PromptPack adds training-free spatial and semantic control to diffusion image editors
The arXiv paper describes Edit Guides that tell diffusion editors where changes may occur and which concepts apply where, without retraining.
In plain words
PromptPack is a method that gives AI image editors finer control over edits. It labels each pixel as Keep, Permit, or Modify, and assigns specific text prompts to specific regions. It works by guiding existing editors at inference time, so no retraining is needed.
What was announced
The paper argues that instruction-driven editing pipelines often leave user intent underspecified, while inpainting editors define where an edit happens but not which concept belongs to which region. PromptPack addresses both gaps with two Edit Guides: an edit-role partition marking pixels as Keep, Permit, or Modify, and a regional-prompt assignment mapping regions to text prompts. For source-object edits, the guides can be derived automatically from a text-grounded SAM 3 mask.
Without retraining, the method injects independently embedded regional prompts into the editor's fixed-length text sequence and uses the guides to mask attention, gate flow updates, and restore protected source pixels. The authors integrated it into FlowEdit FLUX, FlowEdit SD3, and FLUX Kontext, reporting improved concept placement and protected-content preservation on PIE-Bench++ across all three. In a pairwise user study, PromptPack achieved an 87% preference rate over FlowEdit FLUX.
What it does not show
- The evidence is the paper's abstract only; results and the user study are author-reported and not independently confirmed.
- The abstract does not state whether code or integrations are publicly available.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.