Alright, I’m hoping someone here has cracked the code on this. I’m deep in a project where we’re using Leonardo to generate some lifestyle imagery for a nurture campaign. The initial outputs are 90% there—like, the client loves the composition and style, but then comes the inevitable: “Can we make the woman’s jacket blue instead of red?” or “Move that coffee cup slightly to the left?”
I know you can’t just “edit” a pixel in the traditional sense, but I’ve been experimenting with a few workflows and I’m curious what’s working for everyone else.
My current go-to is using the Image Guidance feature with a prompt that’s *almost identical* but with the specific change called out, and then cranking up the guidance strength. It’s hit or miss, though. Sometimes it changes the whole vibe, not just the jacket color. I’ve also had some luck with inpainting on a very small area, but you need a really tight mask and even then, the coherence can be off.
Has anyone found a reliable method for these precise, client-level tweaks? Are we better off using something like Photoshop for the minor stuff and then maybe running it back through Leonardo for style consistency? I’m trying to keep the workflow inside one tool where possible, but maybe that’s not realistic for minor edits. Would love to hear your hacks!
Automate all the things.
Yeah, I've been trying to figure this out too. When you say inpainting with a tight mask, do you find it works better if you use a really descriptive prompt just for that one spot? Like "a dark blue denim jacket" instead of just "blue jacket"?
I'm still new to this, but I've also heard some people will generate a bunch of variations first, then use the one that's closest and do the final tweak in a photo editor to keep things simple. Do you think that's cheating?
Relying solely on inpainting for client revisions is building a house on a foundation of sand. You're right about coherence being the first casualty.
The fundamental error here is attempting to treat a probabilistic output as a deterministic asset. When a client approves an image "90% there," they're approving a specific visual arrangement that the model may be incapable of reproducing with a single-element alteration. Your workflow should reflect that reality.
The pragmatic method is a hybrid pipeline: use the generated image as a baseplate in a proper raster editor (Photoshop, Affinity) to make the precise, client-driven tweaks - color swaps, object moves, text placement. Then, if the edit has damaged the stylized coherence, run *that* edited image back through img2img with a low noise setting to re-harmonize the style. This acknowledges the model's strength (style, texture) while using precise tools for the tasks they're actually good at.
Chasing perfection within the generator for pixel-level edits is a misallocation of effort. You're an artist directing multiple tools, not a priest praying to a single black box. 😉
James K.