| Flux.1 dev       | 12B    | \~20–28 steps         | broad LoRA ecosystem                        |

If Z-Image Turbo is the low-VRAM speed pick and Ideogram 4 is the text/layout
pick, Krea 2 Turbo aims to sit between them — turbo speed without giving up
prompt fidelity, plus the JSON area-prompting workflow Ideogram users like.

## System & VRAM requirements

The pack ships the **fp8** build (`krea2_turbo_fp8.safetensors`) for RTX
4000/3000/2000 cards. On **RTX 5000 (Blackwell)**, swap in
`krea2_turbo_mxfp8.safetensors` for native fp8. A **24 GB** card (e.g. RTX 4090)
runs the fp8 build comfortably at 1080p.

Three files are needed (the installer fetches them):

* `diffusion_models/krea2_turbo_fp8.safetensors` — the 12B Turbo model (fp8)
* `text_encoders/qwen3vl_4b_fp8_scaled.safetensors` — Qwen3-VL 4B encoder
* `vae/qwen_image_vae.safetensors` — Qwen image VAE

> **Important:** the CLIPLoader uses `type=krea2`, which requires native Krea 2
> support in ComfyUI (`comfy/text_encoders/krea2.py`, ComfyUI ≥ v0.26.0). Update
> ComfyUI first if that file is missing.

## Install Krea 2 in ComfyUI

### The fast way — comfyui-mcp + the Panel

From a Claude Code session with the [comfyui-mcp](../) plugin installed:

```
Install the krea2-txt2img-manual pack
```

There are **three** packs (sliced from the KREA2 ULTRA **V2** workflow) — pick by
what you want:

* **`krea2-txt2img-manual`** — a plain prose prompt.
* **`krea2-txt2img-json`** — Ideogram-style structured JSON / area prompting.
* **`krea2-combo`** — a two-pass **detail boost** (turbo LoRA + optional IdeoKrea).

The MCP applies the pack manifest — clones the custom nodes (`rgthree-comfy`,
`ComfyUI-KJNodes`, `ComfyUI-Krea2T-Enhancer`, `ComfyUI-RBG-SmartSeedVariance`;
combo also `ComfyUI_essentials`), downloads the model files, and loads the
workflow. Each is a single-pipeline graph with no group toggles and its one prompt
source active. **V2** wires in the `Krea2T-Enhancer` model detail-boost patch
(active); `RBG_Smart_Seed_Variance` ships bypassed (optional).

### Or one-click

Run the generated `install-windows.bat` (or `install-runpod.sh`) from your
ComfyUI root, restart, and load `workflow.json`.

## Structured JSON prompting

Like Ideogram 4, Krea 2's Qwen3-VL encoder understands **structured prompts** —
per-region descriptions with bounding boxes and palettes — so you can place and
describe subjects deterministically instead of fighting one long prose string.
For this, install the **`krea2-txt2img-json`** pack: its `Ideogram4PromptBuilderKJ`
node drives the encoder directly (nothing to un-bypass). Set ALL the builder
fields and keep palettes empty — a rich palette can render as a literal
color-swatch strip.

## Settings that matter

* **Steps: 8, cfg: 1** — Turbo is distilled; more steps won't help and higher cfg
  will over-cook it.
* **Sampler `er_sde` / scheduler `simple`** — the pack's defaults; sharp results.
* **Resolution 1920×1080** is the shipped default; Krea 2 handles a wide aspect
  range.
* **ImageSharpenKJ** runs before save for crisp output. **Smart Seed Variance**
  ships **bypassed** (optional) — un-bypass it for controlled variations of the
  same prompt (set its seed to `randomize` and tune the variance strength).

## V2: the detail boost + combo

The V2 workflow adds the **`Krea2T-Enhancer`** node — a model-level detail-boost
patch that ships **active** in every pack (bypass it to A/B against the un-boosted
result). The headline is **`krea2-combo`**: a two-pass refine — a first pass, then
a low-denoise second pass (denoise **0.3**, 4 steps `euler`) — with the
**krea2 turbo LoRA** @0.2 on both passes. It saves *both* passes so you can see the
boost. The optional **IdeoKrea** LoRA (`Aitrepreneur/IdeoKrea`) is downloaded but
not wired by default — drop it into the Power Lora Loader's empty slot (start
\~0.5–1.0; it's a test LoRA) for the turbo + IdeoKrea Ideogram-style combo.

## A note on v1's safety-filter node

v1 shipped a `ConditioningKrea2Rebalance` node (the upstream author described it as
removing Krea 2's built-in safety filter), bypassed by default. **The V2 workflow
drops that node entirely**, so it's no longer in these packs — Krea 2's default
behavior is intact.

## Troubleshooting

* **`AssertionError: Torch not compiled with CUDA enabled`** — reinstall Torch
  for your CUDA version: `pip uninstall torch` then
  `pip install --upgrade torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128`
  (swap `cu128` for your CUDA, e.g. `cu126`).
* **`CLIPLoader: 'krea2' not in list`** — your ComfyUI predates native Krea 2
  support; update ComfyUI to ≥ v0.26.0.
* **Washed-out or noisy output** — check steps=8 / cfg=1; Turbo is distilled and
  is sensitive to over-stepping.

## FAQ

**Is Krea 2 free for commercial use?** Yes, under the Krea 2 Community License,
up to 50 seats; larger orgs need an enterprise agreement.

**Raw or Turbo?** Turbo for speed (8 steps); Raw if you want the un-distilled base
for fine-tuning or maximum fidelity.

**Does it need the Qwen models?** Yes — Krea 2 uses a Qwen3-VL 4B text encoder and
the Qwen image VAE, both included in the pack.

## Get it running in one command

Install the [comfyui-mcp](../) plugin, then ask your agent to install the
`krea2-txt2img-manual`, `krea2-txt2img-json`, or `krea2-combo` pack — nodes,
models, and a first render, hands-free.

Sources: [Krea 2 Technical Report](https://www.krea.ai/blog/krea-2-technical-report),
[Krea 2 Open Source](https://www.krea.ai/krea-2-open-source),
[Krea-2-Turbo on Hugging Face](https://huggingface.co/krea/Krea-2-Turbo),
[Krea 2 Community License](https://www.krea.ai/krea-2-licensing).