At Inithouse, a studio shipping a growing portfolio of products in parallel, we build tools that do one thing and do it fast. Pet Imagination generates AI portraits of pets in 9 art styles. Upload a photo, pick a style, get your portrait. No signup, no waiting around. The whole thing runs under 60 seconds.
This post breaks down how the pipeline works: the prompt skeleton, the style matrix, and the tricks we use to keep latency low.
The flow has four stages:
Upload and preprocessing. The user uploads a pet photo. We validate the image (format, size, aspect ratio), compress it if needed, and extract the key visual features. The preprocessing step also detects whether the subject is a dog, cat, or another animal. This matters because prompts need to reference the animal type explicitly for consistent output.
Prompt construction. This is where the style matrix kicks in. We build a prompt from a skeleton template and inject style-specific parameters. More on this below.
Generation. The constructed prompt and the preprocessed image hit the image generation API. We measured average response times across hundreds of generations: most styles return in 15 to 35 seconds depending on complexity.
Delivery. The generated portrait gets served to the user. We delete the original upload after processing. No photos stored on our servers, no accounts to worry about.
Every generation starts from the same base structure:
[STYLE_PREFIX]
A portrait of a [ANIMAL_TYPE] in [STYLE_DESCRIPTION].
The subject should be the focal point, [COMPOSITION_RULES].
[QUALITY_MODIFIERS]
[NEGATIVE_CONSTRAINTS]
The skeleton stays constant. What changes per style are five parameter slots:
STYLE_PREFIX
: sets the overall artistic direction (oil painting, anime cell shading, pencil sketch)STYLE_DESCRIPTION
: detailed style instructions (brushstroke texture, color palette, line weight)COMPOSITION_RULES
: framing and layout rules specific to the styleQUALITY_MODIFIERS
: resolution and detail levelNEGATIVE_CONSTRAINTS
: what to avoid (distortion, extra limbs, text artifacts)By keeping the skeleton fixed and parameterizing only the style layer, we get predictable quality across all 9 styles without maintaining 9 separate prompt chains.
Each of the 9 styles maps to a distinct set of parameters:
| Style | Prefix direction | Key visual trait | Avg. generation time |
|---|---|---|---|
| Renaissance | Classical oil painting | Rich textures, dramatic lighting | ~30s |
| Watercolor | Loose watercolor wash | Soft edges, transparent layers | ~25s |
| Anime | Cell-shaded illustration | Clean lines, large expressive eyes | ~20s |
| Sketch | Graphite pencil drawing | Cross-hatching, paper texture | ~18s |
| Sheriff | Western portrait | Sepia tones, badge and hat | ~28s |
| Wizard | Fantasy illustration | Robes, magical particles | ~32s |
| Astronaut | Sci-fi portrait | Space suit, starfield background | ~30s |
| Final Boss | Video game boss art | Dramatic pose, energy aura | ~35s |
| Blocky | Voxel/pixel art | Cubic geometry, limited palette | ~22s |
The fun styles (Sheriff, Wizard, Astronaut, Final Boss) need more detailed negative constraints to avoid generic fantasy clip art. We found that adding specific composition rules for these styles ("the pet's face must remain recognizable as the original") reduced uncanny output by roughly 40%.
Getting the full pipeline under 60 seconds took some iteration. Here is what moved the needle:
Parallel preprocessing. Image validation and animal detection run concurrently, not sequentially. This shaved about 3 seconds off the preprocessing stage.
Prompt caching. The style parameters rarely change. We precompute and cache the full prompt template for each style. At generation time, we only inject the animal type and any image-specific adjustments. Template assembly takes under 50ms instead of building from scratch each time.
Compression before upload. We resize and compress on the client side before sending to the server. Smaller payload means faster upload, and the generation API handles compressed input fine. We measured no visible quality difference in outputs when compressing input images to 70% quality JPEG.
Timeout and retry strategy. If generation exceeds 45 seconds, we cancel and retry once with a slightly simplified prompt (fewer quality modifiers). About 8% of generations hit the timeout on first attempt; the simplified retry succeeds in under 30 seconds almost every time.
Building image generation at Pet Imagination taught us patterns we applied elsewhere. At Ziva Fotka, our tool that turns still photos into short animated videos, we reused the same parallel preprocessing approach. The latency savings compound when you process thousands of images daily.
We also found that exposing the generation pipeline openly (showing progress stages to the user instead of a blank screen) reduced perceived wait time. Users who see "preprocessing your photo" followed by "generating portrait" report the process feeling faster than users who just see a spinner. We observed the same pattern across multiple products in the portfolio.
The full prompt skeleton and style matrix are what make Pet Imagination consistent at speed. Nine styles, one pipeline, no accounts, photos deleted after processing. If you want to try it: petimagination.com.
We keep shipping tools like this at Inithouse, a studio running parallel product experiments. Pet portraits today, animated photos tomorrow. Each product in the portfolio teaches us something about the next one.