kimi000/quiet-cascade-58

FLUX.2 Klein Base 4B AlphaGRPO checkpoint

This is a complete native Diffusers Flux2KleinPipeline. The EMA LoRA is already merged into the transformer, so FAR and PEFT are not required for inference.

Source experiment: flux2_klein_base_4b_diffusionnft_dvreward_prompt_rubric_v4_3_1_unified_geneval_only_upgrade_4align4text2adaptive2aesthetic_75pct_16prompts_group14_7train_1dvreward_tp1_2node_512px_20step_cfg4_cw

Source checkpoint: step_500.pt

Training profile: 512px, 20 rollout steps, CFG 4, AlphaGRPO DVReward.

python demo.py --prompt "A red cube beside a blue glass sphere."