FLUX.3 Image Klein brings layout control to Diffusers
Hugging Face’s October 15 integration supports generating images, editing references and specifying where elements should appear through boxes in the prompt.
In this article
Generation and editing in one workflow
The Diffusers introduction shows how a prompt can create an image or edit a supplied reference. With several references, the prompt identifies them in their supplied order. The generation-optimized distilled model is intended for image creation and editing; the base version is the route for fine-tuning the model.
Expand the request before generating
Black Forest Labs’ model guidance calls for a prompt upsampler: a language model that expands the request into a full scene description. For an edit, that description incorporates the reference and the intended change. The provider’s own upsampler or another language model can do this; Diffusers does not perform the expansion automatically.
Boxes specify placement, not a pixel lock
In the layout instructions, the scene description names each element, then a structured table supplies its description and bounding box. Coordinates run from zero to one thousand, ordered top, left, bottom, right; moving an element specifies both source and destination boxes. These boxes guide placement and scale rather than cutting off content at an exact boundary.
A suggested small trial is a subject on the left with text on the right: check the resulting positions and whether elements extend beyond their boxes.
The model card warns that edits can change content outside the requested region, and generated lettering can be inaccurate.
Sources(3)
- Diffusers integrationChecked:
- Model guidanceChecked:
- Layout instructionsChecked:
