kimi000/misty-harbor-64

Z-Image Step-500 EMA Diffusers Export

This repository contains a native Hugging Face Diffusers export of a trained Z-Image checkpoint. The public repository name is intentionally neutral; full experiment provenance is recorded below for reproducibility.

Provenance

• Source experiment: zimage_base_diffusionnft_dvreward_prompt_rubric_v4_3_visual_task_program_75pct_28prompts_512px_16step_cfg4_draco
• Formal run: zimage-v43-formal-605
• Source checkpoint: models/step_500.pt
• Global step: 500
• Exported weights: EMA (1,003 EMA entries)
• Source checkpoint SHA-256: 479827b2c5ee7d5ce3b846e59c61fb2f836d04773c82d1402b1bfbc4c76d09be
• Base model: Tongyi-MAI/Z-Image
• Base source revision: b7021aa8a00748a3b536a6feb5975bfe29311329
• Training profile: prompt-rubric v4.3 visual-task-program, 75% setting, 28 prompts per collection, 512 px, 16 rollout steps, CFG 4
• LoRA rank / alpha: 256 / 256

The EMA LoRA weights were merged into the native BF16 Z-Image transformer. The export contains 521 transformer tensors and 6,154,908,736 transformer parameters.

Format

• Pipeline: diffusers.ZImagePipeline
• Transformer: diffusers.ZImageTransformer2DModel
• Precision: BF16
• Scheduler shift: 6.0
• Dynamic shifting: disabled
• Transformer weights: 13 shards
• Text encoder weights: 9 shards
• Maximum shard size: below 1 GB

Verification

The export was reloaded fully offline with stock Diffusers and generated a 512 x 512 smoke image using 4 inference steps, CFG 4, and seed 0. A same-seed Base-model render was generated for comparison:

• Changed channel values: 738,072
• Mean absolute channel delta: 9.250560760498047
• Maximum absolute channel delta: 250

These nonzero deltas verify that the exported transformer contains the trained Step-500 EMA weights rather than an unchanged Base transformer. verification.json, smoke images, the export manifest, and SHA-256 checksums are included.

Loading

import torch
from diffusers import ZImagePipeline

pipe = ZImagePipeline.from_pretrained(
    "kimi000/misty-harbor-64",
    torch_dtype=torch.bfloat16,
)
pipe.scheduler = pipe.scheduler.from_config(
    pipe.scheduler.config,
    shift=6.0,
    use_dynamic_shifting=False,
)
pipe.enable_model_cpu_offload()

image = pipe(
    prompt="A cinematic photograph of a red fox walking through a snowy forest",
    height=1024,
    width=1024,
    guidance_scale=4.0,
    num_inference_steps=50,
).images[0]

See export_manifest.json and verification.json for machine-readable export details.