Dual-model image ref workflow: GPT 2.5 Sunburst for characters, Seedream 5.0 Pro for scenes
techhalla · x · 2026-11-10
Step one of techhalla's AI video pipeline: generating the right image references.
- Characters & lineups: GPT 2.5 Sunburst, praised for character accuracy; the attached images show medieval-fantasy mugshot-style height lineups with strong consistency.
- Scenes: Seedream 5.0 Pro for nailing the tavern's look and lighting feel.
- Full prompts included at the end of the thread.
The dual-model split — one for character consistency, one for scene atmosphere — is a reusable technique.
Related event: Creator shares retro-cinematic AI workflow with Seedance and film tokens(3 posts)→
More from Multimodal
- Midjourney sref 3473377010 delivers charming crayon-textured illustration styles — sergeantsref · 2026-11-10
- Kaigen Studio opening to the public tomorrow, browser-based access confirmed — gdechichi · 2026-11-10
- Early access hands-on: Hailuo AI's H3 Image looks quite impressive — gen_ericai · 2026-11-10
- Redditor Creates a Full Friends Episode With MiniMax H3, Ref2VA and DaVinci Resolve — Thorozar · 2026-11-10
- How Does the Elite Trance Music Channel Keep AI-Generated Characters Consistent Across Cuts? — Marino4K · 2026-11-10
- Original AI character "Mizuki" flight demo shows zero-drift character consistency — EitherCartographer63 · 2026-11-10