Qwen Image 2.1 Turbo 8-step vs original 25-step image quality showdown: 11 test sets in one graph. Proves the turbo LoRA 8-step schedule holds almost par with the original 25-step pipeline while generating far faster.
Main features:
- 11 test branches; each runs the Turbo 8-step pipeline (Turbo LoRA rank 178 + fixed 8-step ManualSigmas schedule, CFG 1, euler) against the original 25-step pipeline (KSamplerAdvanced 25 steps, CFG 3.5, euler) on the same prompt and seed
- rgthree Image Comparer pairs every 8-step result with its 25-step counterpart for direct side-by-side judging
- 11 pre-configured prompt / subject variations covering portraits, scenes and creative subjects
- Final SaveImage (ComfyUI prefix) for the selected outputs
Suggested workflow:
1. Keep the 11 built-in test prompts for the comparison, or swap in your own
2. Run the graph; both pipelines render per branch
3. Judge in the comparer grid: 8-step results track the 25-step originals at a fraction of the time
Dependencies:
- Qwen-Image_2.1.safetensors (base UNET)
- qwen3vl_8b_int8_convrot.safetensors (CLIP)
- qwen_image_2.1_vae_bf16.safetensors (VAE)
- qwen_image_2.1_turbo_lora_avg_rank_178_bf16.safetensors (Turbo LoRA)
- rgthree Node Suite (Image Comparer)
RunningHub workflow page: https://www.runninghub.ai/zh-cn/post/2109550643710304258?inviteCode=rh-v1111
Bilibili tutorial video: https://www.bilibili.com/video/BV14zp96REo9/
RunningHub registration with invite code rh-v1111 (1000 RH coins): https://www.runninghub.ai/?inviteCode=rh-v1111