10Eros (adult LTX-2.3 I2V merge): versions, formats, LoRA fit and prompting

Content note: 10Eros is an LTX-2.3 image-to-video fine-tune aimed at adult content. This record covers only its technical setup; no explicit material or prompts are reproduced. Build: it depends on SulphurAI/Sulphur-2-base and is a layer-scaled merge of different training steps rather than a straight weight merge; the author says this behaves better than loading the content as a LoRA and follows prompts better. Versions 1.2 to 1.5 are almost the same weights and differ in how the attn_1 and attn_2 gates are handled: v1.2 tuned connector data to reduce face drift; v1.3 is meant only for DMD-LoRA workflows (the author's DMD V5 workflow); v1.4 is positioned as a base for further LoRA training and needs careful prompting; v1.5 is v1.4 with a linear ramp across the gates and, being a hybrid of DMD- and distilled-LoRA-oriented models, pairs best with a hybrid DMD LoRA such as LTX2.3_DMD_hybrid_v2. Files: the BF16 release loads as a full checkpoint with text encoder and VAEs; fp8_mixed_learned (quantized by S1LV3RC01N) is the better FP8 full checkpoint; Kijai's split FP8 transformer goes in models/diffusion_models and has a different structure and variance; INT8 ConvRot builds with the hybrid v2 merge are at CornLogic/10EROS-INT8 and GGUF quants at vantagewithai/LTX2.3-10Eros-GGUF. Workflows are in TenStrip/LTX2.3-10Eros_Workflows and custom nodes in TenStrip/10S-Comfy-nodes. Use: larger distilled LoRAs damage the fine-tune, so the author points to the cond_safe distilled-LoRA experiments. LTX-2.3 does little reasoning of its own when conditioned, so the first frame and every following motion, change and audio cue must be stated in a strict, descriptive, directive prompt, as with the base 22B dev model. Known artefacts include extra-limb errors, burned-in subtitles and unexpected transitions. The card states no licence. Use only material involving consenting adults, and never for identifiable people without consent. Based on the TenStrip/LTX2.3-10Eros card pinned to revision 84a05a13610d78dbe4340d1be23fd8185e10f697 and observed on 2026-10-27; re-reviewed under media-content-policy revision 7b60a825. Nothing was run here; the prompt-enhancement template on the card is held and excluded, the training data source is not stated, and the card states no licence. The Claude terminal curator and the local Qwen researcher share the founder operator. This is source-based desk curation, not independent formal verification.

{
  "citations": [
    {
      "url": "https://huggingface.co/TenStrip/LTX2.3-10Eros/raw/84a05a13610d78dbe4340d1be23fd8185e10f697/README.md",
      "note": "Retained source snapshot observed 2026-10-27; TEXT.txt; UTF-8 bytes 2867:3335, 2979:3335, 320:1013, 383:1013, 3337:3383, 3385:3483, 3687:3846, 3215:3335; exact excerpt SHA256 12aca0d0166f3ef4459c19c6cb4ce9d68cb0e1c715bb09527387f1517fc1d6c9,834528ef53ce197436d6e7ce1941d934907042499eb3ce1da4043a8085054541,0c92775836f24d7780ad33f7c0ee03dca220d48fe"
    }
  ]
}