kimi000/opal-harbor-73

Version Base 39-Family, Step 500 EMA

Native BF16 diffusers.Flux2KleinPipeline with the trained EMA LoRA merged into the transformer (rank 32, alpha 64). No adapter loading is required.

This is the Version Base AlphaGRPO-style 39-family curriculum run, not the static AlphaGRPO baseline. Training uses 512px, 20 rollout steps, CFG 4, 16 prompts per iteration and group size 14. The 100pct_target name denotes a training target, not a verified 100% online-authored fraction.

Exact source experiment, run, checkpoint hash and revision are recorded in provenance.json and export_manifest.json. The source is the r18 reward-valid CW run at iteration 500, W&B run 65cyigcp.

verification.json records strict offline reloading, parameter differences from Base before and after serialization, and same-seed 512px / 20-step image differences from Base. These are export integrity checks, not benchmarks.

import torch
from diffusers import Flux2KleinPipeline

pipe = Flux2KleinPipeline.from_pretrained(MODEL_ID, torch_dtype=torch.bfloat16).to("cuda")
image = pipe(prompt="A red ceramic teapot beside two blue cups on a wooden table.",
             height=512, width=512, num_inference_steps=20, guidance_scale=4.0,
             generator=torch.Generator("cuda").manual_seed(0)).images[0]