Dissertation xi 11.6 Ablation studies in step distillation (best viewed in color). For each line, from left to right, the CFG scales starts from 1.0 to 10.5 with interval 0.5. (a) To obtain the same 8-step student model, in direct distillation, the teacher only distills once (16 → 8), while progressive distillation [9, 10] starts from the 64-step teacher, distills 3 times to 8 steps (64 → 32 → 16 → 8). (b) w-conditioned model [10] struggles at achieving high CLIP scores (such as over 0.30) while the original SD-v1.5 and our distilled 8-step SD-v1.5 can easily achieve so. (c) Comparison between vanilla distillation loss L vani dstl , the proposed CFG distillation loss L cfg dstl , and their mixed version L dstl . (d) Effect of adjusting the two hyper-parameters, CFG range and CFG probability, in CFG distillation. As seen, these hyper-parameters can effectively tradeoff FID and CLIP score.