Kontinuous Kontext: Continuous Strength Control for Instruction-based Image Editing

TL;DR

Kontinuous Kontext extends FLUX.1-Kontext (an instruction-driven image editing diffusion model) with a scalar edit-strength input s ∈ [0, 1], turning text-only instruction editing into a slider. A lightweight projector maps (s, pooled CLIP embedding of the instruction) to coefficients in the model’s modulation space; the backbone is adapted with attention-layer LoRAs. Training data is synthetic: source-image + VLM-generated instruction → Flux-Kontext edit → diffusion-based image-morphing interpolation between source and edit → quality filtering, yielding 60K image-edit-instruction-strength quadruplets. The result is a unified slider that works across stylization, attribute, material, background, and shape changes without per-attribute training.

Key claims

Edit strength is naturally encoded in the modulation space of flow-based instruction-driven editing models like FLUX.1-Kontext, and can be exposed as a continuous control by learning a lightweight projector into that space [§1, §3].
The projector is conditioned on both the scalar s and the pooled CLIP embedding of the edit instruction, making the modulation adjustment instruction-adaptive rather than a global “edit intensity” knob [§3].
Synthesizing strength-labeled data via a three-stage pipeline — VLM-proposed instructions, Flux-Kontext edits, diffusion-based morphing between source and edit, and quality filtering — produces 60K usable image-edit-instruction-strength quadruplets [§4, Project page §Dataset construction].
A single trained model handles diverse edit categories (stylization, attribute, material, background, shape) without attribute-specific training [Abstract, §5].
Training uses LoRA on all attention layers in addition to the projector, under a standard diffusion denoising loss [§3, Project page §Method].
The same lightweight-projector-into-modulation-space recipe is conjectured to extend to spatial or temporal intensity fields, not just a global scalar [§6 Conclusion].

Method

The base model is FLUX.1-Kontext (dev), an MMDiT image editing model that already takes a source image + text instruction and produces an edited image. Kontinuous Kontext augments this with a scalar edit strength s. A small projector network takes (s, pooled_CLIP_embed(instruction)) and emits coefficients that are added to the model’s modulation parameters on the text-token side — i.e. it perturbs the AdaLN-style scale/shift signals that the instruction tokens drive. A LoRA is inserted on all attention layers so the backbone can co-adapt with the projector. Training data is synthetic: (1) a VLM proposes edit instructions for each source image; (2) Flux-Kontext produces an edited target; (3) a diffusion-based image morphing model generates intermediate frames between source and target, providing supervision at varying strength levels; (4) extensive filtering removes artifacts and identity drift. Final dataset is ~60K image-edit-instruction-strength quadruplets. The model is trained with the standard diffusion denoising loss on this synthetic supervision.

Results

The paper reports continuous, monotone strength control across diverse edits (style, attribute, material, background, shape) on the project page, with the slider smoothly interpolating from “no change” at s=0 to a fully realized edit at s=1, and supporting extrapolation beyond s=1 in some categories. Headline numbers from the arxiv version aren’t surfaced in the project-page excerpt; the qualitative claim is that one trained model + one slider replaces per-attribute training of separate strength controllers.

Why it’s interesting

This is a clean instance of the recurring recipe the wiki tracks under Parameter-Efficient Finetuning: frozen large pretrained editing DiT + small adapter + a structured conditioning slot that’s not just text. Where FIT: A Large-Scale Dataset for Fit-Aware Virtual Try-On swaps the text-token slot for a measurement embedding and DiffHDR: Re-Exposing LDR Videos with Video Diffusion Models / HDR Video Generation via Latent Alignment with Logarithmic Encoding (LumiVid) reshape the input distribution with a log-gamma transform, Kontinuous Kontext inserts a scalar axis directly into the modulation space — arguably the most economical of the three because no input-space change is needed. It also relates to Qwen-Edit-2509-Multiple-angles LoRA — community camera-control adapter for Qwen-Image-Edit-2509 in spirit (closed-set continuous control surfaced on top of a frozen image-edit DiT) but exposes a single dense slider rather than a discrete language menu. The synthetic-data pipeline — using a second generative model (diffusion morphing) to manufacture strength labels that don’t exist in real data — is a useful pattern for any future “expose a continuous control axis” project.