Andrii Bidochko

• Updated: June 12, 2026
• 6 min read

BlazeEdit: Generalist Image Editing on Mobile Devices with Image-to-Image Diffusion Models

Direct Answer

BlazeEdit is a lightweight, general‑purpose image‑to‑image diffusion model designed to run entirely on mobile devices. By stripping away text‑conditioning and consolidating multiple editing tasks into a single 195 M‑parameter network, BlazeEdit delivers near‑real‑time, privacy‑preserving image manipulation on smartphones.

Background: Why This Problem Is Hard

Modern diffusion models have set new standards for photorealistic generation, but they typically require hundreds of millions to billions of parameters and intensive GPU workloads. The consequences are twofold:

• Server‑side bottlenecks: Cloud inference incurs latency, bandwidth costs, and scaling challenges for high‑volume consumer apps.
• Privacy concerns: Sending raw user photos to remote servers exposes sensitive visual data to potential interception or misuse.

Efforts to shrink text‑to‑image models for mobile deployment have produced architectures in the 0.5 B–1 B parameter range. While these are impressive engineering feats, they remain too large for many mid‑tier smartphones, consume excessive RAM, and still rely on a textual prompt pipeline that is unnecessary for many editing scenarios (e.g., object removal or relighting). The industry therefore faces a gap: a truly on‑device, multi‑task editor that is both fast and small enough to fit within the memory envelope of everyday phones.

What the Researchers Propose

The BlazeEdit team reframes the problem from “text‑guided generation” to “task‑specific image transformation.” Their proposal consists of three intertwined ideas:

1. Eliminate text conditioning: By removing the language encoder and cross‑attention layers, the model reduces both parameter count and computational overhead.
2. Multi‑task consolidation: A single diffusion backbone is trained jointly on five common editing primitives—object removal, outpainting, tone correction, relighting, and sticker generation—using task‑specific conditioning vectors instead of free‑form text.
3. Mobile‑first architecture: The network adopts depth‑wise separable convolutions, quantization‑aware training, and a streamlined scheduler that completes a full denoising pass in under 300 ms on a Pixel 10.

In essence, BlazeEdit acts as a “Swiss‑army‑knife” diffusion engine that can be called by any on‑device UI component, delivering consistent quality across a spectrum of editing operations without ever leaving the handset.

How It Works in Practice

From a developer’s perspective, integrating BlazeEdit follows a clear, modular workflow:

1. Input preparation: The user selects an image and optionally draws a mask (for removal or outpainting) or chooses a preset style (for tone correction).
2. Task token selection: The app maps the UI action to a compact task token (e.g., REMOVE_OBJ, OUTPAINT, RELIGHT). This token replaces the textual prompt used in conventional diffusion pipelines.
3. Diffusion inference: BlazeEdit receives the image, mask, and task token, then runs a 195 M‑parameter UNet through a 20‑step scheduler. Each step refines the latent representation, guided solely by the task token.
4. Post‑processing: The final latent is decoded to an RGB image, optionally blended with the original using the mask to preserve untouched regions.

The key differentiator is the removal of a language model branch, which cuts inference time by roughly 40 % and shrinks the download size to under 200 MB. Moreover, because all computation stays on the device, the user’s photo never traverses the network, satisfying stringent privacy regulations such as GDPR and CCPA.

Evaluation & Results

The authors benchmarked BlazeEdit against two baselines: a 0.8 B‑parameter text‑to‑image diffusion model (run on a server) and a handcrafted mobile editing suite (e.g., traditional Photoshop‑style filters). Evaluation covered both quantitative metrics (FID, LPIPS) and human preference studies across the five tasks.

• Speed: On a Pixel 10, BlazeEdit completed a full edit in 290 ms, whereas the server‑based baseline incurred a 1.2‑second round‑trip latency.
• Memory footprint: The on‑device model occupied ~650 MB of RAM at peak, comfortably below the 1 GB limit of most modern Android devices.
• Quality: Across object removal and outpainting, BlazeEdit achieved an average FID of 12.4, only 1.8 points higher than the larger server model, while surpassing the handcrafted suite by a margin of 6.7 points.
• User study: In a blind 30‑participant test, 68 % of users preferred BlazeEdit’s results over the baseline, citing “natural lighting” and “seamless background fill” as primary reasons.

These findings demonstrate that a compact, text‑free diffusion model can deliver competitive visual fidelity while dramatically improving latency and privacy—a trade‑off previously thought unattainable.

Why This Matters for AI Systems and Agents

For AI‑powered agents that interact with visual content—such as autonomous photo assistants, e‑commerce recommendation bots, or AR‑enhanced chat interfaces—BlazeEdit opens new design space:

• Edge‑centric pipelines: Agents can perform on‑device visual reasoning and editing without offloading data, reducing bandwidth costs and latency spikes.
• Privacy‑first user experiences: By guaranteeing that raw images never leave the handset, developers can comply with data‑protection policies while still offering sophisticated visual transformations.
• Unified task handling: A single model replaces a suite of specialized tools, simplifying orchestration logic in multi‑modal agents.
• Scalable deployment: Because the model fits within typical mobile storage limits, it can be bundled with apps or delivered via OTA updates, ensuring consistent performance across device generations.

These advantages align directly with the capabilities of the UBOS platform overview, which emphasizes modular AI components that run at the edge, enabling developers to build privacy‑preserving agents without reinventing the diffusion stack.

What Comes Next

While BlazeEdit marks a significant step forward, several avenues remain open for exploration:

• Dynamic task expansion: Future work could introduce plug‑in style adapters that let developers add niche editing operations (e.g., style transfer) without retraining the entire backbone.
• Cross‑modal conditioning: Integrating low‑dimensional audio or gesture cues could allow hands‑free editing in AR glasses or wearables.
• Hardware acceleration: Leveraging emerging mobile NPUs or Tensor cores could push inference below 100 ms, enabling real‑time video editing.
• Robustness to diverse inputs: Extending training to low‑light, high‑ISO, and non‑RGB formats would broaden applicability in professional photography workflows.

Developers interested in extending BlazeEdit’s capabilities can explore the Workflow automation studio to prototype custom pipelines that combine on‑device diffusion with other UBOS AI services, such as voice‑driven commands via the ElevenLabs AI voice integration.

References

For a complete technical description, see the original pre‑print: BlazeEdit paper.

Andrii Bidochko

CTO UBOS

Andrii Bidochko is an AI entrepreneur and researcher focused on AI agents, reinforcement learning, and autonomous systems. He writes about the technologies shaping the future of machine intelligence, from frontier models and agent architectures to real-world AI applications.