OzzyGT/krea2_reference_blocks

Krea 2 reference-image modular blocks

Custom Modular Diffusers blocks that let Krea 2 take reference images. Its text encoder is the full Qwen3-VL, vision tower included, so a reference can condition the generation through the encoder's vision path without any extra model. Two community edit LoRAs go further and feed the reference into the transformer as clean VAE latents.

reference_mode picks the path:

reference_mode="off"       Qwen3-VL vision path only. No LoRA, works on the stock checkpoint
reference_mode="append"    vision + clean reference tokens after the target. Ostris style-reference LoRA
reference_mode="prepend"   vision + clean source before the target. Identity-Edit LoRA

With reference_mode="off" and no images, the blocks are stock text-to-image. No mode patches the transformer.

Loading & running

import sdnq  # needed to load the quantized text encoder
import torch

from diffusers import ModularPipeline
from diffusers.utils import load_image


pipe = ModularPipeline.from_pretrained(
    "OzzyGT/krea2_reference_blocks", trust_remote_code=True
)
pipe.load_components(dtype=torch.bfloat16)
pipe.to("cuda")

image = pipe(
    prompt="the same cat, sitting on a stone wall at sunset",
    reference_images=load_image("cat.png"),
    height=1024,
    width=1024,
    output="images",
)[0]
image.save("reference.png")

The LoRA modes need their adapter loaded first:

pipe.load_lora_weights("ostris/krea2_turbo_style_reference", weight_name="krea2_style_reference.safetensors")

image = pipe(
    prompt="the same subject in a snowy forest",
    reference_images=load_image("subject.png"),
    reference_mode="append",
    output="images",
)[0]

For loading a different checkpoint or swapping components, see Modular pipeline in the diffusers docs.

Reference inputs

reference_images            one image or a list
reference_subject_strength  scales content and subject. 1.0 default, 0 drops it
reference_style_strength    scales texture and style. 1.0 default, 0 drops it
grounding_px                longest side of the view the VLM sees, 768 default

Either strength takes one value for every reference, or a list matching them one to one, so a reference can give its look without its content or the other way round. grounding_px is the identity-versus-adherence dial: lower follows the prompt, higher holds the reference, 0 is native resolution.

Masks

reference_masks restricts a reference to part of the picture, painted being the part to use. One mask per image, or None for a slot without one.

reference_mask_mode decides what happens to the rest:

"exclude_blank"  default. Mask the vision attention and blank the masked-out tokens
"exclude"        mask the attention only, leaving the masked-out tokens in the sequence
"deemphasize"    blank the tokens only, which attenuates the region instead of removing it

Masks always apply to the vision path. In the LoRA modes the clean reference latents are unmasked by default, so the edit still sees the whole image and a masked edit can show hints of the unpainted parts. Set mask_reference_latents=True to drop those patches from the sequence too.

These blocks only work with Krea 2, Turbo or raw.

Credits

• ostris/krea2_turbo_style_reference -- the style-reference LoRA behind append
• conradlocke/krea2-identity-edit -- the Identity-Edit LoRA behind prepend
• Both recipes come from ai-toolkit; append follows its reference/edit setup, prepend its predict_velocity_edit.