README.md

---
library_name: diffusers
tags:
  - modular-diffusers
  - krea
  - image-to-image
  - image-editing
base_model: krea/Krea-2-Turbo
---

# Krea 2 reference-image edit — Modular Diffusers blocks

Custom [Modular Diffusers](https://huggingface.co/docs/diffusers/main/en/modular_diffusers/overview)
blocks that reproduce the [`ostris/Krea2OstrisEdit`](https://huggingface.co/ostris/Krea2OstrisEdit)
reference-image ("edit") workflow for **Krea 2**, loadable as remote code on top of stock `diffusers`.

```python
import torch
from transformers import Qwen3VLProcessor
from diffusers import ClassifierFreeGuidance
from diffusers.modular_pipelines import ModularPipelineBlocks

blocks = ModularPipelineBlocks.from_pretrained("diffusers-modular/krea2-edit", trust_remote_code=True)
pipe = blocks.init_pipeline("krea/Krea-2-Turbo")        # weights from the base repo
pipe.load_components(torch_dtype=torch.bfloat16)
pipe.update_components(processor=Qwen3VLProcessor.from_pretrained("Qwen/Qwen3-VL-4B-Instruct"))
pipe.update_components(guider=ClassifierFreeGuidance(guidance_scale=0.0, use_original_formulation=True))
pipe.to("cuda")

from PIL import Image
image = pipe(
    prompt="a white yeti with horns reading a book",
    image=Image.open("reference.png"),   # one or more reference images
    num_inference_steps=8, mu=1.15, output="images",
)[0]
```

## What it does

Reference images condition generation two ways (matching how the Ostris AI-Toolkit edit LoRAs train):

1. **Qwen3-VL prompt embedding** — a coarse view of each reference is embedded into the text
   conditioning through the vision tower.
2. **Clean VAE latents at flow time t=0** — each reference is VAE-encoded and appended to the
   transformer sequence as clean tokens on its own rotary frame axis, so the noisy image tokens
   attend to it at every block.

The bundled `Krea2Transformer2DModel` adds a small, backward-compatible `ref_seq_len` argument to the
Krea 2 transformer forward (t=0 modulation of the reference span; those tokens are excluded from the
predicted velocity). With `ref_seq_len=0` it is numerically identical to plain Krea 2 text-to-image.

## Files

- `block.py` — entry point (`Krea2EditBlocks`), referenced by `config.json`'s `auto_map`.
- `transformer_krea2.py` — the Krea 2 transformer (with the `ref_seq_len` edit path).
- `modular_blocks_krea2*.py`, `encoders.py`, `before_denoise.py`, `denoise.py`, `decoders.py`,
  `inputs.py`, `modular_pipeline.py` — the modular blocks.