Local LLM plus video models turn a novel into a trailer: 48-scene pipeline needs help

OpenEvidence9680 · reddit · 2026-09-12

The author already runs a pipeline in their CMS for generating custom book covers, and now wants to go further: a short trailer plus a long-form film for Guards! Guards! by Pratchett, a favorite book never adapted to film.

Current setup: a local LLM reads the whole book and produces a character list with descriptions, which Krea2 turns into character sheets (not ideal, but acceptable for a test). The same LLM writes a screenplay in three acts totaling 48 scenes, each with an action description and a Krea-generated still; the pipeline then generates clips to be joined. The author has made only a handful of 5-second videos before and used very basic txt2img for the covers.

Minimax is giving trouble (likely bad prompting), so they switched to ltx2.5, which does a decent job, but they started this morning and have only one scene done; a dragon head on a dragon's wing is not the worst issue. Hardware is two 5060 Ti 16GB cards plus 64GB RAM until a new PC arrives.

They are asking for advice on which models to use, available speedups, which models handle fantasy better, LoRA suggestions, whether LTX 2.5 supports first and last frames like 2.3, and any workflows that could streamline this.

Related event: Maker Uses Local LLM and Video Models to Adapt Novel into 48-Scene Film(2 posts)→

Original post →

More from Multimodal

Multimodal channel →