MOSS Image Sampler Tests: dpmpp_sde + radiance_shift Follows Prompts Best at cfg 3
mikemend · reddit · 2026-10-17
- The author benchmarked multiple samplers/schedulers on MOSS Image in ComfyUI beyond Euler.
- Usable samplers: heun, heunpp2, expheun2x0, dpmppsde, res2s, res2m. Usable schedulers: simple, sgmuniform, beta, normal, linearquadratic, beta57, sigmoidoffset, powershift, radianceshift, bongtangent.
- Best prompt adherence: dpmppsde/radianceshift and dpmpp2msde/radianceshift.
- Test setup: 8-step SFT lora at strength 0.7, 1536x1536, 28 s/image on an RTX 3090, good realism even at cfg 3.
- Gotchas: the stock workflow's shift node only produced noise with the ComfyUI models — delete or disable it; eulerancestralcfgpp, lcm and ersde skewed toward generating Asian women in testing.
Models and base workflow via Kijai's Comfy-Org/MOSS-Image on Hugging Face.
More from Multimodal
- Seedance 2.5 keeps Midjourney reference style remarkably well in video gen test — CurieuxExplorer · 2026-10-17
- Creative AI videos made with Gemini Omni & Flow, prompts shared — michaelrabone · 2026-10-17
- Midjourney style code --sref 1602964516 yields cinematic giant-monster imagery — tisch_eins · 2026-10-17
- CEB SAM 3D Body Loader for Blender goes free: full AI mocap & animation studio — ssh4net · 2026-10-17
- A Fully Reusable Video Prompt: Turning Four Reference Photos into a 15-Second Surrealist Short — umesh_ai · 2026-10-17
- Runway video packs cinematic destruction into 15 seconds, sparks realism debate — umesh_ai · 2026-10-17