Trigger: <code>jack-o pose</code>Full retrain of the jack-o pose LoRA, focused on the two weaknesses reported against v1: pose/style entanglement at high weight, and poor obedience to camera-angle instructions (especially anything other than a straight front shot).What changed under the hood<ul><li>Trigger word now properly wired into the trainer, not just present in captions. This changes what the model learns from the 5% caption-dropout steps — instead of associating "no text" with the dataset's average look, those steps now reinforce the trigger phrase alone.</li><li><code>content_or_style: content</code> — timestep sampling now favours the steps that encode structure and composition over the ones that encode surface texture. This is the main lever behind the reduced style bleed at high weight.</li><li>Horizontal flip augmentation enabled — the pose is left/right symmetric, so this doubled effective coverage for free.</li><li>Dataset grew from ~65 to 81 images, with additions targeted specifically at the angle/subject combinations that were thin or missing: true overhead shots, elevated three-quarter views, rear and rear-three-quarter angles at eye level and elevated height, and more non-female subjects (men, robots, androids, dolls, figurines).</li><li>Every caption rewritten to one consistent template — natural sentence, trigger phrase grammatically embedded, camera height and angle from a closed vocabulary, no pose anatomy in the text (that's the trigger's job now), no quality/style adjectives.</li><li>Full caption-accuracy audit before training — went back through the dataset image-by-image and corrected several front/rear mislabels, plus a handful of captions that had drifted onto the wrong image entirely after files got swapped during earlier dataset iterations.</li></ul>What to expect differently from v1<ul><li>Style holds up better as you push weight toward 1.0 — medium (photo, anime, painting, product-shot) stays intact instead of collapsing toward the dataset's dominant studio-photo look.</li><li>Non-front camera angles are meaningfully more obedient, especially rear and rear-three-quarter at eye level and elevated height. Straight overhead is improved but still the weakest angle — it has the least training coverage of any camera position.</li><li>Non-human and male subjects hold the pose more reliably instead of drifting toward the dataset's dominant female anatomy.</li></ul>Recommended weightStill 0.6–0.8 for most prompts. If v1 needed 0.9–1.0 to lock the pose in on a hard prompt, try v2 at 0.7–0.8 first — the trigger binds harder now, so the same weight goes further.Changelog vs v1Training config<ul><li><code>trigger_word</code>: unset → <code>jack-o pose</code></li><li><code>content_or_style</code>: <code>balanced</code> → <code>content</code></li><li><code>flip_x</code>: <code>false</code> → <code>true</code></li></ul>Dataset<ul><li>65 → 81 images</li><li>Added: true overhead shots (3→8), elevated-height shots at side/rear azimuths (previously near-zero), additional male and non-human subjects (2 → 10+)</li><li>Camera-angle distribution rebalanced — front-heavy skew reduced from ~70% toward ~47% of the set</li></ul>Captions<ul><li>Rewritten wholesale to a single consistent template (subject → trigger phrase → setting → camera height/angle → medium)</li><li>Removed all pose-anatomy descriptions from caption text (previously present in ~35% of captions, competing with the trigger for the pose concept)</li><li>Removed quality/style adjectives (<code>polished</code>, <code>cinematic</code>, <code>dramatic</code>, etc.) throughout</li><li>Corrected 6 front/rear/side mislabels found on manual re-verification</li><li>Corrected 5 captions that had become mismatched to the wrong image after dataset file swaps</li></ul>Known limitations carried over<ul><li>Straight overhead remains the least-covered camera angle — expect more retries there than at any other position</li><li>Very high weight (0.9–1.0) on prompts stacking an unusual medium, unusual angle, and non-human subject simultaneously can still show minor style pull</li></ul>Example prompt(and you can see their in the WF, attached to this post):<pre><code>A woman with a short haircut, wearing a Stellar Blade cosplay costume, in a jack-o pose on the metal floor of a dimly lit underground command compartment; She turns her head toward the camera, smiling invitingly. The shot from behind is taken directly from the top-rear, overhead, full-length, a color photograph with cool blue-white lighting from above.</code></pre><pre><code>image in cartoon style. A woman with a short haircut, wearing a Stellar Blade cosplay costume, in a jack-o pose on the metal floor of a dimly lit underground command compartment; She turns her head toward the camera, smiling invitingly. The shot is taken directly from the top three-auqrter angle, overhead, full-length, a color photograph with cool blue-white lighting from above.</code></pre>

Files with SHA256: c1c622a1daf2c040f5c7782698fc486eb3bb4a9dfbcc27320efd5bfd39be70ee