kimi000/amber-river-83

Native Diffusers Model Export

This public repository contains a native Diffusers export of an EMA training checkpoint. The neutral repository name is intentional; complete provenance is recorded below.

Provenance

• Source experiment: zimage_base_diffusionnft_dvreward_prompt_rubric_v4_2_repetition_aware_75pct_28prompts_512px_16step_cfg4_cw
• Formal run: 20260825T155918Z-f7140a084f34
• Source checkpoint: models/step_500.pt
• Global step: 500
• Checkpoint weight: EMA
• Source checkpoint SHA-256: 5492cd71133dea4b8998998538adebd4c596806eca1dd467d7741477d361e6a2
• Base model: Tongyi-MAI/Z-Image
• Base revision: 47049b35ef3fa1478747b48bd0912ce28cdd7731
• Training profile: DiffusionNFT with prompt-rubric v4.2 repetition-aware DVReward, 75% setting, 28 prompts per collection, 512 px, 16 rollout steps, CFG 4
• LoRA merge: rank 256, alpha 256, merged into BF16 EMA transformer weights

Verification

The checkpoint and export were validated on CW:

• Checkpoint top-level keys: ema, global_step, model_state_dict_g, world_size
• EMA keys: 1003
• Exported transformer tensors: 521
• Exported transformer parameters: 6,154,908,736
• Native classes: diffusers.ZImagePipeline and diffusers.ZImageTransformer2DModel
• Strict transformer state loading: passed
• Offline Diffusers reload: passed
• Fixed scheduler shift: 6
• Dynamic shifting: disabled
• Transformer shards: 13
• Text encoder shards: 9
• Maximum shard size: 1 GB
• Deterministic 512 px smoke: passed
• Smoke changed channel values versus Base: 732,612
• Smoke mean absolute channel delta versus Base: 9.525258382161459
• Smoke maximum absolute channel delta versus Base: 212

See export_manifest.json, verification.json, and artifact_checksums.sha256 for machine-readable provenance and integrity data.

Loading

import torch
from diffusers import ZImagePipeline

pipe = ZImagePipeline.from_pretrained(
    "kimi000/amber-river-83",
    torch_dtype=torch.bfloat16,
).to("cuda")

pipe.scheduler.register_to_config(shift=6.0, use_dynamic_shifting=False)
if hasattr(pipe.scheduler, "set_shift"):
    pipe.scheduler.set_shift(6.0)

steps = 50
sigmas = torch.linspace(1.0, 1.0 / steps, steps).tolist()
image = pipe(
    "A cinematic photograph of a red fox walking through a snowy forest",
    height=1024,
    width=1024,
    num_inference_steps=steps,
    sigmas=sigmas,
    guidance_scale=4.0,
    generator=torch.Generator(device="cuda").manual_seed(0),
).images[0]

The included demo.py provides the same fixed-shift inference setup.