Krea 2 is licensed under the Krea 2 Community License Agreement.
For more information, visit https://krea.ai/krea-2-licensing.

------------------------------------------------------------------------
This repository contains a MODIFIED DERIVATIVE of the Krea Model
(krea/Krea-2-Turbo). Modifications by avlp12 (ALISVOLATPROPRIIS):

  - Re-implementation of the Krea 2 Turbo transformer, Qwen3-VL-4B text
    encoder, and flow-matching sampler for Apple MLX.
  - 8-bit (group-size 64) quantization of the transformer's 28
    attention + SwiGLU blocks (sensitivity-graded; all other paths bf16).

This is an independent, unofficial port. It is NOT an official Krea
product and is NOT endorsed by or affiliated with Krea.ai, Inc.

A full copy of the Krea 2 Community License Agreement is included in the
LICENSE file in this repository. Your use of these weights is governed by
that agreement (including, e.g., the < $1,000,000 USD annual-revenue
threshold for Commercial Use and the content-filtering requirements).
------------------------------------------------------------------------

------------------------------------------------------------------------
FURTHER MODIFICATIONS in the krea2-mlx-studio fork, by Casey Thomas
(github.com/Cthomasdesign):

  - LoRA support: a runtime low-rank adapter applied to the (quantized)
    transformer (krea2/lora.py), Krea2Pipeline.set_loras(), CLI flags,
    and web-UI controls. The key-mapper handles both the official Krea /
    diffusers and the ai-toolkit / PEFT LoRA naming conventions.

This fork is independent and unofficial. It is NOT endorsed by or
affiliated with avlp12 (ALISVOLATPROPRIIS) or Krea.ai, Inc.
------------------------------------------------------------------------

------------------------------------------------------------------------
FURTHER MODIFICATIONS in this repository (ComfyUI-Krea2-MLX), by Casey
Thomas (github.com/Cthomasdesign):

  - The `krea2` engine (transformer, text encoder, sampler, LoRA, safety
    filter) is VENDORED into this repository as `krea2_engine/`, so this
    node package has no external git dependency on krea2-mlx-studio.
    The unused HF-Transformers-based text encoder variant
    (text_encoder_hf.py) and the auto-download-by-precision helpers
    (BUILDS/resolve_weights) were dropped as dead code not used by the
    ComfyUI nodes, which always load an explicit local transformer file
    chosen via the model loader node.
  - Performance modifications to the vendored engine (v0.4.0):
    native grouped-query attention in the fused SDPA kernels (no K/V
    head repetition); step-invariant conditioning (text fusion, RoPE
    tables, attention masks) hoisted out of the denoising loop; padded
    text tokens trimmed before the DiT; dynamic-length text encoding
    (pad to longest-in-batch with preserved absolute RoPE positions);
    and a per-prompt embedding cache in the pipeline. Mathematically
    faithful to the reference computation (validated); note that the
    sequence-length changes alter bf16 kernel rounding order, so fixed
    seeds render equal-quality but not bit-identical images vs earlier
    versions. Set KREA2_EXACT_LEGACY=1 for the pre-0.4 exact behavior.
  - ComfyUI node wrappers (nodes.py): model loading from
    ComfyUI/models/krea2, LoRA stacking from ComfyUI/models/loras,
    generation with progress/cancel support, and memory unloading.
  - Image-conditioning (v0.5.0+): img2img and in-context edit paths,
    and (v0.7.0+) image-grounded text conditioning via the Qwen3-VL
    vision tower.

------------------------------------------------------------------------
THIRD-PARTY CODE (krea2_engine/vision/): the Qwen3-VL vision tower
implementation is vendored from mlx-vlm (github.com/Blaizzy/mlx-vlm),
Copyright © 2025 Prince Canuma, MIT License:

  MIT License. Permission is hereby granted, free of charge, to any
  person obtaining a copy of this software and associated documentation
  files (the "Software"), to deal in the Software without restriction,
  including without limitation the rights to use, copy, modify, merge,
  publish, distribute, sublicense, and/or sell copies of the Software,
  and to permit persons to whom the Software is furnished to do so,
  subject to including the above copyright notice and this permission
  notice in all copies or substantial portions of the Software. THE
  SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND.

The vendored files (krea2_engine/vision/vision.py, config.py) are
adapted only to be self-contained (inlined a small SDPA helper, local
config import). The vision weights themselves are part of the Qwen3-VL
text encoder shipped with krea/Krea-2-Turbo.
------------------------------------------------------------------------

This integration is independent and unofficial. It is NOT endorsed by or
affiliated with avlp12 (ALISVOLATPROPRIIS) or Krea.ai, Inc.
------------------------------------------------------------------------