AI Image Generation (Kandinsky) for Mobile App

Implementing AI Image Generation (Kandinsky) in a Mobile App

Kandinsky is a Russian model from Sber AI (KANDINSKY 3.1 at time of writing). Main practical advantage for products serving Russian-speaking audiences: native understanding of Russian prompts without translation. "Sunset over birch forest" on Kandinsky works the same way English prompts work on Western models — no quality loss from translation.

Available integration methods

Fusionbrain API (api.fusionbrain.ai) — official API from Kandinsky developers. Free tier, REST, relatively stable. Most integrations use this.

Replicate — Kandinsky 2.2 and 3 available as community models. Stable API, but may be older model version.

HuggingFace Inference API — kandinsky-community/kandinsky-3. Sufficient for prototypes.

For production — Fusionbrain API with custom backend proxy.

Fusionbrain API: protocol specifics

API uses two-stage model: create task first, then poll status.

class KandinskyService(private val apiKey: String, private val secretKey: String) {

    // Step 1: get model ID
    suspend fun getModelId(): String {
        val response = httpClient.get("https://api-key.fusionbrain.ai/key/api/v1/models") {
            header("X-Key", "Key $apiKey")
            header("X-Secret", "Secret $secretKey")
        }
        val models = response.body<List<FusionBrainModel>>()
        return models.first { it.name == "Kandinsky" }.id.toString()
    }

    // Step 2: create generation task
    suspend fun createTask(modelId: String, prompt: String, width: Int = 1024, height: Int = 1024): String {
        val params = JSONObject().apply {
            put("type", "GENERATE")
            put("numImages", 1)
            put("width", width)
            put("height", height)
            put("generateParams", JSONObject().apply {
                put("query", prompt)
            })
        }

        // Multipart request
        val requestBody = MultipartBody.Builder()
            .setType(MultipartBody.FORM)
            .addFormDataPart("model_id", modelId)
            .addFormDataPart(
                "params",
                "params.json",
                params.toString().toRequestBody("application/json".toMediaType())
            )
            .build()

        val response = OkHttpClient().newCall(
            Request.Builder()
                .url("https://api-key.fusionbrain.ai/key/api/v1/text2image/run")
                .header("X-Key", "Key $apiKey")
                .header("X-Secret", "Secret $secretKey")
                .post(requestBody)
                .build()
        ).execute()

        return JSONObject(response.body!!.string()).getString("uuid")
    }

    // Step 3: polling
    suspend fun pollResult(taskUuid: String): Bitmap? {
        repeat(30) {
            delay(3000)
            val response = OkHttpClient().newCall(
                Request.Builder()
                    .url("https://api-key.fusionbrain.ai/key/api/v1/text2image/status/$taskUuid")
                    .header("X-Key", "Key $apiKey")
                    .header("X-Secret", "Secret $secretKey")
                    .get()
                    .build()
            ).execute()

            val json = JSONObject(response.body!!.string())
            if (json.getString("status") == "DONE") {
                val images = json.getJSONArray("images")
                val base64 = images.getString(0)
                val bytes = Base64.decode(base64, Base64.DEFAULT)
                return BitmapFactory.decodeByteArray(bytes, 0, bytes.size)
            }
        }
        return null
    }
}

Response comes as base64 string in images field — not URL. Decode to Bitmap / UIImage directly on client. Save to internal storage if history is needed.

Generation parameters

Kandinsky supports:

• width / height: 256 to 1024, multiples of 64. Optimal: 768x768 or 1024x1024
• style: DEFAULT, ANIME, PORTRAIT, NATURE, REALISTIC (availability depends on model version)
• negativePromptDecoder: negative prompt — list of what should not be present

val params = JSONObject().apply {
    put("type", "GENERATE")
    put("numImages", 1)
    put("width", 768)
    put("height", 1024)
    put("style", "PORTRAIT")
    put("generateParams", JSONObject().apply {
        put("query", "portrait of young woman in Russian traditional costume, detailed, realism")
    })
    put("negativePromptDecoder", "blurry, artifacts, deformation, text, watermark")
}

Russian prompt vs English

Kandinsky understands Russian without quality degradation. But in practice, for technical descriptions (architecture, mechanisms), English prompts yield more precise results — model trained on mixed corpus, technical terms better represented in English. For artistic, landscape, portrait scenarios — Russian works excellently.

For maximum quality — prompt in both languages (if UI allows), Kandinsky processes both.

Integration via Replicate (alternative)

let replicateBody: [String: Any] = [
    "version": "ai-forever/kandinsky-3:...",
    "input": [
        "prompt": prompt,
        "negative_prompt": negativePrompt,
        "num_steps": 50,
        "guidance_scale": 4.0,
        "scheduler": "DDPMScheduler",
        "width": 1024,
        "height": 1024
    ]
]

Replicate provides more predictable response time (8–20 sec) than Fusionbrain during peak hours.

Common mistakes

FAIL status without explanation from Fusionbrain — usually prompt violates content policy or too short (under 3 words). Minimum prompt for stable work — 5–10 words of description.

Decoding base64 on main thread — UI block. Always background thread: DispatchQueue.global().async (iOS) or Dispatchers.Default (Android).

Timeline

Basic Fusionbrain API integration with UI — 3–4 days. Styles, generation history, gallery saving, content policy error handling — 8–12 days.