Edit the photo, not the person.

Start from your photo

Open Sogni Studio or Sogni Web, or run it through the Sogni Creative Agent, then pick Krea 2 Identity Edit LoRA v1.2 from the model list. It wants three inputs:

• Input: the photo you want to edit.
• Context: one or two images the edit draws from, such as a garment product shot, a pose sketch, a face or a prop.
• Instruction: one plain sentence describing the change.

Three dials steer how literal the edit is. Grounding controls how strongly the original photo is preserved: raise it to protect the parts that should stay, lower it when the model is too timid to change anything. Reference Boost is the main likeness dial: raise it when the face or garment drifts. Reference B Boost weights the second image in multi-image edits: raise it when a pose sketch is being ignored. The starting values already work for almost everything below, so leave them alone until a render tells you otherwise.

Swap the outfit from a product shot

Drop a plain e-commerce product shot, here a black zip bodysuit on a gray background, into the first Context slot, leave the dials at their defaults, and describe the end state you want:

Notice what it kept without being told: her face and expression, the double buns with star clips, the white boots, the framing, the neon puddle reflections. And notice what it understood about the garment, raglan seams, mock collar, front zipper, all transferred from a flat catalog photo and re-lit for a wet street at dusk. That is the “identity” in Identity Edit: it holds for the person and the product, which is what makes it usable for real try-on work.

Layer clothing back on

Identity editing chains: every output is a clean starting point for the next edit. Take Step 2’s output as the new input, keep the bodysuit image in the context slot, and describe the layered look you want:

Two things worth noticing. The coat was gone from the input, yet the instruction alone brought it back, because the sentence names it unambiguously. And the layering is physically plausible: the coat sits open on the shoulders, the fur overlaps the bodysuit’s edges, the lighting matches.

If you ever get a merged or repainted garment instead of true layering, name the order explicitly (“under”, “over”, “open”). The model follows spatial words closely.

Transfer a pose from a drawing

The context slots accept anything that reads as a pose: a stock illustration, a stick figure, an outline sketch, a quick doodle. It needn’t match your subject’s body, clothes or style, since the model extracts only the skeleton of the pose. Take Step 3’s output as the input and drop in a flat vector drawing of a Warrior II yoga pose:

The whole figure re-poses, arms extended, legs in a lunge, head turned to profile, and the outfit built in steps 2-3 comes along: the coat sleeves follow the arms, the fur shifts with the shoulders, the boots plant on the asphalt. The camera, the intersection and her face survive one more round-trip. Three edits deep, and she is still recognizably the same person.

Write better prompts

This model is instruction-tuned, so it wants a stage direction, not a classic image prompt. Forget quality tags, style soup and negative prompts, the winning format is one short, declarative sentence describing the end state. The rules that consistently pay off:

• Describe the result, not the operation. “She wears the black bodysuit” beats “replace her dress with a bodysuit”. State the world as you want it after the edit; the model works out what to change.
• Never write the word “reference.” You attach the extra image, but in the sentence you just name the thing plainly, “the black bodysuit”, “the green knit scarf”, “the drawing”. The model links your noun to the image you added; talking about “the reference” only confuses it.
• Name it the way you’d point at it. Describe the object as a person looking at both pictures would: color + item (“the lime fur coat”), or what the sketch is (“the outline drawing”, “the tree pose”).
• One change per prompt. Chain single edits instead of stacking three into one sentence, that’s exactly what steps 2-4 do.
• Use spatial words for layering and placement: “under the coat”, “over the shoulders”, “in her left hand”.
• Don’t re-describe what should stay. Preservation is the default, naming the hair or the street just invites the model to repaint them.
• Present tense, third person, “She wears…”, “The woman now holds…”, “The head is now…”, matches the training data and lifts the hit rate.

When a prompt is right but the render drifts, fix it with the dials rather than more words: raise Reference Boost when likeness slips, lower Grounding when the model is too timid to change anything, and raise Reference B Boost when the second image (usually the pose) is ignored.