Fix Krea2 atlas cue prompt formatting

This commit is contained in:
2026-06-30 22:01:03 +02:00
parent ff484aa27c
commit 337bbb10f1
7 changed files with 366 additions and 10 deletions
+1 -1
View File
@@ -326,7 +326,7 @@ scene. The corrected test pattern keeps the coworking location intact:
editing the generator, wait for `sxcp_eval_in` to advance to the new turn, and
compare each image against the atlas verticality criteria. The useful axis is
`nadir-angle` or `bird's-eye` plus standing male POV, nearby floor plane
dominating the image, one woman directly below between the viewer's feet, and
dominating the image, the woman directly below between the viewer's feet, and
top-down office anchors. Avoid `plumb-line` and `map` in generator prompts
because Krea2 can literalize them as drawn graphics.
- `2026-06-29`: For quick wording-axis search, prefer a batched prompt-probe
+2 -2
View File
@@ -1449,7 +1449,7 @@
"result": "accepted",
"decision": "provisional_generator_patch",
"baseline_prompt_summary": "Earlier top-view probes on the same seed used vertical-shaft or generic steep-overhead wording and could still read too frontal, too horizontal, or too dependent on long office depth cues.",
"candidate_prompt_summary": "Prompt-only axis loop tested near-vertical floor-plane, plumb-line/map, nadir-angle, and bird's-eye standing male POV wording. The accepted axis is nadir-angle or bird's-eye standing male POV, viewer looking almost straight down from torso to floor, nearby carpet/floor plane dominating, one woman kneeling directly below between the viewer's feet, top-down office anchors, and a short centered vertical shaft column.",
"candidate_prompt_summary": "Prompt-only axis loop tested near-vertical floor-plane, plumb-line/map, nadir-angle, and bird's-eye standing male POV wording. The accepted axis is nadir-angle or bird's-eye standing male POV, viewer looking almost straight down from torso to floor, nearby carpet/floor plane dominating, the woman kneeling directly below between the viewer's feet, top-down office anchors, and a short centered vertical shaft column.",
"observation": "Turns 67, 69, and 70 on sampler seed 4242424242 produced the first clearly atlas-like verticality after the user rejected horizontal probes: viewer abdomen/shorts/feet anchor the lower edge, the woman is directly below between the viewer's feet, floor plane and nearby office furniture read as top-down anchors, her hair crown/forehead/shoulders/hands/knees are visible from above, and mouth contact stays centered. Turn 68 showed that plumb-line and map are unsafe generator words because Krea2 literalized them into drawn graphics. Mirror the nadir-angle/floor-plane wording as a provisional generator patch, but keep the catalog route candidate until another source or seed repeats it.",
"baseline_image": "/media/unraid/comfyui/output/agent_bridge/img_e421319637bc45f0bb49bfc486c7f4ad.png",
"candidate_image": "/media/unraid/comfyui/output/agent_bridge/img_ea922fa7bd6642f5bbe76b50f48b558b.png",
@@ -1457,7 +1457,7 @@
{
"source_case": "turn67 near-vertical floor-plane",
"candidate_image": "/media/unraid/comfyui/output/agent_bridge/img_56d6738989794a569cfe45d79dd96e88.png",
"observation": "Strong vertical floor plane, one woman below the viewer, table and chair legs as top-down anchors, and coherent centered contact."
"observation": "Strong vertical floor plane, the woman below the viewer, table and chair legs as top-down anchors, and coherent centered contact."
},
{
"source_case": "turn68 plumb-line map",
+2 -2
View File
@@ -795,7 +795,7 @@ Works better:
- `nearby carpet/floor plane dominates the image`
- `viewer abdomen, shorts, thighs, and feet frame the lower foreground`
- `shaft appears as a short centered vertical column from the foreground`
- `one woman kneels directly below the viewer between his feet`
- `the woman kneels directly below the viewer between his feet`
- `hair crown, forehead, shoulders, hands, and knees are visible from above`
- `desk legs, chair wheels, carpet texture, and floor seams as top-down office anchors`
- `mouth seals around the centered shaft`
@@ -822,7 +822,7 @@ Correction probes: a later same-seed source-`46` probe showed that vertical
center-shaft wording and generic steep-overhead wording can still feel too
frontal or horizontal. A short axis loop found the stronger terms: `nadir-angle`
or `bird's-eye` paired with `standing male POV`, `floor plane dominates`, nearby
top-down office anchors, `one woman directly below between his feet`, and
top-down office anchors, `the woman directly below between his feet`, and
`short centered vertical column`. The `plumb-line` and `map` terms produced
good geometry but literal drawn artifacts, so they should stay out of generator
wording.