velvet-comet-51

Velvet Comet 51

Native Diffusers export of a trained FLUX.2 Klein Base 4B checkpoint.

Provenance

• Source experiment: flux2_klein_base_4b_diffusionnft_alphagrpo_dvreward_alphagrpo20k_512px_20step_cfg4_16prompts_group14_7train_1dvreward_tp1_2node_nrt
• Source checkpoint: models/step_500.pt
• Global step: 500
• Export weights: EMA
• Source checkpoint SHA-256: 2fc9365453f4ff19ce28493d83361dcb94bbe90b9474eabd1a3c65a3a13eb00c
• Base model revision: a3b4f4849157f664bdbc776fd7453c2783562f4d
• LoRA rank / alpha: 32 / 64
• Export format: diffusers.Flux2KleinPipeline, BF16, max shard size 1GB

Verification

The exporter matched and merged 60 default LoRA pairs. Relative to the Base transformer, 60 tensors and 2,447,174,844 elements changed; maximum absolute parameter delta was 0.0116424560546875. Offline reload reproduced the same deltas. The fixed-seed 512px smoke image differed from Base in 780,590 channel values.

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

Loading

import torch
from diffusers import Flux2KleinPipeline

pipe = Flux2KleinPipeline.from_pretrained(
    "kimi000/velvet-comet-51",
    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]