Menu

TailSteer reduces drift in multi‑turn image editing; History Forcing stabilizes long‑horizon video generation
📰

TailSteer reduces drift in multi‑turn image editing; History Forcing stabilizes long‑horizon video generation

DEV Community: machinelearning·Papers Mache·6 days ago
#jN2LmRbe
Reading 0:00
15s threshold

Multi‑turn image editing and long‑horizon video rollouts still suffer from drift that compounds after just a handful of steps, making iterative creative tools brittle. TailSteer preserves the early denoising dynamics while History Forcing injects historic context, effectively halting that drift.

Earlier solutions either retrain on multi‑turn data—rewriting the pretrained priors—or rely on inference hacks that do not scale to large diffusion backbones. Those workarounds trade off fidelity for stability and often require hand‑crafted rollout pipelines.

TailSteer cuts cumulative drift across 15‑20 successive edits. It learns tail‑focused corrections from on‑policy rollouts, leaving the early denoising trajectory untouched; experiments across the FLUX, SenseNova, and Qwen families show a substantial reduction while largely preserving base‑model performance [1] .

TailSteer maintains the generative and editing abilities of the original model.…

Continue reading — create a free account

Join HashtagPLUS to read full articles, follow hashtags, vote, and join the conversation.