jayark/sammy-krea-v4-loras

Sammy Krea 2 LoRAs

Two separate Krea 2 RAW LoRAs trained for the Sammy animation project.

Weights

File	Status	Purpose	Trigger	SHA-256
sammy-character-krea2-200.safetensors	Current	Sammy character identity	sammy	b30d4439b6dfcb6f7488a80db0f701c1089d8573f9f5e6caf149e31a7b7dce18
sammy-character-krea2-1000.safetensors	Archived	Prior Sammy character checkpoint	sammy	50947ce7a55b8a515a4f987b632b93e2ff0104c6c33639368550ece9c61ef1b2
sammy-style-krea2-1000.safetensors	Current	Sammy visual style	sammy style	d78f8aa878405f0b67716ac5b3345afde671e429e921d709177ee4533356511e

Training recipe

• Base model: krea/Krea-2-Raw
• Trainer: musubi-tuner 0.3.4, commit 8934cfb
• LoRA type: lora_krea2
• Rank and alpha: 64 / 64
• Learning rate: 3e-4
• Optimizer: AdamW8bit
• Precision: bf16
• Resolution: 1024, bucketed
• Current character checkpoint: step 200
• Current style checkpoint: step 1,000
• Character dataset: 25 image-caption pairs
• Style dataset: 27 image-caption pairs

Inference

The weights use musubi-tuner's native Krea 2 LoRA format. The production route is the private underscoreanimation/sammy-krea-lora Replicate model, which exposes named sammy-character, sammy-style, and none adapters. A deterministic test found that Fal's Krea LoRA endpoint accepted these files without applying their weights.

The repository contains weights and training metadata only. Training images and captions are excluded.