Version Base v0.1 Seasonal 39-Family, Step 600 EMA
Complete native BF16 diffusers.Flux2KleinPipeline. Trained EMA LoRA weights
(rank 32, alpha 64) are merged into the transformer. No FAR or PEFT runtime
is needed for inference.
This is the 39-family Version Base v0.1 Seasonal curriculum experiment, not the static
AlphaGRPO baseline. Training: 512px, 20 rollout steps, CFG 4, 16 prompts per
iteration, group size 14. 100pct denotes a target, not a verified
100% online-authored fraction.
Source run: vb-v01-isolated-lifecycle-20260913. W&B run: fb24f2db64306fd0bf7dc90c08823fa6.
Exact experiment, source checkpoint hash, source revision, configuration hash
and the complete family list are recorded in provenance.json and
export_manifest.json.
verification.json records strict offline reloading, nonzero parameter
differences from Base before and after serialization, and same-seed
512px/20-step image differences from Base. These checks validate the export;
they are not benchmark results or evidence of quality improvement.
import torch
from diffusers import Flux2KleinPipeline
pipe = Flux2KleinPipeline.from_pretrained(
"kimi000/cedar-cloud-83", 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]