I’m sorry, but LTX still isn’t a professionally viable filmmaking tool

A detailed critique from a professional filmmaker has sparked debate about the real-world viability of AI video tools, specifically targeting the open-source model LTX 2.3. While acknowledging the team's effort, the author argues a significant hype-reality gap exists. The core issue is that text-to-video generation, LTX's flagship feature, is fundamentally misaligned with professional needs. Pros require precise control, consistency, and verifiable outputs, not random, non-deterministic clips. The critique states bluntly that text-to-video will "never be that useful for professional workflows" because businesses cannot rely on unpredictable results or verify the provenance of generated content, risking copyright issues.

The post details specific, deal-breaking technical flaws. LTX suffers from severe overfitting, where the model regurgitates elements from its training data, like spontaneously adding 'Big Think,' 'Washington Post,' or 'TED' logos to generated videos. More critically, the control mechanisms essential for professional adoption are described as "broken." Image-to-video is inconsistent, LORA (Low-Rank Adaptation) training for custom styles is unreliable, and features for character consistency or pose tracking are lacking. Without functional fine-grained control via img2vid, LORAs, or video-to-video, the tool cannot integrate into serious filmmaking pipelines where artists need to direct outputs, not just hope for a usable result from dozens of generations.