TailSteer: Drift-Resistant Multi-Turn Image Editing from CUHK MMLab
MMLab-CUHK · hf · 2026-11-06
TailSteer, from CUHK MMLab with SenseTime Research and Shanghai AI Lab, tackles cumulative visual drift in multi-turn image editing. The final-stage training framework keeps early denoising dynamics close to the base model and learns tail-focused preservation corrections from on-policy rollouts, preserving pretrained generative and editing abilities. Comparisons show stable results across local editing, identity preservation, and mixed-type editing over many turns.
More from Multimodal
- Shared ComfyUI base workflow for the MOSS image model gets community love — Ill_Resolve8424 · 2026-11-06
- 2D-to-3D character design: a flat warrior sketch turned photoreal 3D render — azed_ai · 2026-11-06
- AI short film "Metora's Natural Selection" showcases current video gen — finalbosu · 2026-11-06
- Home motion capture in 15 minutes with GVHMR, Blender MCP and Claude — sidahuj · 2026-11-06
- Dream II: a surreal 1940s noir horror short film made with AI — Material-Wonder-4624 · 2026-11-06
- Tencent Hy4 renders a Millennium Falcon in 2h16m for just $2.50 — TencentHunyuan · 2026-11-06