SyFeee/LTX-2.3-SyFe-Plain-AV-LoRA

SyFe LTX-2.3 Plain and Audio-Video LoRAs

Self-trained SyFe LoRA checkpoints for Chinese-drama generation, bilingual prompting, and joint audio-video experiments on LTX-2.3 22B-dev.

Checkpoints

Run	Purpose	Rank	Steps	Status
tv306954_run01	First multi-character show baseline	64	3,000	Kept baseline; final loss 0.2897
v5_unified	Rich 747-clip Chinese-drama corpus	32	4,000	Shipped plain LoRA
official_704_cuval	Native 1280x704 official-trainer control run	32	3,000	Experimental control
SyFe_Bilingual_Plain_704	Bilingual 704p production stack component	32	10,000	Deployed checkpoint
av_lora_v1_production	Joint audio-video experiment	64	4,000	Archived
av_lora_v2_production	Joint audio-video P10.5 experiment	64	4,000	Kept with caveats

The AV checkpoints can generate ambient music, breathing, and sound effects, but they did not produce reliable intelligible dialogue. They should not be described as standalone voice-cloning models; use the ID-LoRA release for reference-voice conditioning.

Run folders include final weights and available training configs. v5_unified and the bilingual deployment checkpoint are release copies whose original preprocessing data is not included.

Use is subject to the LTX-2 community license.