---
license: apache-2.0
library_name: diffusers
pipeline_tag: text-to-image
base_model:
  - RudySen/Krea2-Realism-Checkpoint
tags:
  - krea2
  - text-to-image
---

# Krea2 Realism (diffusers transformer)

A diffusers-format copy of the transformer from [RudySen/Krea2-Realism-Checkpoint](https://huggingface.co/RudySen/Krea2-Realism-Checkpoint) (`Krea2Realism_V1.safetensors`), a realism finetune of [Krea 2 Turbo](https://huggingface.co/krea/Krea-2-Turbo) by **RudySen**. All credit for the model goes to RudySen and Krea.

The original FP8 ComfyUI file was dequantized (per-tensor `weight_scale`) to bfloat16 and renamed to the diffusers `Krea2Transformer2DModel` layout. No weights were retrained or merged.

This repo only holds the transformer. Use the text encoder, tokenizer, VAE and scheduler from Krea 2 Turbo.

```python
import torch
from diffusers import Krea2Pipeline, Krea2Transformer2DModel

transformer = Krea2Transformer2DModel.from_pretrained(
    "mrfakename/Krea2-Realism-diffusers", subfolder="transformer", torch_dtype=torch.bfloat16
)
pipe = Krea2Pipeline.from_pretrained(
    "krea/Krea-2-Turbo", transformer=transformer, torch_dtype=torch.bfloat16
).to("cuda")

image = pipe(
    "a candid street photo of an old man reading a newspaper at a cafe, soft morning light",
    height=1792, width=1152, num_inference_steps=6, guidance_scale=0.0,
).images[0]
```

Recommended settings from the original card: 6 steps (4 to 8), no CFG (`guidance_scale=0.0` in diffusers, CFG 1.0 in ComfyUI), about 2MP (for example 1152 × 1792).

Use of the base model is also subject to the [Krea 2 Community License](https://krea.ai/krea-2-licensing) and the [Krea Acceptable Use Policy](https://www.krea.ai/krea-2-use-policy).