Same prompt, 3 models: SDXL vs Krea 2 Turbo vs MOSS-Image-SFT anime comparison
lajonquillebleu · reddit · 2026-10-18
A Reddit user ran a controlled anime image comparison across SDXL (WAI-Illustrious v17), Krea 2 Turbo, and MOSS-Image-SFT using identical booru-tag prompts — no style prefixes, quality tags, negative prompts, or LoRAs. Four scenes × two seeds at 1024×1024:
- SDXL (WAI-Illustrious v17): ComfyUI, 28 steps, CFG 5, Euler a, CLIP skip 2, on an L40S
- Krea 2 Turbo: 4-step turbo q8 via mflux, running locally on a Mac, guidance 1.0
- MOSS-Image-SFT: 5B, 50 steps, CFG 3, uncond skip layer 22, on an H100
Each model ran at its recommended defaults, so step counts aren't equalized. All four prompts (cherry blossom portrait, rooftop night scene, rainy cafe pair, forest sword action) are included for direct comparison of composition, lighting, and motion handling.
More from Multimodal
- Head-to-head: OpenAI's new image model blue edges out 2.5 flare in Berlin police test — flowersslop · 2026-10-18
- Uncensored Qwen-Image LoRA collection trends on Hugging Face — raubatz · 2026-10-18
- AI Recreation of Original Footage Draws Community Attention — mhu99 · 2026-10-18
- AI video is now so real its own creator can't tell where fake ends — Aiden_Tech_Ai · 2026-10-18
- AI image creativity: a lightbulb you grow, prompt included — umesh_ai · 2026-10-18
- Y2K designer toy blind-box portrait prompt shared with three sample renders — azed_ai · 2026-10-18