SCAIL 2 survives character swaps, object permanence, and physics tests better than expected
blackmixture · reddit · 2026-07-29
SCAIL 2 handles character swaps, object permanence, and physics better than expected
The author stress-tests SCAIL 2 across scenarios that usually break generative video demos: character swaps, complex actions, prop swaps, novel interactions, relighting, 2D motion transfer, and object permanence.
Key findings
- Character swaps are the strongest use case. Results improve a lot when the reference image is first edited into roughly the same pose/framing as the driving video.
- Object permanence held up surprisingly well. In one car clip, the vehicle leaves frame and returns without collapsing into mush.
- It can invent plausible motion. In a live-action Zuko example, fire is generated in a believable arc even though no fire appears in the source footage.
- Physics is a standout. A flower-to-wine-glass swap preserves hand tracking, liquid sloshing, and even refraction/distortion through transparent glass.
- Text is still weak. A speed sign in the background degrades into gibberish.
Workflow and access
- Everything was done in Mix Studio, a free, open-source local interface on top of ComfyUI.
- The author says the workflow is: edit an image, set it as the first frame, choose video mode SCAIL 2, and generate.
- Hardware used: Dell Pro Precision T2 with an NVIDIA RTX 6000 Pro, taking about 2–3 minutes per generation.
- A video tutorial and more examples are linked in the post.
More from Multimodal
- Seedance 2.5 workflow: character reference sheets + structured prompt yields a 30s single-take fake tourist video — techhalla · 2026-10-16
- AI couple moves into a castle ruin: full Magnific + Seedance 2.5 prompt thread behind the viral clip — techhalla · 2026-10-16
- Smoky Velvet Noir: a reusable image-gen prompt template with three moody examples — LudovicCreator · 2026-10-16
- VideoGen-Agent: An 8B Multimodal Agent Turns Video Generation Into a Tool-Use Loop — jiqizhixin · 2026-10-16
- Redditor builds impressive seaplane sim project with Sonnet 5.5 + Blender — Wise_Shame_2853 · 2026-10-16
- Redditor automates LoRA dataset captioning for 300 photos with Muse AI — Wide_Director_8897 · 2026-10-16