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

1. Introduction

• Text conveys what change to make but not its extent, while prior continuous-control approaches are often narrow-domain or require per-attribute training.This leaves a need for unified control across diverse edits without training separate attribute models.Source: §1 · p. 2
• Kontinuous Kontext supports gradual control from no edit to a fully realized edit, including stylization, snowfall intensity, local attributes, and object size.Figure 2 also illustrates simultaneous attribute, hair-color, and structure changes, plus changing a car’s size.Source: §1 · p. 2, Figure 2 · p. 2
• Kontinuous Kontext pairs an edit-strength scalar with the instruction and uses a lightweight projector to produce modulation coefficients that refine the instruction.The projector takes the scalar and instruction embeddings as input and outputs coefficients in the model’s modulation space.Source: §1 · p. 2
• The training data is synthesized as source-image, instruction, edited-image, and strength tuples, then filtered for identity preservation and smooth transitions.The pipeline generates instructions and edits, creates intermediate strengths, and filters noisy sequences and artifacts.Source: §1 · pp. 2–3
• Experiments report finely controlled edits across local attributes, materials, and appearance; global style, environment, and lighting; and geometric shape changes, including unseen facial attributes and body shapes.The reported scope spans diverse instruction-driven editing tasks.Source: §1 · p. 3

Instruction-driven Image Editing

• Instruction-driven editing models perform diverse tasks but lack control over edit extent, limiting fine-grained adjustments.Instruction-based editing was introduced by Instruct-Pix2Pix.Source: §2 · p. 3

Discovering Continuous Control in Generative Models

• GAN and VAE latent traversals support fine-grained attribute manipulation, but remain restricted to narrow domains.The cited traversal methods include [1] [40].Source: §2 · p. 3
• Diffusion latent traversals are challenging because denoising lacks a compact latent space, text embeddings are not smooth [18], and LoRA interpolations remain computationally expensive and concept-specific [10] [15] [16].These approaches seek latent- or weight-space directions with continuous variation.Source: §2 · p. 3

Adding Continuous Control for Image Editing

• Continuous-control methods using rendered synthetic data, token-predicting encoders, or GAN-diffusion adapters remain limited to a single attribute or object category.The section cites examples across these directions [8] [31] [39], [5] [7] [34], and [33].Source: §2 · p. 3

Image Interpolation

• Diffusion-based morphing can produce artifacts and abrupt transitions, while video inbetweening yields abrupt imaginative edits and motion blur.The cited morphing methods interpolate diffusion features [6] [50], and cited video approaches include [44].Source: §2 · p. 3

3. Method

• The method pairs synthetic source–instruction–strength–target-edit examples with a modified instruction-driven model that accepts scalar strength alongside the edit instruction.Source: §3 · pp. 3–4

3.1. Dataset

• 110K Subject200K images seed diverse global and local edit instructions from Qwen LVLM; Flux Kontext generates each corresponding full-strength edit.Source: §3.1 · p. 4
• Freemorph generates intermediate edits, but its non-smooth latent space and inversion errors motivate endpoint reconstruction and extensive data filtering.Source: §3.1 · p. 4
• Filtering thresholds trajectory uniformity by KL divergence and uses LPIPS distances to reject inversion-altered edits and near-no-op outputs, reducing the dataset from 110K to 64K.Source: §3.1 · p. 5

3.2. Kontinuous Kontext

• Scaling text-token modulation changes edit strength while preserving image identity; the projector uses pooled CLIP embeddings to calibrate strength offsets and improve transition smoothness.Source: §3.2 · p. 5, Figure 7 · p. 6
• Training tunes attention-projection LoRA and the projector with standard flow matching, randomly dropping slider conditioning with probability 0.1.Source: §3.2 · p. 6

4. Experiments

• Evaluation uses PIEbench edits across seven categories, excluding add/remove, and measures trajectory smoothness and instruction following.Smoothness is measured by maximum triangle deficit per sequence; instruction following uses CLIP directional similarity aggregated over edit strengths.Source: §4 · p. 6

4.1. Baseline Comparisons

• Our method produces gradual, identity-preserving edits and the highest linearity across interpolation baselines, measured by Pearson correlation |r|.Diffmorpher and Freemorph can distort or remove objects, while Wan may show the full edit prematurely at intermediate strengths.Source: §4.1 · pp. 6–7, Figure 9 · p. 7
• Against domain-specific methods, our method yields smoother, more consistent transitions and works out of the box for new attributes.ConceptSliders often produce weak edits, while MARBLE can jump abruptly to the final edit at lower strengths on complex real images.Source: §4.1 · p. 7, Table 2 · p. 8, Figure 10 · p. 8

4.2. Ablations

• Ablations show that text-token conditioning and removing pooled text embeddings weaken strength control, while effective data filtering improves smoothness and text alignment.Text-token conditioning produces abrupt transitions; removing pooled text embeddings yields weaker, non-smooth edits and inferior δsmooth and CLIP-dir. scores.Source: §4.2 · pp. 7–8, Table 3 · p. 8

4.3. User Study

• The reported user-study win rates (%) indicate that our method consistently outperforms all baselines across all criteria, delivering faithful edits with superior quality.Morphing-based methods often appear smooth because of continuous transitions, but can suffer from artifacts or missed edits.Source: Figure 11 · p. 8, §4.3 · p. 8

5. Discussion and Conclusions

• The method provides smooth, fine-grained edit-intensity control without sacrificing the underlying editing model's strong baseline capabilities.The discussion frames edit intensity as naturally encoded in the modulation space and uses a lightweight projector to control it across diverse edits.Source: §5 · p. 8
• Because it builds on Flux Kontext, the method inherits difficulty with precise geometric edits such as accurate object rotation or translation.Consistent extrapolating edits can also fail for large transformations.Source: §5 · p. 8
• The authors suggest spatial or temporal intensity fields as possible extensions of the lightweight modulation-space control mechanism.Source: §5 · p. 8

A.1. Implementation Details.

• The model trains a slider projector with a rank-4 LoRA on all attention layers at 512 × 512 resolution, using 66K filtered edit trajectories.Source: §A.1 · p. 10
• Models are trained for 110,000 iterations on eight NVIDIA A100 (80GB) GPUs, using effective batch size 8, constant learning rate 2 × 10^-5, and 10% slider-conditioning dropout.Training takes about 72 hours.Source: §A.1 · p. 10
• Inference uses Flux Kontext's default Euler scheduler for 28 steps, and generation time is similar to Flux Kontext because the added projector is a small MLP.Source: §A.1 · p. 10

A.2. Dataset Generation

• The pipeline draws 110K source images from Subject200K and uses Qwen-VLM to generate image-specific instructions sampled across global and local edit categories.Categories include stylization, scene reimagination, environment change, material and appearance editing, attribute modification, and shape morphing.Source: §A.2 · p. 10, Figure 13 · p. 12
• Flux Kontext generates full-strength edits from the source images and instructions; cases where it fails to make a meaningful edit are filtered using LPIPS.Source: §A.2 · pp. 10, 13
• Freemorph generates five intermediate edits between source and full edit, using LLaVA captions and diffusion-feature interpolation during denoising.The resulting trajectories serve as targets for intermediate edit strengths.Source: §A.2 · p. 12
• Filtering removes poor inversions, non-smooth trajectories, and weak edits; trajectory smoothness is assessed from adjacent-image distances using KL divergence from a discrete uniform distribution.A 0.15 KL threshold filters non-smooth trajectories, and LPIPS thresholds detect poor inversion and weak edits.Source: §A.2 · pp. 12–13
• Filtering thresholds are selected by manually inspecting about 100 edit sequences per category, with VLM-based threshold selection suggested for adding data from new sources.Source: §A.2 · p. 13

• The projector is a four-layer MLP that takes a 1536-dimensional input combining the scale value and pooled CLIP text embedding.The scale and text components each contribute 768 dimensions.Source: §A.3 · p. 13
• Its 6144-dimensional output is split into 3072-dimensional offsets for the modulation parameters Δyscale and Δyshift.Source: §A.3 · p. 13
• The scale value is sinusoidally encoded to 128 dimensions, linearly mapped to 768 dimensions, then concatenated with the text embedding; projector overhead is negligible.Source: §A.3 · p. 13

• Scaling text modulation parameters by v = (0.5, 1.3) produced varied edits, but v did not directly correlate with edit strength.Source: §A.4.1 · p. 16
• The experiment varied the text-token modulation parameters using scaling values v = (0.5, 1.3).Source: §A.4.1 · p. 16
• The authors identify a learned, calibrated mapper such as the slider projector as necessary for accurate strength control through modulation parameters.Source: §A.4.1 · p. 16

A.4.2. Evaluation of identity preservation

• Across edit strengths, Kontinuous Kontext follows the desired similarity decay while keeping even the strongest edits close to the source; MARBLE and Concept Sliders fail to match that trend or preserve identity.The comparison uses CLIP-Image and DINO-Image similarity; Figure 17 shows the identity-preservation trends.Source: §A.4.2 · p. 16, Figure 17 · p. 17

A.4.3. Faithfulness and quality tradeoff.

• Across edit strengths, the method shows monotonic, smooth transitions in faithfulness and identity preservation, with higher faithfulness and less identity change than baselines at s = 0.50.Baselines either diverge from the source at small strengths or are non-monotonic.Source: §A.4.3 · p. 16, Figure 18 · p. 17

A.5. Ablation study

• Injecting slider output into text embeddings causes abrupt jumps, while omitting pooled text input yields non-smooth trajectories; modulation-space injection with instruction-adaptive projection produces smooth trajectories.Figure 19 illustrates the architecture ablation.Source: §A.5 · p. 16, Figure 19 · p. 18

A.6.1. Smoothness of the edit sequence

• In a user study, 20 volunteers judged 40 sequence pairs; among first- and second-order smoothness configurations using LPIPS and DreamSim, δ2 (DreamSim) aligns best with preferences, capturing consistency in the rate of change.Source: §A.6.1 · p. 18

A.7. Qualitative Comparison

• Interpolation baselines can introduce artifacts or omit details, while WAN can produce abrupt jumps; the study also compares against domain-specific methods and inference-time baselines.The reported inference-time baselines fail to produce smooth transitions and significantly distort input identity.Source: §A.7 · p. 19, Figure 22 · p. 20, §A.8 · p. 20

A.9. Failure case - Extrapolation beyond the training strength s > 1

• Beyond s = 1, the method may fail to apply edits or reduce their extent; evaluation uses PIEBench categories except add/remove-object edits, which are discrete.The evaluation also uses selected generated images for Concept Sliders and 40 PIEBench material-editing images for MARBLE.Source: §A.9 · pp. 21–22, §A.10 · pp. 23, 25

A.11. Compositional Editing

• Kontinuous Kontext supports multi-turn editing, and its strength-conditioned projector is described as compatible with recent instruction-based editors that use DiT blocks.The paper demonstrates multi-turn editing by applying two instructions in sequence.Source: §A.11 · p. 25, §A.12 · p. 25

A.13. Latency and Computational Cost

• Inference takes 60s versus 56s for the base Flux Kontext model on one NVIDIA A6000 GPU with 28 denoising steps; the method uses the same or fewer diffusion evaluations than baselines except WAN.The trainable projector and LoRA parameters account for approximately 1% of the base model's parameters.Source: §A.13 · p. 25, Table 5 · p. 25