quiet-cascade-57

Diffusers Model Export

This public repository contains a native Diffusers export of an EMA training checkpoint. The neutral repository name is intentional; complete provenance is recorded below.

Provenance

• Source experiment: flux2_klein_base_4b_diffusionnft_dvreward_prompt_rubric_v4_3_visual_task_program_16prompts_group14_7train_1dvreward_tp1_2node_512px_20step_cfg4_nrt
• Formal run: flux2-v43-formal-70bfccfe
• Source checkpoint: models/step_500.pt
• Global step: 500
• Checkpoint weight: EMA
• Source checkpoint SHA-256: 1e8e9d503c7f6bdfedbc41ae8a2d2fd49b8b03beb796b27c8de3c965cb1a3f84
• Base model: black-forest-labs/FLUX.2-klein-base-4B
• Base revision: a3b4f4849157f664bdbc776fd7453c2783562f4d
• Training profile: DiffusionNFT with prompt-rubric v4.3 visual-task-program DVReward, 512 px, 20 rollout steps, CFG 4
• LoRA merge: rank 32, alpha 64, 60 adapter pairs, merged into BF16 EMA transformer weights

Verification

The export was strictly loaded with native Diffusers in offline mode and passed a deterministic 512 px smoke generation.

• Strict transformer load: passed
• Offline pipeline reload: passed
• Transformer tensors changed from Base: 60
• Changed transformer elements: 2,434,099,540
• Transformer L2 delta from Base: 30.570273992298265
• Maximum absolute parameter delta: 0.009368896484375
• Smoke mean absolute channel delta from Base: 51.595113118489586
• Transformer shards: 9
• Text encoder shards: 9
• Maximum shard size: 1 GB

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/quiet-cascade-57",
    torch_dtype=torch.bfloat16,
)
pipe.to("cuda")

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