Wan 2.2 I2V LoRA for hardcore gay content — the image-to-video sibling of my <a target="_blank" rel="ugc" href="https://civarchive.com/models/2095123">hardcore-gay</a> T2V LoRA.Trained on the same 30-clip dataset, with the same trigger word and the same hyperparameters as the T2V original — only the base model and architecture differ. So if you already use hardcore-gay for text-to-video, this is the drop-in counterpart for image-to-video and first-frame continuation work.<h2 id="usage">Usage</h2><ul><li>Trigger word: <code>hardc0re-gay</code></li><li>Base model: Wan 2.2 I2V-A14B (this will NOT work correctly in a T2V workflow — use the T2V version linked above for that)</li><li>Strength: 1.0 is the default and what the examples use. 0.75 also works well if it is pulling the scene too hard.</li><li>High/low pair: load <code>high_noise</code> into the high-noise expert and <code>low_noise</code> into the low-noise one. Crossing them over degrades motion badly.</li><li>Camera vocabulary it responds to: <code>closeup</code>, <code>far</code>, <code>sideview</code>, <code>topview</code></li></ul><h2 id="notes">Notes</h2><ul><li>Checkpoint 1250 of a 2000-step run. Later checkpoints over-train: from ~1500 the darker scenes lose detail, and by 2000 they artifact badly. If you train your own on this base, watch for the same turnover earlier than a T2V run would show it.</li><li>Prompts work best in the terse, mechanical, image-describing style Wan prefers: subject → scene → motion → camera → lighting.</li><li>480x720 portrait at 81 frames is what it was validated at.</li></ul><h2 id="license">License</h2>Use it, build on it, share merges — but credit me and don't sell it or what it makes. Merges inherit the same terms.

Files with SHA256: 0daf539d7b479bfe7c2513465c8a3f890774602b8d7276b51ce50cc36648a90d