How AI-Powered Size Extension Boosts Image Workflow Efficiency

AI size‑extension, also known as intelligent dimension scaling, uses artificial‑intelligence techniques to flexibly enlarge or adjust visual content while preserving quality, clarity, and visual effect, even improving image fidelity.

1. Analyze Requirements

The brand’s main visual KV layout is relatively fixed with predictable patterns (vertical‑to‑horizontal, horizontal‑to‑vertical, etc.). We looked for ways to break these patterns.

2. Implementation Method

Tool: libulibu online

Implementation: train a targeted LoRA model for specific scenarios – the Kontext workflow.

We enable the LoRA to learn capabilities that the base model lacks and set a trigger prompt to activate the custom Kontext LoRA model and workflow.

3. Design Process

Kontext LoRA Model Training

Kontext advantages: fast, simple, efficient; supports text editing, layout adjustment, and graphic editing.

LoRA: a lightweight fine‑tuning model inserted between the loader and sampler to add specific functions and enhance the large model.

Training method: use groups of input‑image + output‑image pairs, along with labeled prompts, so the model learns the transformation pattern (e.g., face swap, style change, angle change) that matches our layout‑adjustment needs.

Four Steps: Data Collection & Labeling – Training – Testing

Data Collection + Labeling

Labeling core: set a trigger phrase (usually functional) and describe the output image in English. Example keywords: “convert vertical layout to horizontal”, “remove text while keeping elements unchanged”, etc.

Compute Consumption & Training Steps

27 groups – 6,646 compute points (2,000 steps) – ~1‑2 hours.

60 groups – 29,907 compute points (9,000 steps, 150 steps per group) – ~3‑5 hours.

1,000 compute points can generate ~1,000 images or run 5 training cycles (e.g., 20 images × 15 repeats × 10 epochs).

Training Release

Public / private (60 groups / 27 groups).

LoRA Test Workflow

Model strength: 0.7‑1.2, steps: 20‑42.

Base algorithm: Kontext (paired with F.1 Kontext dev_fp8).

Function: multi‑size adaptation.

Keyword example: “convert vertical layout to horizontal, scale theme, delete text, keep other elements unchanged”.

Using our custom Kontext size‑adaptation model + workflow improves material production efficiency.

10% of assets in the commercial project have been AI‑processed.

20% AI‑enabled assets save 0.5‑1 workday.

Achieved batch processing of proportional and model‑specific materials.

Case: BYD Series Commercial Effect

10 car models, 10 pages, each with a main KV image and a 1:1 guide image.

Parameters: model strength 0.7, steps 42, denoising 1; using 60‑group / 27‑group size‑expansion models.

AI quickly processes image dimensions and removes text, enabling uniform secondary adjustments and reducing repetitive retouching time.

1:1 guide image – export proportional material for landing pages (low‑requirement, batch‑produced).

Keywords for guide image generation:

Keep vertical layout unchanged, shorten element spacing.

Optimize layout, improve compactness.

Delete text, keep elements unchanged.

The AI size‑extension experiment significantly improved workflow efficiency, though challenges remain such as handling low‑quality images, model/data quality, and professional entry barriers.

With ongoing AI advancements, these issues are expected to be resolved, leading to broader applications, higher‑quality extensions, and greater commercial value.