kimi000/opal-garden-73

FLUX.2 Klein 4B DiffusionNFT Checkpoint

This repository contains a native Diffusers export of an internally trained FLUX.2 Klein Base 4B checkpoint. The public repository name is intentionally neutral; the exact training provenance is recorded below.

Provenance

• Source experiment: flux2_klein_base_4b_diffusionnft_dvreward_prompt_rubric_v4_3_qwen_hps_qwen3vl_16prompts_group14_7train_1dvreward_tp1_2node_512px_20step_cfg4_cw
• Source configuration: flux2_klein_base_4b_diffusionnft_dvreward_prompt_rubric_v4_3_qwen_hps_qwen3vl_16prompts_group14_7train_1dvreward_tp1_2node_512px_20step_cfg4_cw.yml
• Source run: flux2-v43-qwen-hps-qwen3vl-cw-formal
• Frozen run-code revision: cc30b9c80c8011cb74684318fe9ba62cdacad423
• Training checkpoint: step_500.pt
• Exported weight: EMA
• Source checkpoint SHA-256: cbc911b6bd173ec7815a842f4f0e1e07142bddd188e4bb2c25ffb48fdf64deb6
• Base model: black-forest-labs/FLUX.2-klein-base-4B
• Base revision: a3b4f4849157f664bdbc776fd7453c2783562f4d

Training Profile

• Prompt/rubric variant: v4.3 Qwen-Image/HPS/Qwen3-VL
• Resolution: 512 x 512
• Rollout steps: 20
• Guidance scale: 4.0
• Prompts per optimizer collection: 16
• Group size: 14
• Distributed topology: two nodes, each with seven policy ranks and one TP1 reward replica
• Reward model: Qwen3-VL-30B-A3B-Instruct DVReward
• LoRA rank / alpha: 32 / 64

HPS data is used to inform the prompt/rubric skills; no HPS reward model is loaded during training.

Export Validation

• Native pipeline: diffusers.Flux2KleinPipeline
• Native transformer: diffusers.Flux2Transformer2DModel
• Weight dtype: BF16
• Transformer parameters: 3,875,544,576
• Transformer shards: 9, each below 1 GB
• Text encoder shards: 9, each below 1 GB
• Strict transformer reload: passed
• Fully offline pipeline reload: passed
• Changed transformer tensors relative to Base: 60
• Changed transformer elements relative to Base: 2,444,775,823
• Transformer L2 delta from Base: 31.969998284036485
• Maximum absolute weight delta: 0.009521484375
• Deterministic 512 px smoke image: passed
• Smoke changed channel values relative to Base: 785,077
• Smoke mean absolute channel delta: 59.61954879760742

export_manifest.json, verification.json, and artifact_checksums.sha256 provide machine-readable provenance and integrity records.

Loading

import torch
from diffusers import Flux2KleinPipeline

pipe = Flux2KleinPipeline.from_pretrained(
    "kimi000/opal-garden-73",
    torch_dtype=torch.bfloat16,
)
pipe.to("cuda")

image = pipe(
    "A cinematic photograph of a red fox walking through a snowy forest",
    num_inference_steps=20,
    guidance_scale=4.0,
).images[0]

Users are responsible for complying with the upstream FLUX.2 model license and terms.