Modelpublic

kimi000/copper-reef-47

sourceHugging Faceotherupdated 22d agoView on Hugging Face

0likes14downloads

Model Card

FLUX.2 Klein 4B Step-500 EMA Diffusers Export

This repository contains a complete native Hugging Face Diffusers export of a trained FLUX.2 Klein Base 4B checkpoint. The public repository name is intentionally neutral; full experiment provenance is recorded below for reproducibility.

Provenance

—Source experiment: flux2_klein_base_4b_diffusionnft_dvreward_prompt_rubric_v4_3_geneval_composition_16prompts_group14_7train_1dvreward_tp1_2node_512px_20step_cfg4_nrt
—Experiment variant: prompt-rubric v4.3, GenEval composition-only
—Formal run: flux2-v43-geneval-formal-0903f430-r2
—Source run code revision: 0903f430244d330b7c18bdba2f9560491b1a6898
—Source checkpoint: models/step_500.pt
—Global step: 500
—Exported weights: EMA
—Source checkpoint SHA-256: 8fa1aef3c311f422f407eb0b06b956581da86cc877a76d3e372c3e6ee608c1b1
—Base model: black-forest-labs/FLUX.2-klein-base-4B
—Base revision: a3b4f4849157f664bdbc776fd7453c2783562f4d
—Training profile: 16 prompts, group size 14, 512 px, 20 rollout steps, CFG 4, 7 policy ranks and 1 DVReward replica per node across 2 nodes
—LoRA rank / alpha: 32 / 64

The EMA LoRA weights were merged into the native BF16 FLUX.2 transformer. The export contains 169 transformer tensors and 3,875,544,576 transformer parameters. DiffusionNFT old-policy training-state tensors were excluded from the inference artifact.

Format

—Pipeline: diffusers.Flux2KleinPipeline
—Transformer: diffusers.Flux2Transformer2DModel
—Precision: BF16
—Scheduler: 1,000 training timesteps, dynamic shifting enabled, deterministic sampling
—Transformer weights: 9 shards
—Text encoder weights: 9 shards
—Maximum shard size: below 1 GB

Verification

The export was strictly reloaded offline with stock Diffusers and generated a 512 x 512 smoke image using 4 inference steps, CFG 4, and seed 0. The merged transformer differs from Base in 60 tensors and 2,422,759,106 elements, with L2 delta 28.82849469865312 and maximum absolute delta 0.008819580078125.

A same-seed Base render was also generated:

—Changed channel values: 730,266
—Mean absolute channel delta: 28.308038075764973
—Maximum absolute channel delta: 249

These checks 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

python

import torch
from diffusers import Flux2KleinPipeline

pipe = Flux2KleinPipeline.from_pretrained(
    "kimi000/copper-reef-47",
    torch_dtype=torch.bfloat16,
)
pipe.enable_model_cpu_offload()

image = pipe(
    prompt="A red cube beside a blue glass sphere.",
    height=1024,
    width=1024,
    guidance_scale=4.0,
    num_inference_steps=20,
).images[0]

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