From 32791ee83234dab9774ed09b431b818f49234405 Mon Sep 17 00:00:00 2001 From: Victor Kuznetsov Date: Sun, 27 Sep 2026 15:49:57 -0700 Subject: [PATCH] Record the FLUX.2 klein Google brackets from four Gemini accounts Three of four Google carriers clear with klein at or below 0.20, where it matches qwen-zimage's PSNR at 0.35, beats its LPIPS, and keeps the objects qwen-zimage redraws. The fourth is detected at 0.10 and inconclusive at 0.20, so there is no Google operating point yet. Co-Authored-By: Claude Opus 5.5 --- docs/chroma1-engine-research.md | 19 +++++++++++++------ docs/research-sweep-2026-09.md | 5 +++-- 2 files changed, 16 insertions(+), 8 deletions(-) diff --git a/docs/chroma1-engine-research.md b/docs/chroma1-engine-research.md index 14eb21d..44fe39e 100644 --- a/docs/chroma1-engine-research.md +++ b/docs/chroma1-engine-research.md @@ -673,10 +673,17 @@ the two limiting carriers, against `qwen-zimage` at 22.4-24.0 dB and OpenAI klein is the worse engine. It is fast, under one second per image on a warm H100. -Google is unmeasured. At 0.10 klein kept far more of the Google carriers -(LPIPS 0.04-0.07 against `qwen-zimage` 0.11-0.19 at 0.35, which also redraws -objects), but the Gemini Verify AI tool returned an empty answer on the -first upload and "a tool error or quota limit" on the second, so no Google -verdict exists yet. Outputs, reports and oracle records: -`.local-eval/flux2-klein-2026-09-27/`. +Google, through Gemini on four accounts (the first returned "a tool error +or quota limit" and later "the usage quota for checks on similar images has +been exceeded", so the rest went to three others, each result recorded +against its own slot): one carrier `(0.05, 0.10]`, one clean at 0.10 with +0.05 unchecked, one `(0.10, 0.20]` ("Part of this image was edited or +generated with Google AI" at 0.10), and one `detected` at 0.10 and +"inconclusive" at 0.20, so its floor is above 0.20 or unsettled. At 0.20 +klein matched `qwen-zimage` at 0.35 on PSNR (17.0-20.3 against 17.6-21.0 dB) +and beat it on LPIPS (0.079-0.152 against 0.111-0.191), and it kept the +objects `qwen-zimage` redraws (the meatball bowl, the sign lettering). That is +the one place klein could win, but no Google operating point exists until the +fourth carrier settles above 0.20 and a margin is added. Outputs, reports and +oracle records: `.local-eval/flux2-klein-2026-09-27/`. diff --git a/docs/research-sweep-2026-09.md b/docs/research-sweep-2026-09.md index bb1bd11..a4bc143 100644 --- a/docs/research-sweep-2026-09.md +++ b/docs/research-sweep-2026-09.md @@ -61,8 +61,9 @@ This page is an index. Each finding lives in its canonical page. ([removal research](synthid-removal-research.md#color-restoration-after-regeneration-2026-09-27)). FLUX.2 klein 4B was measured the same day: stock inpainting copies the source through its reference tokens and keeps OpenAI SynthID, and without - them it clears OpenAI only at lower fidelity than `qwen-zimage`; its Google - verdict waits on the Gemini tool + them it clears OpenAI only at lower fidelity than `qwen-zimage`. On Google it + keeps objects `qwen-zimage` redraws, but one of four carriers is not yet + clean at 0.20 ([Chroma1 research](chroma1-engine-research.md#flux2-klein-as-a-global-stage-2026-09-27)). Still open: Z-Image-Turbo as global stages (with an Apple Silicon arm through mflux),