How to run Krea 2 locally.

Turbo or Raw.

Krea 2 is Krea’s own 12.8B model, trained from scratch and released on 22 June 2026. It isn’t related to Flux.1 Krea Dev, the older model Krea made with Black Forest Labs. It comes in two versions.

• Turbo is the one to generate with. Eight steps, CFG 1, sizes from 1K to 2K.
• Raw is the undistilled base, meant for training LoRAs and further tuning. It takes 52 steps at CFG 3.5, follows a negative prompt, and was trained up to 1K.

Krea’s own advice is to train on Raw and run on Turbo. Users on the model pages agree that Raw alone makes weak pictures.

Files you need.

One model, one text encoder, one VAE. The encoder is Qwen3-VL 4B, and the VAE is the one from Qwen-Image.

Turbo

• Model krea2_turbo_fp8_scaled.safetensors ComfyUI/models/diffusion_models/ or krea2_turbo_int8_convrot, 13.5 GB. Full precision: krea2_turbo_bf16, 26.3 GB
  13.1 GB Download
• Text encoder qwen3vl_4b_fp8_scaled.safetensors ComfyUI/models/text_encoders/ full precision: qwen3vl_4b_bf16.safetensors, 8.9 GB
  5.2 GB Download
• VAE qwen_image_vae.safetensors ComfyUI/models/vae/
  0.3 GB Download

Raw

• Model krea2_raw_bf16.safetensors ComfyUI/models/diffusion_models/ smaller: krea2_raw_fp8_scaled 13.1 GB, krea2_raw_int8_convrot 13.5 GB. Same encoder and VAE
  26.3 GB Download
• LoRA krea2_turbo_lora_rank_64_bf16.safetensors ComfyUI/models/loras/ optional: makes Raw behave like Turbo
  0.5 GB Download

Which 13 GB file: in a 96-image test against bf16 on an RTX 4090, int8 convrot came closest, ahead of mxfp8 and fp8. int8 needs PyTorch built for CUDA 13.0 (cu130) to run fast, and it helped on AMD too. On an RTX 50-series card there is also krea2_turbo_nvfp4 at 7.7 GB, about 15% faster than fp8 on a 5090. Comfy-Org also hosts nine style LoRAs; three from the launch were pulled at Krea’s request.

models/
├── diffusion_models/
│   └── krea2_turbo_fp8_scaled.safetensors
├── text_encoders/
│   └── qwen3vl_4b_fp8_scaled.safetensors
└── vae/
    └── qwen_image_vae.safetensors

GGUF: not yet

realrebelai’s Krea 2 GGUFs run from 7.2 GB (Turbo Q4_K_M) to 13.6 GB (Q8_0). The main branch of ComfyUI-GGUF rejects their krea2 architecture (issue #464), so they need a fork for now. A Qwen3-VL 4B GGUF encoder loads, but without its vision part.

What fits your computer.

Turbo at about 1 megapixel. The text encoder runs first and ComfyUI moves it out of the way before sampling, so the model file sets the limit. A model that doesn’t fit still runs with part of it in system RAM, only slower.

• 6 GBNo

  Not a realistic fit. Z-Image Turbo or Flux.2 Klein 4B are.

• 8 GBOffloads

  Turbo fp8 with ComfyUI’s dynamic VRAM and at least 24 GB of system RAM. An RTX 3060 Ti takes about 30 s per image.

• 12 GBTight

  Turbo fp8: an RTX 3060 12 GB made 1280 × 720 in 37 s at 8 steps, peaking near 11.8 GB.

• 16 GBFits

  Turbo fp8 or int8 comfortably. The size most users settle on.

• 24 GBFits

  Turbo up to 2K, or Raw in fp8 or int8. An RTX 3090 makes a 1024 image in about 13 s.

• 32 GBFits

  Turbo or Raw at full precision (26.3 GB).

• Mac 32 GBNo

  The bf16 Turbo needs about 31 GB resident, and the smaller files don’t work on a Mac yet.

• Mac 64 GB+Fits

  Turbo in bf16: an M1 Max with 64 GB took about 212 s per 1024 image at 8 steps.

• 8 GB graphics cards
• RTX 3060 12 GB
• 16 GB cards
• RTX 5060 Ti 16 GB
• 24 GB cards
• Mac with Apple Silicon

Set it up.

1. 1

   Update ComfyUI

   Krea 2 support arrived in ComfyUI at launch, in June 2026. Anything older doesn’t know the krea2 encoder type. ComfyUI Desktop updates itself; the portable build has an update script. In a manual install:

   Terminal, in the ComfyUI folder Copy

   git pull
   pip install -r requirements.txt

2. 2

   Download the three files

   Model, text encoder and VAE from the lists above. Comfy-Org’s repo needs no sign-in.

3. 3

   Put them in their folders

   Model in diffusion_models, encoder in text_encoders, VAE in vae. Restart ComfyUI so the loaders list them.

4. 4

   Open the Krea 2 template

   In the template browser, pick Krea-2: Text to Image, or Krea-2 Int8: Text to Image for the int8 file. It comes with a style LoRA and an LLM prompt enhancer switched on; bypass either if you don’t want them. If it says it could not load subgraphs, update ComfyUI’s frontend.

5. 5

   Check the loaders

   Load Diffusion Model gets the Krea 2 file. Load CLIP gets the Qwen3-VL 4B encoder with type krea2. Load VAE gets qwen_image_vae. Turbo needs no ModelSampling node: ComfyUI’s default shift for Krea 2, 1.15, is Krea’s own Turbo value.

6. 6

   Write a long prompt

   Krea 2 likes long, detailed, natural sentences. Put text you want in the picture in quotes. Negations like “without flame” tend to fail; describe what you want instead, such as “an unlit torch”.

Settings that work.

Turbo

Steps

8

CFG

1

Sampler

euler

Scheduler

simple

Size

1024 to 2048

Shift

1.15, default

Negative

none

Text encoder

Load CLIP, krea2

Krea’s README runs Turbo at 8 steps with CFG off and a fixed shift of 1.15, which is what Comfy’s template does. For 2K, stay near 2 megapixels.

Raw

Steps

52

CFG

3.5

Sampler

euler

Scheduler

simple

Size

up to 1024

Shift

from image size

Negative

yes

Latent

EmptyLatentImage

These are the settings in Krea’s README, and ComfyUI’s Krea 2 docs also say 52 steps. The 28 steps and CFG 4.5 in Krea’s inference.py are generic defaults, not the Raw recipe. Raw’s shift grows with the image size, from 0.5 to 1.15; in ComfyUI, add ModelSamplingFlux with max shift 1.15 and base shift 0.5. For the negative prompt, replace the template’s ConditioningZeroOut with a second text encode.

The Krea2T Enhancer

ComfyUI-Krea2T-Enhancer is an optional community node by capitan01R. It patches Krea 2’s text path to follow the prompt more closely. It goes between the model loader and the sampler, with strength 1.0 by default, from 0 (off) to 2. The author calls it experimental and suggests starting at 1.0 and going lower when you stack LoRAs. The pack also has a node for (phrase:weight) prompt weighting, since normal weighting doesn’t map onto Krea 2.

How fast.

Times after the first run, which loads the models. The 4090 reached 1.9 it/s with ComfyUI’s --fast flag. On AMD, try the int8 file.

When it goes wrong.

Out of memory on 8 to 12 GB

Use the fp8 or int8 file, keep dynamic VRAM on, and have 24 to 32 GB of system RAM.

The template says it could not load subgraphs

The frontend is older than the template. Update ComfyUI’s frontend, or build the plain graph: loaders, text encode, empty latent, KSampler, VAE decode.

Raw makes mush at any CFG

Use 52 steps at CFG 3.5, keep it at 1024 or below, or try the Turbo LoRA at 0.6 as above.

Black images on a Mac

Start ComfyUI with --fp16-vae. If the fp8 file is rejected, use the bf16 one.

ComfyUI-GGUF says the architecture is unknown

Its main branch doesn’t support krea2 yet. Use the safetensors files.

Template LoRAs missing (warmpastel, plasmoid, coolblue)

Krea had them removed. Pick one of the other style LoRAs.

Raw’s shift stays after you bypass ModelSamplingFlux

A dynamic VRAM bug on low-memory cards. Restart ComfyUI, or start it with --disable-dynamic-vram.

Raw is a base for training, and the Turbo LoRA trick came out of the first thread. The second found 8 GB workable with dynamic VRAM and 24 GB of system RAM.

Questions.

Sources: Krea 2 repository and README, Krea 2 Turbo model card, Krea 2 licensing, ComfyUI Krea 2 tutorial, Krea 2 speed and quality thread [1], comfylab Krea 2 guide [2], RTX 3060 low-VRAM test [3], Krea 2 Turbo VRAM thread [4], Krea 2 on an M1 Max [5], quantisation comparison, Krea2T Enhancer.