Capstone: Crafting a Prompt Iteration Workflow

a coffee shop interior

⚠ Problem: exactly the failure imgai1-3's own warn-box predicted — five words map to a huge, poorly-differentiated region of embedding space. The result is a plausible but generic coffee shop, with no warmth, no barista, no signage, and no connection to "Ember & Oak" at all.

a cozy artisan coffee shop interior, a barista standing behind a wooden counter, hand-lettered chalkboard menu board on the wall behind [subject]; warm rustic photorealistic style [style]; wide shot, counter centered, chalkboard visible in background [composition]; warm golden interior lighting, soft window light from the left [lighting]; 35mm lens, shallow depth of field [camera/lens]; photograph [medium]

⚠ Problem: much stronger overall, but the chalkboard menu text comes back as garbled, illegible lettering — exactly the mechanical limitation imgai1-2 and imgai1-8 explained (letters are learned as visual texture, not discrete symbols). No amount of rewording the subject line fixes this on its own.

negative prompt: blurry, extra limbs, deformed hands, watermark, oversaturated, cartoon, illustration
CFG scale: 8

⚠ Problem observed at this pass: the barista's hand near the espresso machine shows a fused-finger artifact — the anatomical failure mode imgai1-8 covered in depth.

✓ Per imgai1-5's own CFG formula, the negative prompt's own embedding replaces the unconditional baseline, actively steering away from "deformed hands" at every denoising step, in addition to the general quality terms — a first, partial mitigation.

a cozy artisan coffee shop interior, a barista standing behind a wooden counter, (hands resting on the counter, not visibly gripping anything:1.2), hand-lettered chalkboard menu board on the wall behind ...

✓ Per imgai1-7's own weighting mechanism, this doesn't add an anatomical rule the model never learned (imgai1-8's own warn-box on this exact point) — it does, however, steer the pose itself toward a simpler, less articulated hand position, which reduces how often the fuzziest, highest-variability poses (the ones most prone to fused-finger artifacts) get generated in the first place.

The base checkpoint's own default style leaned slightly more illustrative than the brief's "photorealistic" requirement wanted. Per imgai1-5, this is exactly the situation a checkpoint swap exists for: switching to a community checkpoint fine-tuned specifically toward photorealistic interior photography changes the model's own underlying visual instincts before a single prompt word is reconsidered, rather than trying to fight the base checkpoint's own bias with ever-more-specific style language.