Prompt adherence, built on LTX-2.5.
A new text encoder and prompt enhancer let LTX-2.5 follow complex, multi-element instructions from shorter, simpler prompts. Self-host the open weights to test it on your own production prompts.
.svg)
Key capabilities
Complex instructions, simpler prompts
Improve prompt adherence across multiple characters, specific actions, and camera direction using shorter prompts than earlier LTX versions required.
New text encoder and prompt enhancer
Simplify AI video prompt engineering with a custom prompt enhancer that expands brief descriptions before a Gemma 4 12B text encoder processes them.
Multiple characters, one prompt
Use simpler AI video prompt engineering to direct multiple characters and specific actions in one prompt without exhaustive manual detail.
Complex results from plain language
Combine camera movement, action, and style in a short prompt. The best AI video prompts can be clear and direct without being engineered mega-prompts.
Fewer retries
AI video prompts with nuanced instructions are followed correctly on the first pass more often, reducing rewrites and retries.
Test it on your own prompts
Self-host LTX-2.5 and test your most complex prompts — multiple characters, specific camera moves, and layered actions — instead of judging adherence from an example.
How it works
Input
Technical characteristics:
- Prompt — A prompt with complex, multi-element instructions: multiple characters, specific actions, camera direction, and style.
- Prompt enhancer + text encoder — A custom prompt enhancer expands brief descriptions into richer cinematic direction before a Gemma 4 12B text encoder processes them for generation.
- Base — Built on the LTX-2.5-22B base model.
Output
Technical characteristics:
- Accurate generation — A generation that follows complex, multi-element instructions from a shorter, simpler prompt than earlier LTX versions required.
What users are building with LTX
FAQs
What is prompt adherence in AI video generation?
Prompt adherence measures how closely the output matches specific instructions in the prompt, including characters, actions, camera direction, and style, rather than producing a generic interpretation.
What changed in LTX-2.5?
A new Gemma 4 12B text encoder and a custom prompt enhancer. Together, they let the model follow complex creative instructions from shorter, simpler prompts.
Can I evaluate this myself?
Yes. LTX-2.5 has open weights and can be self-hosted, so you can test prompt-following directly against your own production prompts.
Is there a prompting guide?
Yes. This page covers the model's prompt-adherence capability. For practical guidance on how to write AI video prompts, see the separate LTX Prompting Guide.
