stabilityai

/

stable-diffusion-3.5-large-turbo

like 684

Follow

Stability AI 37.7k

A small idea about control in img2img / SD:

We often rely on a single global strength, but in practice it bundles together very different needs.
Changing the background, adding effects, modifying the outfit, or preserving the face are not the same kind of transformation at all.

I’m wondering whether the better direction would be a per-component strength model, something like:

• identity_strength
• anatomy_strength
• outfit_strength
• background_strength
• effects_strength
• props_strength

Even if the final sampler only accepts one strength, these separate scores could still be used upstream to drive:

• the final prompt
• visual priorities
• masking
• the choice between light or strong img2img
• specialized passes

So the idea is basically to replace one global slider with a finer semantic control layer.

To me, that seems much closer to the actual kinds of transformations we ask these models to perform.

Curious whether anyone here has already experimented with something along these lines.