Degrading latents don't matter:分段 fresh latents + FL2V 接缝修复跑通长视频
roychodraws · reddit · 2026-10-03
Reddit user roychodraws shares an open-source long-video generation workflow (degraderepo) that sidesteps latent degradation in video diffusion models. By exploiting a static camera and a mostly stationary subject, the workflow generates fresh latents every 10-15 seconds sequentially, reducing the latent decay problem to a simple seam problem. A second workflow then uses FL2V to resample a small section of video across each seam to hide it. Minor artifacts (e.g. a chair) remain fixable with reference images.
More from Multimodal
- Runway MCP one-shots a corporate training video in 15 minutes, typically $10k to make — tlakomy · 2026-10-03
- Hands-On: Qwen's New Image Model Delivers a Midjourney-Like Look — Puzzled-Valuable-985 · 2026-10-03
- AI-Generated Halo x Elden Ring x Overwatch Mashup Wows Gamers — adonis_singh · 2026-10-03
- CUTROOM: an open-source local video editor with AI-assisted rough cuts — Competitive_Might709 · 2026-10-03
- A meta-prompt that turns any LLM into a YuE2 music prompt builder — Fabulous_Broad_115 · 2026-10-03
- Full prompt shared for a 15-second one-shot magical winter AI video — umesh_ai · 2026-10-03