lora_rank64 lora_alpha64 timestep_shift_mu1.15 key_formatcomfyui (diffusion_model.*, lora_down/lora_up) lora_scale1.0 methodprogressive distillation (teacher: Krea 2 Turbo, 8-step, guidance 0.0) formatpt checkpointchk00010000 training_samples10000 descriptionA LoRA for Krea 2 Turbo that reduces the minimum usable step count from 8 to 4 while maintaining good quality results. Load it on top of Krea 2 Turbo, run 4 steps instead of 8, keep guidance at 0.0. base_modelkrea/Krea-2-Turbo training_datahttps://huggingface.co/datasets/Lakonik/t2i-prompts-3m trained_modules224 block linears (attn q/k/v/gate/out, ff gate/up/down x28) + 4 global (time_embed.linear_1, time_embed.linear_2, time_mod_proj, final_layer.linear) comfyui_cfg1.0 converted_fromkrea2_turbo_4step_rank_64_lora_chk00010000.safetensors hardwaresingle RTX 3090 24GB schedulersimple source_formatdiffusers (lora_A/lora_B) guidance_scale0.0 inference_steps4 samplereuler statuswork in progress