SumomoLee/Krea2-Identity-Long-Caption-R32-3ep

Krea2 Identity Long Caption R32 3ep

This repository contains a Krea 2 LoRA trained for long-caption adherence, artist identity, character identity, and broader visual coverage. It is not a standalone diffusion model and requires the krea/Krea-2-Raw base model.

Model details

• Base model: krea/Krea-2-Raw
• Network implementation: Musubi Tuner networks.lora_krea2
• Rank / alpha: 32 / 32
• Trainable Krea 2 modules: 257
• Text-fusion low-LR modules: 33
• LoRA+ ratio: 2.5
• Main learning rate: 1e-4
• Text-fusion learning rate ratio: 0.02
• Precision: BF16
• Final saved step: 16932 (configured terminal step: 16929)
• Final average loss reported by the training run: 0.08579

Excluded module patterns:

first
last\.linear
tmlp\..*
txtmlp\..*
tproj\.1

Training data notes

The training mixture combines curated illustration data with approximately 10,000 aesthetic real-world images. Artist and character identities are represented in natural-language captions, including @artist identity tokens where applicable.

For the final continuation from step 14000 to completion, images carrying a structured ai-generated tag were excluded. Earlier training stages may have included those images. The final continuation dataset reported 171,564 repeated training items and 5,330 optimizer batches per epoch across four GPUs.

Usage

Load the .safetensors file as a Krea 2 LoRA with Musubi Tuner's networks.lora_krea2 network implementation. The LoRA is intended for the raw Krea 2 base model. It is not a replacement for the base DiT, VAE, or text encoder.

The model does not use one universal trigger word. Prompts can use the artist and character identities present in the training captions, including artist tokens written in the @artist_name form.

Experimental PatchGAN test checkpoint

krea2_identity_long_caption_r32_3ep_patchgan_test_step500.safetensors is an experimental test checkpoint derived from the primary 3ep LoRA. It was trained with a DINOv3 PatchGAN objective to test whether adversarial tuning can restore skin, hair, fabric, metal, glass, and other fine real-world textures that may be softened by illustration-heavy SFT data.

This file contains a complete replacement LoRA state. Load it instead of krea2_identity_long_caption_r32_3ep.safetensors; do not stack the two LoRAs. The checkpoint was selected at 500 discriminator updates (166 generator updates) because later checkpoints began adding meaningless high-frequency detail.

• Stored precision: BF16

This checkpoint is provided strictly as a test. Although it can produce more visible local texture, it may also exaggerate pores and wrinkles, add scratches or colored speckles, alter facial appearance, and over-detail anime linework or decorations. The primary 3ep LoRA remains the default, more conservative model.

Experimental checkpoint integrity:

SHA256 527e6d6a1150f9edba91574a08f26e721937498b54f1968040726c26dc527569

Integrity

SHA256 05f423c7bc50e8ffeadd0e16007d1e8582cbf7960c375aa58856cefbf604e36a

Limitations and license

The training distribution remains illustration-heavy despite the real-world supplement. Artist and character rendering quality varies with representation, visual complexity, and prompt composition.

Use of this LoRA is subject to the license and usage terms of the Krea 2 base model. The base model declares an other license on Hugging Face.