Prompt Adherence Feels Worse Than Base LTX 2.3
No offense to the author — this is just based on my own testing. Compared to the base LTX 2.3 model, this fine-tune seems to follow prompts less reliably. I’m seeing more drift from instructions, weaker scene/action consistency, and outputs that don’t match the requested details as closely. Curious if others are seeing similar results or if there are settings/workflows that improve adherence. I am using 10Eros_10SNodes_LikenessGuideHelper_I2V_v3.2 workflow and the prompt enhancement using grok.
Not sure what kind of outputs you're doing but sounds like something incredibly niche or not being done correctly. Sounds more complex than what that workflow is for, use an LTX director workflow for something ambitious all that workflow is for is 5-15 second one shot animations of an image.
same for me. Things which worked on V1, dosnt works on 1.2 anymore.
Also my gore loras dont doing anything anymore. Maybe its better for fuckery stuff, but not for action scenes with blood etc.
I am using the FP8 Version.
If loras don't work at normal low strength they may all need higher strength in 1.2, even over 1.0, up to 1.2-1.5 strength. v5 workflow also enables the whole point of 1.2 which is using the memory tensors from the joyAI model. It's all meant for strict conditioning and unique input consistency. If you're just trying to craft outputs which have evolution you may stick to v1.