BlazeEdit: Mobile Image Editing Diffusion

• BlazeEdit is a mobile-first image-to-image latent diffusion model that supports five editing modes using structured visual signals.
• It employs a compact 195M-parameter architecture, eliminating text encoders through task-specific mask conditioning for low latency performance.
• Innovative techniques like decoder pruning, quantization, and two-step distillation enable real-time, privacy-preserving edits on edge devices.

Searching arXiv for BlazeEdit and closely related mobile diffusion work to ground the article in current literature. arXiv search query: "BlazeEdit Generalist Image Editing on Mobile Devices with Image-to-Image Diffusion Models" BlazeEdit is a generalist image-to-image latent diffusion model for on-device image editing introduced in "BlazeEdit: Generalist Image Editing on Mobile Devices with Image-to-Image Diffusion Models" (Deng et al., 27 May 2026). It is designed for mobile deployment and consolidates five editing modes—object removal, outpainting, tone correction, relighting, and sticker generation—within a single model of 195 million trainable parameters. Its defining premise is that many practical image editing tasks do not require text-based guidance: instead, user intent is expressed through an image, a binary mask, and a small task-specific constant. This removes the need for an external text encoder and yields a pure image-to-image design oriented toward low latency, reduced memory overhead, and privacy-preserving inference on edge devices (Deng et al., 27 May 2026).

1. Design premise and problem setting

BlazeEdit addresses a specific systems constraint in diffusion-based visual generation: modern diffusion models often achieve high generation quality at the cost of massive parameter counts, which typically force server-side inference with corresponding computational and privacy costs (Deng et al., 27 May 2026). The paper situates this against recent mobile-oriented text-to-image systems that remain relatively large, typically ranging from 0.5B to 1B parameters, and argues that practical editing workloads often expose intent directly through masking or framing rather than through natural-language prompts (Deng et al., 27 May 2026).

The model therefore adopts a pure image-to-image formulation. The input is always a triplet $(x, m, \tau_i)$, where $x \in \mathbb{R}^{3 \times 512 \times 512}$ is the original image, $m \in \{0,1\}^{1 \times 512 \times 512}$ is a binary mask, and $\tau_i \in \mathbb{R}$ is a small task-specific constant indicating which of the five editing modes should be performed (Deng et al., 27 May 2026). By concatenating $[x; \tau_i \cdot m]$ and conditioning the generative process on that representation, BlazeEdit treats task selection and spatial user intent as explicit structured inputs rather than as latent information to be extracted from text.

A common assumption in mobile diffusion is that generality requires text conditioning. BlazeEdit is organized around the opposite claim for a narrower but practically important regime: where the user has already specified the edit region or framing, task control can be handled entirely through the scaled mask. The paper states that this design sheds on the order of $0.1$–$2$ billion parameters that text encoders normally impose, making it appropriate for on-device editing workloads (Deng et al., 27 May 2026).

2. Network architecture

BlazeEdit follows the latent diffusion paradigm and is built around four major components: a frozen autoencoding pair $(\mathcal{E}, \mathcal{D})$, a jointly trained image-and-mask encoder $f_\theta$, and a U-ViT–style denoiser $\epsilon_\theta$ (Deng et al., 27 May 2026). The fixed autoencoder maps images into a compact latent space of size $x \in \mathbb{R}^{3 \times 512 \times 512}$0 and back. The encoder $x \in \mathbb{R}^{3 \times 512 \times 512}$1 is never finetuned; its cost is absorbed into the device’s read-only assets. The decoder $x \in \mathbb{R}^{3 \times 512 \times 512}$2 is aggressively pruned to 6 million parameters and then retrained to preserve reconstruction quality (Deng et al., 27 May 2026).

The image-and-mask encoder $x \in \mathbb{R}^{3 \times 512 \times 512}$3 takes a 4-channel input $x \in \mathbb{R}^{3 \times 512 \times 512}$4 and produces a multiscale feature tensor that conditions the denoiser. The denoiser $x \in \mathbb{R}^{3 \times 512 \times 512}$5 contains roughly 189 million parameters and uses a hybrid layout: at higher spatial scales $x \in \mathbb{R}^{3 \times 512 \times 512}$6 it employs residual convolutional blocks to preserve local structure, while at lower resolutions $x \in \mathbb{R}^{3 \times 512 \times 512}$7 it uses self-attention layers to capture global context (Deng et al., 27 May 2026). Together, $x \in \mathbb{R}^{3 \times 512 \times 512}$8 sum to 195 million trainable parameters; $x \in \mathbb{R}^{3 \times 512 \times 512}$9 is present at deployment but excluded from that parameter budget because it is frozen (Deng et al., 27 May 2026).

This architecture is notable for collapsing multiple heterogeneous editing objectives into a single conditional backbone. Object removal and outpainting require geometric and semantic completion; tone correction and relighting are largely photometric transformations; sticker generation introduces stylization while preserving subject identity. BlazeEdit does not instantiate separate task-specific generative models for these operations. Instead, the same backbone is reused across all five tasks, with the task-specific constant $m \in \{0,1\}^{1 \times 512 \times 512}$0 serving as the mode selector (Deng et al., 27 May 2026).

3. Diffusion process and optimization objective

BlazeEdit inherits the standard forward and reverse diffusion equations of latent diffusion models (Deng et al., 27 May 2026). Let $m \in \{0,1\}^{1 \times 512 \times 512}$1 denote the clean latent of the target image $m \in \{0,1\}^{1 \times 512 \times 512}$2. The forward process is

$m \in \{0,1\}^{1 \times 512 \times 512}$3

where

$m \in \{0,1\}^{1 \times 512 \times 512}$4

and $m \in \{0,1\}^{1 \times 512 \times 512}$5 is a fixed noise schedule (Deng et al., 27 May 2026). The denoiser $m \in \{0,1\}^{1 \times 512 \times 512}$6 is trained to recover $m \in \{0,1\}^{1 \times 512 \times 512}$7 from $m \in \{0,1\}^{1 \times 512 \times 512}$8 and conditioning features $m \in \{0,1\}^{1 \times 512 \times 512}$9.

At inference, the reverse update is

$\tau_i \in \mathbb{R}$0

Training uses mean-squared error on the predicted noise in both pretraining and downstream multitask finetuning (Deng et al., 27 May 2026). During self-supervised pretraining, the model sees random masks of varying shapes—patches, strokes, and boundary paddings—on raw images $\tau_i \in \mathbb{R}$1, and minimizes

$\tau_i \in \mathbb{R}$2

In finetuning on task $\tau_i \in \mathbb{R}$3 with true input/output pair $\tau_i \in \mathbb{R}$4, the mask is replaced by $\tau_i \in \mathbb{R}$5 and the latent is formed as $\tau_i \in \mathbb{R}$6, giving the objective

$\tau_i \in \mathbb{R}$7

Two properties are central. First, no classifier-free guidance is used. Second, no external textual signals are introduced: task control is entirely through the scaled mask (Deng et al., 27 May 2026). To reduce runtime, the fully finetuned model is then subjected to a two-step distribution-matching distillation with adversarial objectives, compressing the usual $\tau_i \in \mathbb{R}$8 denoising steps into only two while preserving visual fidelity according to the paper’s qualitative assessment (Deng et al., 27 May 2026).

4. On-device deployment profile

The deployment strategy is a stack of mutually reinforcing efficiency choices (Deng et al., 27 May 2026). Eliminating text encoders saves approximately $\tau_i \in \mathbb{R}$9–$[x; \tau_i \cdot m]$0 billion parameters compared to SnapFusion, SnapGen, or MobileDiffusion. Latent diffusion reduces the working spatial resolution from $[x; \tau_i \cdot m]$1 pixels to a $[x; \tau_i \cdot m]$2 latent grid. Decoder pruning reduces the VAE decoder to 6 million parameters with negligible quality loss. The U-ViT hybrid uses convolutional ResNet blocks at high resolution and efficient self-attention at low resolution to balance memory and compute. Quantization converts the denoiser and decoder to 8-bit integer arithmetic to leverage the Pixel 10 Edge TPU. Distillation reduces inference cost by approximately 90% relative to 50+ step sampling (Deng et al., 27 May 2026).

These design choices yield a concrete mobile execution profile. The model completes a full inference pass in 290 ms on a Pixel 10, with an end-to-end latency breakdown of 45 ms for the $[x; \tau_i \cdot m]$3 encoder, 190 ms for the two-step $[x; \tau_i \cdot m]$4 denoiser, and 55 ms for the $[x; \tau_i \cdot m]$5 decoder (Deng et al., 27 May 2026). Its download footprint in 8-bit form is under 200 MB, described as roughly half the size of competing mobile text-to-image denoisers plus text encoders (Deng et al., 27 May 2026).

Component	Function	Reported size or latency
$[x; \tau_i \cdot m]$6	image-and-mask encoder	45 ms
$[x; \tau_i \cdot m]$7	U-ViT denoiser	roughly 189M parameters; 190 ms
$[x; \tau_i \cdot m]$8	pruned decoder	6M parameters; 55 ms

The result is a model positioned for real-time editing on edge devices rather than asynchronous cloud execution. The paper frames this as a privacy-preserving deployment mode because the editing pipeline can remain on-device (Deng et al., 27 May 2026).

5. Editing tasks and empirical behavior

BlazeEdit is finetuned on five datasets spanning distinct editing regimes (Deng et al., 27 May 2026). For object removal, the model is trained on approximately 20K manually curated before/after pairs; the user provides a mask over the unwanted object, and the model fills in a semantically and structurally consistent background, including cast shadows. For outpainting, approximately 5K high-quality examples are used; given the original image and a mask on the new border region, the model extrapolates beyond the frame and changes aspect ratios seamlessly. Tone correction uses approximately 3M synthetic pairs from a high-capacity teacher; a full-image mask signals global color, exposure, and white-balance adjustments. Relighting uses approximately 100K portrait pairs with synthetic shadow augmentations; the face region mask directs the model to remove harsh shadows and harmonize lighting. Sticker generation uses approximately 100K stylized pairs via a text-to-image teacher; a subject mask specifies the region from which to generate a high-quality, identity-preserving artistic sticker (Deng et al., 27 May 2026).

Task	Training data	Conditioning role of mask
Object Removal	$[x; \tau_i \cdot m]$9K manually curated before/after pairs	unwanted object region
Outpainting	$0.1$0K high-quality examples	new border region
Tone Correction	$0.1$1M synthetic pairs from a high-capacity teacher	full-image global adjustment
Relighting	$0.1$2K portrait pairs with synthetic shadow augmentations	face region
Sticker Generation	$0.1$3K stylized pairs via a text-to-image teacher	subject region

The paper reports qualitative evidence of strong structure and semantic coherence across all tasks, as illustrated in Figure 1 (Deng et al., 27 May 2026). It also explicitly notes the absence of FID or user-study numbers. Instead, the empirical support is based on internal ablations. Two findings are highlighted: masked-reconstruction pretraining yields 5–10% better sample fidelity on object removal and outpainting than training from scratch, and the universal task-signaling mechanism matches or exceeds single-task baselines despite parameter sharing across five editing objectives (Deng et al., 27 May 2026). In addition, the two-step distillation is reported to incur no noticeable degradation in either fine details or overall realism.

A plausible implication is that BlazeEdit treats generality not as open-ended prompt following but as robust parameter sharing across a bounded family of mask-conditioned transformations. This differs from text-to-image mobile diffusion systems whose flexibility is tied to language conditioning, and it helps explain why the paper emphasizes generalist editing rather than general image generation.

6. Limitations and prospective extensions

BlazeEdit’s current resolution ceiling is $0.1$4 (Deng et al., 27 May 2026). The paper further states that extremely large outpainting ratios, such as doubling both width and height, can still produce small artifacts or incoherent borders. In scenes with very complex geometry, the two-step sampling procedure may occasionally leave faint residual noise (Deng et al., 27 May 2026). These limitations are consistent with the trade-offs implied by compact latent diffusion, aggressive pruning, and very short reverse trajectories.

The future directions named in the paper are progressive multi-resolution editing, such as $0.1$5 crops; temporal consistency for video editing; and user-guided brush controls within the same unified network (Deng et al., 27 May 2026). This suggests an effort to preserve the central design principle—a single compact, mask-conditioned image-to-image backbone—while extending it to higher-resolution and temporally coherent settings.

Within the scope defined by the paper, BlazeEdit is best understood as a mobile-first diffusion editor that replaces text conditioning with structured visual conditioning and task constants. Its technical contribution lies less in introducing a new diffusion formalism than in integrating latent diffusion, a multitask conditioning scheme, pruning, quantization, and two-step distillation into a single on-device system for five distinct editing modes under a 195M-parameter trainable budget (Deng et al., 27 May 2026).