Winnougan/10Eros-INT8-Convrot

10 Eros v1.2 — LTX 2.3 Explicit Video Finetune

10 Eros is an explicit video generation finetune built on LTX 2.3, merged with Sulphur-2-base weights using layer-scaled merges across different training steps. It is not a straight weight merge — this approach produces significantly better prompt adherence and more natural motion than loading a LoRA on top of a base model.

Designed primarily for Image-to-Video (I2V) use. Loads as a full checkpoint with CLIP and VAEs in BF16.

⚠️ This model generates explicit adult content. You must be 18+ to use it. Intended for legal adult content creation only.

---

What's New in v1.2

• Leveraged tuned connector data to reduce face drift during long generations
• Improved handling of long prompts and director-style prompt structures
• Applied Sulphur EXP weights on top of v1 to sharpen explicit motion fidelity
• Known issues (carried over from v1): mistaken extra anatomy, subtitles appearing in frame, unexpected transitions — these are LTX 2.3 base model limitations and are still present

---

Model Files

Format	Location
BF16 full checkpoint (CLIP + VAE included)	This repo
FP8 Transformer split files (diffusion_models)	Kijai/LTX2.3_comfy
INT8 ConvRot quantized	[Winnougan HuggingFace — link TBD]

Load the BF16 as a checkpoint. The Kijai FP8 split files go inside diffusion_models.

---

Required LoRA:

• Grab the official 10Eros lora here. If you don't load it, the resulting video will be fuzzy: https://huggingface.co/TenStrip/LTX2.3_Distilled_Lora_1.1_Experiments/tree/main

Workflows

Workflows available here: 👉 TenStrip/LTX2.3-10Eros_Workflows

---

Prompting Guide

LTX 2.3 has very little autonomous reasoning when conditioned on a first frame. You must explicitly describe everything — initial scene, all motion, body part movements, composition changes, audio, and dialogue. If you don't ask for it, you won't get it.

Recommended LoRAs

Use the cond_safe distilled LoRAs to aid generation. Larger distilled LoRAs will interfere with the finetune:

👉 LTX2.3 Distilled LoRA 1.1 Experiments

Prompt Enhancement

For best results, use an uncensored LLM (Grok or similar) with this foreword to generate your prompt from a reference image:

Click to expand prompt enhancement template