card: document img2img, with a per-checkpoint comparison at strength 0.75

README.md

@@ -3,6 +3,7 @@ license: unknown
 pipeline_tag: text-to-image
 tags:
   - text-to-image
+  - image-to-image
   - stable-diffusion-xl
   - sdxl
   - checkpoint
@@ -63,6 +64,34 @@ Values this build is run with in practice. They are read out of a live config wh
 
 ---
 
+## img2img
+
+Both checkpoints do `img2img` as well as `txt2img`. The same source image, the same prompt, the same seed — run through each checkpoint at the production `strength` of **0.75**.
+
+<table>
+<tr>
+<td align="center"><img src="https://huggingface.co/ChrisColeTech/sdxl/resolve/main/samples/img2img-source.png" width="250" alt="source image"><br><sub><b>source</b></sub></td>
+<td align="center"><img src="https://huggingface.co/ChrisColeTech/sdxl/resolve/main/samples/img2img-juggernaut.png" width="250" alt="Juggernaut XL v9 img2img at strength 0.75"><br><sub><b>Juggernaut XL v9</b></sub></td>
+<td align="center"><img src="https://huggingface.co/ChrisColeTech/sdxl/resolve/main/samples/img2img-illustrious.png" width="250" alt="Illustrious Anime v4 img2img at strength 0.75"><br><sub><b>Illustrious Anime v4</b></sub></td>
+</tr>
+</table>
+
+**Source** — `a red apple on a rustic wooden table beside a window, soft daylight, visible wood grain, studio photograph, sharp focus`. Generated separately, 1024×1024, seed 7.
+
+**img2img prompt** — `the same scene rendered as a vivid oil painting, thick visible brushstrokes, rich saturated colour, painterly texture`, at 1024×1024, **30 steps**, **strength 0.75**, seed 7. Roughly 170 s per image on an A100 80GB.
+
+### What the two checkpoints do differently
+
+This pair is the reason both are in this repo, and it is worth reading before you pick one.
+
+- **Juggernaut XL v9 keeps the subject.** The apple stays the apple — same object, same central composition, same scale — and the paint treatment is applied over it. The setting drifts (the window and table became a shoreline), which is expected: at `strength 0.75` only a quarter of the source survives the denoise, and a photographic checkpoint spends that budget on the subject.
+- **Illustrious Anime v4 does not.** At the same strength the apple becomes a different object entirely — a red lidded jar. That is not a fault to fix with settings; it is an anime checkpoint being handed a photographic source and a painting prompt, which is off-domain in both directions, so it re-imagines rather than restyles.
+
+**Practical rule:** for img2img over photographs, use Juggernaut. Use Illustrious for img2img only when the source is already illustration, or lower `strength` to ~0.3–0.4 so more of the source survives to anchor it.
+
+**On `strength`:** it is the fraction of the image the sampler is free to rewrite, so it trades style against fidelity directly. Below ~0.3 you get little more than a gloss pass; above ~0.8 the source is a loose suggestion and the composition may not survive. 0.75 is the production value and is chosen for restyling, not for subtle edits.
+---
+
 ## Layout
 
 Each `.safetensors` under `combined/` is a complete SDXL checkpoint — unet, both text encoders and the VAE in one file — so any standard SDXL loader can take it directly. The shared CLIP-L and OpenCLIP-G tokenizers live under `config/tokenizer/` and `config/tokenizer_2/`. There is no per-component split here.