{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreigyyffsny2rgwz77k56rmqgfzypn2ao2qthtqotdgadk4bwfycovu",
    "uri": "at://did:plc:pgryn3ephfd2xgft23qokfzt/app.bsky.feed.post/3mkx7uevi6w72"
  },
  "path": "/t/multi-image-edit-3-refs-artifacts-at-true-cfg-fine-on-lightning-reference-content-dependent/175726#post_1",
  "publishedAt": "2026-05-03T11:11:26.000Z",
  "site": "https://discuss.huggingface.co",
  "tags": [
    "https://codeshare.io/5eLXqK"
  ],
  "textContent": "Setup\n\nQwen-Image-Edit-2511 BF16 via ComfyUI\nTextEncodeQwenImageEditPlus with 3 reference images (face close-up + body front + body back)\nOutput 1024×1536 (non-square 2:3)\nSampler: res_3m + bong_tangent (RES4LYF)\n\nBehavior\nWith full CFG (2.7, 33 steps): generation reliably breaks on some reference sets — mixed artifacts (identity drift, color/texture corruption, anatomy distortion). Same parameters and seeds produce clean output on other reference sets.\nWith Lightning 4-step (true_cfg=1): every reference set is clean.\nPattern\n\n1 reference (face only) → always clean, both modes\n3 references → clean on some characters, broken on others — content-dependent\nAll references are Z-Image Turbo outputs, same prompt structure, identical dimensions\nFailing sets tend to contain high-frequency content (curly hair, darker skin texture); working sets tend to be lower-frequency (straight hair, lighter skin). To be clear: this is about the rendered references, not the character identity itself.\n\nWhat I’ve tried (no fix, or partial only)\n\nI don’t even remember what I tried, but I tried a lot of things that seemed possible, and none of them worked. The workflow is below.\n\nhttps://codeshare.io/5eLXqK\n\nQuestion\nIs this a known interaction between multi-ref token packing and the true-CFG noise_pred * (cond_norm / noise_norm) rescale path? Specifically:\n\nDoes Qwen2.5-VL’s 384² downscale produce per-token norm outliers on high-frequency reference content that get amplified across denoising steps once true CFG is active?\nIs multi-image reference (3+ refs) currently only stable at distilled-CFG / Lightning, or is there a recommended setup for full CFG?",
  "title": "Multi-image edit (3 refs): artifacts at true CFG, fine on Lightning — reference-content dependent"
}