{
"$type": "site.standard.document",
"bskyPostRef": {
"cid": "bafyreihxurwumd7jefvys6bcsfmotk7zdhrjlskatwdtvc63kkcztoxawq",
"uri": "at://did:plc:pgryn3ephfd2xgft23qokfzt/app.bsky.feed.post/3ml5ajo6ijg52"
},
"path": "/t/multi-image-edit-3-refs-artifacts-at-true-cfg-fine-on-lightning-reference-content-dependent/175726#post_3",
"publishedAt": "2026-05-05T20:53:14.000Z",
"site": "https://discuss.huggingface.co",
"textContent": "To be honest, I’m going to put this on hold for now. Because it works well with all the characters I’m trying to create—even the anime character with the fox tail :D. It only had issues with one of them. If I remove the negative prompt and tweak the prompts a bit, it works out. But it’s still very fragile. I’ll set aside more time later to find a solution for this. But for now, I’m skipping it.\n\nThanks anyway for the detailed explanation. I’ve read some of it, and I’ll read the rest when I get back to this problem.\n\nBut actually, I don’t think we even need a reference image of the back (unless there’s a tattoo or something). Because I think the front view can produce an optimal result in rear shots as well, based on the waist-to-hip ratio.\n\nBut I’d like to add something. In my opinion, Qwen Image Edit 2511 really does produce results with consistent quality. Maybe I haven’t even managed to find the optimal settings yet, but it’s great. Since it generates scenes from scratch, it retains that AI-generated look. That’s why I’m trying out some LoRa’s that work well with text-to-image generation.",
"title": "Multi-image edit (3 refs): artifacts at true CFG, fine on Lightning — reference-content dependent"
}