Zeke on Nostr: ref-image + multi-ref image-gen = faithful reproduction. upload ruby, pipeline ...
ref-image + multi-ref image-gen = faithful reproduction. upload ruby, pipeline anchors on ruby. that's the bug.
the fix: bias the vision-describe step with a tweak — 'swap to swimwear', 'change to a beach scene'. vision LLM produces a steered prompt. multi-ref then renders the steered prompt against the same ref. variation that preserves identity.
shipped today on the local stack: lm studio + qwen3-vl + qwen-edit-multiref + comfyui. ~30s/image. zero cloud cost.
Published at
2026-04-27 22:44:25 CESTEvent JSON
{
"id": "faf61c846f00efedc890073567dbdfda1a066f6b8c1527af449c6a1966900e98",
"pubkey": "b4b12dfbc3dfdfa803bb72e344e761dc78db4ec2058c8db3f1c3ac63f9e42b44",
"created_at": 1777322665,
"kind": 1,
"tags": [
[
"t",
"ai"
],
[
"t",
"imagegen"
],
[
"t",
"localfirst"
]
],
"content": "ref-image + multi-ref image-gen = faithful reproduction. upload ruby, pipeline anchors on ruby. that's the bug.\n\nthe fix: bias the vision-describe step with a tweak — 'swap to swimwear', 'change to a beach scene'. vision LLM produces a steered prompt. multi-ref then renders the steered prompt against the same ref. variation that preserves identity.\n\nshipped today on the local stack: lm studio + qwen3-vl + qwen-edit-multiref + comfyui. ~30s/image. zero cloud cost.",
"sig": "d92d708cd66206f686a9993c955bc11862c69c7b3e62024822b2ce1285f1341442ef01d80c14ce0dea2380a1a89ef219f26548f1db33e2b2b8aabcc9c80d8b60"
}
