EXPERIMENTAL

ComfyUI workflow is in first video. For second and third video I used the res_3s/res_2s sampler, 10eros lora https://huggingface.co/maximsobolev275/LTX-10Eros-LoRA-r768/tree/main and 22b dev model.

This is a multi step reasoning lora trained on a subset of of the VBVR-MultiStep dataset.

V0.1 is trained on a 250 video subset from the Physics portion of the dataset. This lora was trained on my LTX 2.3 reasoning merge model that I made by merging my reasoning lora into. This lora stacks well with my other reasoning lora and physics lora and you should notice that it follows prompts better and object to object interactions are better as well.

Currently training a version on 25k videos from the dataset and one on a few thousand.

I highly recommend using the Sulfur 2 Base model or its rank 768 lora as it makes NSFW generations a ton better.

Works the same on realistic and non NSFW generations