card: fixture prompt matched to the source; state plainly why it is a weak demo

README.md

@@ -93,25 +93,38 @@ settings are forgiving. Below ~0.3 you get little more than a gloss pass; above
 ~0.85 the source becomes a loose suggestion.
 
 
-### The repo's own test fixture
 
-`tests/sdxl/test_06_source.png` with the test plan's `DEFAULT_PROMPT`, through both checkpoints.
+### The repo's own test fixture — and why it is a poor demo
+
+`tests/sdxl/test_06_source.png` through both checkpoints, with a prompt describing
+what is actually in the source.
 
 <table>
-<tr><td><img src="https://huggingface.co/ChrisColeTech/sdxl/resolve/main/samples/fixture-source.png" width="250" alt="test_06 fixture source"></td><td><img src="https://huggingface.co/ChrisColeTech/sdxl/resolve/main/samples/fixture-juggernaut.png" width="250" alt="Juggernaut on the fixture source"></td><td><img src="https://huggingface.co/ChrisColeTech/sdxl/resolve/main/samples/fixture-illustrious.png" width="250" alt="Illustrious on the fixture source"></td></tr>
-<tr><td><sub>**source** — `tests/sdxl/test_06_source.png`, a synthetic gradient</sub></td><td><sub>**img2img prompt:** `a colorful vibrant landscape painting, mountains lake sunset clouds, thick oil paint brushstrokes, high detail, 4k ultra detailed` — **Juggernaut XL v9**, 1024×1024, **25 steps**, **guidance 7.5**, **strength 0.75**, seed 42</sub></td><td><sub>**img2img prompt:** `a colorful vibrant landscape painting, mountains lake sunset clouds, thick oil paint brushstrokes, high detail, 4k ultra detailed` — **Illustrious Anime v4**, 1024×1024, **25 steps**, **guidance 7.5**, **strength 0.75**, seed 42</sub></td></tr>
+<tr><td><img src="https://huggingface.co/ChrisColeTech/sdxl/resolve/main/samples/fixture-source.png" width="250" alt="test_06 fixture source, a synthetic gradient with a flat circle"></td><td><img src="https://huggingface.co/ChrisColeTech/sdxl/resolve/main/samples/fixture-juggernaut.png" width="250" alt="Juggernaut on the fixture source"></td><td><img src="https://huggingface.co/ChrisColeTech/sdxl/resolve/main/samples/fixture-illustrious.png" width="250" alt="Illustrious on the fixture source"></td></tr>
+<tr><td><sub>**source** — `tests/sdxl/test_06_source.png`, a synthetic gradient with a flat circle</sub></td><td><sub>**img2img prompt:** `an oil painting of an abstract composition with a large glowing golden circle at the centre over a flowing rainbow gradient, thick visible brushstrokes, impasto ridges catching the light, visible canvas texture` — **Juggernaut XL v9**, 1024×1024, **25 steps**, **guidance 7.5**, **strength 0.75**, seed 42 — about 174 s on an A100 80GB</sub></td><td><sub>**img2img prompt:** `an oil painting of an abstract composition with a large glowing golden circle at the centre over a flowing rainbow gradient, thick visible brushstrokes, impasto ridges catching the light, visible canvas texture` — **Illustrious Anime v4**, 1024×1024, **25 steps**, **guidance 7.5**, **strength 0.75**, seed 42 — about 174 s on an A100 80GB</sub></td></tr>
 </table>
 
-**A caveat about this fixture, not the model.** The source is a synthetic gradient
-with a flat circle. The result gains real vertical brushwork and holds luminance
-(1.00× and 1.03× of source), but the prompt's *content* — mountains, lake, sunset —
-never appears, because at `strength 0.75` the source's geometry dominates and there
-is no semantic content to reinterpret.
-
-That also makes this fixture unable to detect the failure that matters: with no
-subject in the source, subject loss is invisible. It is why `strength 0.90` passes
-here and discards the subject on a photograph. **Validate img2img against a real
-photograph as well.**
+**Neither result is a good advertisement, and the source is why.** Juggernaut gives
+the gradient real vertical brushwork; Illustrious barely changes it at all. In both,
+the circle stays **perfectly flat** — no impasto — even though the prompt explicitly
+asks for ridges catching the light on it.
+
+That is not a model or a settings problem. A large region of uniform colour gives
+the denoiser nothing to work with at `strength 0.75`: flat reconstructs to flat.
+The gradient varies, so it takes texture. Use the apple comparison above to judge
+these checkpoints; this fixture is included because the test suite uses it, not
+because it shows the model at its best.
+
+Two consequences worth carrying into any testing you do:
+
+- **A synthetic source cannot detect subject loss.** With no subject present, the
+  failure that matters is invisible — which is how `strength 0.90` passes here and
+  discards the subject entirely on a photograph.
+- **`histogram_correlation` is meaningless on this source.** Measured against it,
+  the metric comes out 0.001 and −0.017, because a near-uniform histogram cannot
+  correlate with anything. On a real photograph the same runs score 0.73–0.78. The
+  test plan reads a negative value as "colors unrelated — check VAE scaling", so
+  run against its own fixture the suite would send you hunting a bug that is not there.
 
 ## Recommended settings