Forost/grim-all-v3-zimage-lora

Grim All v3 — Z-Image LoRA

LoRA trained from Tongyi-MAI/Z-Image (non-Turbo) with AI Toolkit v0.11.2.

Available weights

• grim_all_v3_zimage_lora_epoch1.safetensors — first complete epoch, step 11,738.
• grim_all_v3_zimage_lora_epoch2.safetensors — second complete epoch, cumulative step 23,476.

Both files contain 480 LoRA tensors. Epoch 1 and epoch 2 are preserved as separate immutable weights so they can be compared directly.

Training summary

• LoRA rank / alpha: 64 / 64
• Resolution: 1024 px with aspect-ratio buckets
• Optimizer: AdamW 8-bit, weight decay 1e-4
• Learning rate: 5e-5
• Precision: BF16
• Transformer and text encoder loading: qfloat8
• Dataset: 11,735 images
• Caption presentations per image: one short caption and one long caption
• Effective examples per epoch: 23,470
• Batch size: 2
• Text encoder: frozen
• Gradient checkpointing: enabled

The source images were restored with HYPIR and selectively color graded. Short and long caption presentations were both included explicitly in every epoch.

Usage

import torch
from diffusers import ZImagePipeline

pipe = ZImagePipeline.from_pretrained(
    "Tongyi-MAI/Z-Image",
    torch_dtype=torch.bfloat16,
).to("cuda")

pipe.load_lora_weights(
    "Forost/grim-all-v3-zimage-lora",
    weight_name="grim_all_v3_zimage_lora_epoch2.safetensors",
)

To use the earlier checkpoint, replace the weight name with grim_all_v3_zimage_lora_epoch1.safetensors.