DWARP open-sourced: Stable Diffusion video restyling with optical flow and per-pixel depth control

PurzBeats · x · 2026-10-04

spiritform released DWARP, an open-source tool that repaints videos frame by frame into painted, drawn, or dreamed styles on your local GPU.

Highlights:

• Uses Stable Diffusion per frame, but instead of starting from scratch it carries the previous painted frame forward along the video's optical flow, so brushstrokes travel with objects and results don't flicker.
• Its DepthDiff feature uses a luma/depth map to set per-pixel denoise strength: keep a face photographic while dissolving the room into paint, or paint only the shadows.
• One-screen workflow: drop a clip and render, with presets and local deployment supported.

Original post →

More from Multimodal

• Suspected Nano Banana 2.5 spotted on AI Arena under codename resplendent_flash — mornaji · 2026-10-04
• Grok Imagine quietly adds intermediate keyframes for video generation — mark_k · 2026-10-04
• W.I.T.C.H. fan trailer made 95% locally: full AI video pipeline on a 4070 + 5060ti — No_Taste_4102 · 2026-10-04
• Balancing inpainting cadence in video: more frequent inpaints mean less smear but more hectic detail — matdryhurst · 2026-10-04
• User claims Suno has achieved 'musical AGI', composing better than 95% of people — astralmatrix · 2026-10-04
• "Bad Footage" Makes AI Video More Believable: A UFO Vlog Demo with Full Prompt — LudovicCreator · 2026-10-04

Multimodal channel →