Seedance 2.5 still misspells text in videos despite prompt emphasis
AthleteArtistic3121 · reddit · 2026-08-25
A long-time user of video generation models reports that even the strongest model, Seedance 2.5, still has weak prompt understanding. The most ridiculous basic failure: the model misspells words appearing inside the video, and re-emphasizing the spelling in the prompt didn't fix it. The author jokes that LLMs may have hit a wall, but video generation models clearly haven't reached theirs yet.
More from Multimodal
- The math: 5 rerolls at 480p plus a 4K upscale costs $5 vs $9 at 720p — HeyAmit_ · 2026-10-15
- Dreamina workflow: generate at 480p, upscale to 4K, cut video costs 44% — HeyAmit_ · 2026-10-15
- One-prompt GPT-6 Astra writes two Palestrina-style motets in LilyPond — deanwball · 2026-10-15
- DiffSynth-Studio Adds OPD Workflow: Few-Step Diffusion Training in ~2 GPU-Hours — bdsqlsz · 2026-10-15
- AI film and TV hit the market: Prime Video picks up AI spin-off, pilots screen at Mipcom — lmoroney · 2026-10-15
- Ramayana film taps DNEG's massive VFX and AI-assisted production for global release — CurieuxExplorer · 2026-10-15