13 files changed, 2003 insertions, 0 deletions

diff --git a/internal/image/doc.go b/internal/image/doc.go
new file mode 100644
index 0000000..ae5d586
--- /dev/null
+++ b/internal/image/doc.go
@@ -0,0 +1,2 @@
+// Package image provides provider-neutral image generation and download helpers.
+package image
diff --git a/internal/image/download.go b/internal/image/download.go
new file mode 100644
index 0000000..7a196fa
--- /dev/null
+++ b/internal/image/download.go
@@ -0,0 +1,223 @@
+package image
+
+import (
+	"context"
+	"fmt"
+	"io"
+	"os"
+	"path/filepath"
+	"strings"
+)
+
+// DownloadOptions configures image download behavior.
+type DownloadOptions struct {
+	OutputDir         string
+	OverwriteExisting bool
+	CreateDir         bool
+	FileNamePattern   string
+	MaxSizeBytes      int64
+}
+
+// DefaultDownloadOptions returns sensible defaults for image downloads.
+func DefaultDownloadOptions() *DownloadOptions {
+	return &DownloadOptions{
+		OutputDir:         "./images",
+		OverwriteExisting: false,
+		CreateDir:         true,
+		FileNamePattern:   "{word}_{source}",
+		MaxSizeBytes:      10 * 1024 * 1024,
+	}
+}
+
+// Downloader handles image downloads from search results.
+type Downloader struct {
+	provider ImageProvider
+	options  *DownloadOptions
+}
+
+// NewDownloader creates a new image downloader.
+func NewDownloader(provider ImageProvider, options *DownloadOptions) *Downloader {
+	if options == nil {
+		options = DefaultDownloadOptions()
+	}
+	return &Downloader{
+		provider: provider,
+		options:  options,
+	}
+}
+
+// DownloadImage downloads a single image to the specified path.
+func (d *Downloader) DownloadImage(ctx context.Context, result *SearchResult, outputPath string) (err error) {
+	if d == nil || d.provider == nil {
+		return fmt.Errorf("image provider is required")
+	}
+	if result == nil {
+		return fmt.Errorf("search result is required")
+	}
+
+	dir := filepath.Dir(outputPath)
+	if d.options != nil && d.options.CreateDir && dir != "" && dir != "." {
+		if err := os.MkdirAll(dir, 0o755); err != nil {
+			return fmt.Errorf("create output dir %q: %w", dir, err)
+		}
+	}
+
+	if d.options != nil && !d.options.OverwriteExisting {
+		if _, err := os.Stat(outputPath); err == nil {
+			return fmt.Errorf("output file exists: %s", outputPath)
+		}
+	}
+
+	reader, err := d.provider.Download(ctx, result.URL)
+	if err != nil {
+		return fmt.Errorf("download %q: %w", result.URL, err)
+	}
+	defer func() {
+		_ = reader.Close()
+	}()
+
+	file, err := os.Create(outputPath)
+	if err != nil {
+		return fmt.Errorf("create output file %q: %w", outputPath, err)
+	}
+	defer func() {
+		if closeErr := file.Close(); err == nil && closeErr != nil {
+			err = fmt.Errorf("close output file %q: %w", outputPath, closeErr)
+		}
+	}()
+
+	if d.options != nil && d.options.MaxSizeBytes > 0 {
+		written, copyErr := io.CopyN(file, reader, d.options.MaxSizeBytes)
+		if copyErr != nil && copyErr != io.EOF {
+			_ = os.Remove(outputPath)
+			return fmt.Errorf("write output file %q: %w", outputPath, copyErr)
+		}
+
+		if written == d.options.MaxSizeBytes {
+			var probe [1]byte
+			if n, probeErr := reader.Read(probe[:]); n > 0 || probeErr != io.EOF {
+				_ = os.Remove(outputPath)
+				return fmt.Errorf("image exceeds max size %d bytes", d.options.MaxSizeBytes)
+			}
+		}
+	} else {
+		if _, err = io.Copy(file, reader); err != nil {
+			_ = os.Remove(outputPath)
+			return fmt.Errorf("write output file %q: %w", outputPath, err)
+		}
+	}
+
+	if err := file.Sync(); err != nil {
+		return fmt.Errorf("sync output file %q: %w", outputPath, err)
+	}
+
+	if attribution := d.provider.GetAttribution(result); attribution != "" {
+		attrPath := strings.TrimSuffix(outputPath, filepath.Ext(outputPath)) + "_attribution.txt"
+		if err := os.WriteFile(attrPath, []byte(attribution), 0o644); err != nil {
+			fmt.Fprintf(os.Stderr, "Warning: failed to save attribution: %v\n", err)
+		}
+	}
+
+	return nil
+}
+
+// DownloadBestMatch downloads the best matching image for a query.
+func (d *Downloader) DownloadBestMatch(ctx context.Context, query string) (*SearchResult, string, error) {
+	opts := DefaultSearchOptions(query)
+	opts.PerPage = 5
+	return d.DownloadBestMatchWithOptions(ctx, opts)
+}
+
+// DownloadBestMatchWithOptions downloads the best matching image for given search options.
+func (d *Downloader) DownloadBestMatchWithOptions(ctx context.Context, opts *SearchOptions) (*SearchResult, string, error) {
+	if d == nil || d.provider == nil {
+		return nil, "", fmt.Errorf("image provider is required")
+	}
+	if opts == nil {
+		return nil, "", fmt.Errorf("search options are required")
+	}
+
+	searchOpts := *opts
+	searchOpts.PerPage = 5
+
+	results, err := d.provider.Search(ctx, &searchOpts)
+	if err != nil {
+		return nil, "", fmt.Errorf("search images: %w", err)
+	}
+	if len(results) == 0 {
+		return nil, "", fmt.Errorf("no images found for %q", opts.Query)
+	}
+
+	for i, result := range results {
+		filename := d.generateFileName(opts.Query, &result, i)
+		outputDir := "./images"
+		if d.options != nil && d.options.OutputDir != "" {
+			outputDir = d.options.OutputDir
+		}
+		outputPath := filepath.Join(outputDir, filename)
+
+		if err := d.DownloadImage(ctx, &result, outputPath); err == nil {
+			return &result, outputPath, nil
+		} else {
+			fmt.Fprintf(os.Stderr, "Warning: failed to download image %d: %v\n", i+1, err)
+		}
+	}
+
+	return nil, "", fmt.Errorf("no downloadable images found for %q", opts.Query)
+}
+
+func (d *Downloader) generateFileName(word string, result *SearchResult, index int) string {
+	filename := ""
+	if d != nil && d.options != nil {
+		filename = d.options.FileNamePattern
+	}
+	if filename == "" {
+		filename = "{word}_{source}"
+	}
+
+	filename = strings.ReplaceAll(filename, "{word}", sanitizeFileName(word))
+	if result != nil {
+		filename = strings.ReplaceAll(filename, "{source}", result.Source)
+		filename = strings.ReplaceAll(filename, "{id}", result.ID)
+	}
+	filename = strings.ReplaceAll(filename, "{index}", fmt.Sprintf("%d", index))
+
+	ext := ""
+	if result != nil {
+		ext = filepath.Ext(result.URL)
+		if strings.HasPrefix(result.URL, geminiDataPrefix) {
+			ext = ".png"
+		} else if ext == "" || len(ext) > 5 {
+			ext = ".jpg"
+		}
+	}
+
+	if filepath.Ext(filename) == "" {
+		filename += ext
+	}
+
+	return filename
+}
+
+func sanitizeFileName(name string) string {
+	replacer := strings.NewReplacer(
+		"/", "_",
+		"\\", "_",
+		":", "_",
+		"*", "_",
+		"?", "_",
+		"\"", "_",
+		"<", "_",
+		">", "_",
+		"|", "_",
+		" ", "_",
+		".", "_",
+	)
+
+	sanitized := replacer.Replace(name)
+	if len(sanitized) > 50 {
+		sanitized = sanitized[:50]
+	}
+
+	return sanitized
+}
diff --git a/internal/image/download_test.go b/internal/image/download_test.go
new file mode 100644
index 0000000..588d875
--- /dev/null
+++ b/internal/image/download_test.go
@@ -0,0 +1,144 @@
+package image
+
+import (
+	"context"
+	"io"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+type mockDownloaderProvider struct {
+	results       []SearchResult
+	searchErr     error
+	payload       string
+	attribution   string
+	searchQueries []string
+	downloadURLs  []string
+}
+
+func (m *mockDownloaderProvider) Name() string { return "mock" }
+
+func (m *mockDownloaderProvider) Search(_ context.Context, opts *SearchOptions) ([]SearchResult, error) {
+	if opts != nil {
+		m.searchQueries = append(m.searchQueries, opts.Query)
+	}
+	if m.searchErr != nil {
+		return nil, m.searchErr
+	}
+	return append([]SearchResult(nil), m.results...), nil
+}
+
+func (m *mockDownloaderProvider) Download(_ context.Context, url string) (io.ReadCloser, error) {
+	m.downloadURLs = append(m.downloadURLs, url)
+	return io.NopCloser(strings.NewReader(m.payload)), nil
+}
+
+func (m *mockDownloaderProvider) GetAttribution(*SearchResult) string {
+	return m.attribution
+}
+
+func TestDownloaderGenerateFileName_DataURIUsesPNG(t *testing.T) {
+	t.Parallel()
+
+	d := NewDownloader(&mockDownloaderProvider{}, &DownloadOptions{
+		FileNamePattern: "{word}_{source}",
+	})
+	result := &SearchResult{
+		URL:    "data:image/png;base64,AAAA",
+		Source: Gemini,
+	}
+
+	if got := d.generateFileName("ябълка", result, 0); got != "ябълка_gemini.png" {
+		t.Fatalf("generateFileName() = %q, want %q", got, "ябълка_gemini.png")
+	}
+}
+
+func TestDownloadImageWritesAttribution(t *testing.T) {
+	t.Parallel()
+
+	provider := &mockDownloaderProvider{
+		payload:     "image-bytes",
+		attribution: "attribution text",
+	}
+	d := NewDownloader(provider, &DownloadOptions{
+		OutputDir:         t.TempDir(),
+		CreateDir:         true,
+		OverwriteExisting: false,
+		FileNamePattern:   "{word}_{source}",
+		MaxSizeBytes:      10 * 1024 * 1024,
+	})
+
+	outputPath := filepath.Join(d.options.OutputDir, "ябълка_gemini.png")
+	if err := d.DownloadImage(context.Background(), &SearchResult{
+		URL:    "https://example.com/image.png",
+		Source: Gemini,
+		ID:     "1",
+	}, outputPath); err != nil {
+		t.Fatalf("DownloadImage() error = %v", err)
+	}
+
+	data, err := os.ReadFile(outputPath)
+	if err != nil {
+		t.Fatalf("ReadFile() error = %v", err)
+	}
+	if string(data) != "image-bytes" {
+		t.Fatalf("downloaded file = %q, want %q", string(data), "image-bytes")
+	}
+
+	attrPath := strings.TrimSuffix(outputPath, filepath.Ext(outputPath)) + "_attribution.txt"
+	attr, err := os.ReadFile(attrPath)
+	if err != nil {
+		t.Fatalf("ReadFile(attribution) error = %v", err)
+	}
+	if string(attr) != "attribution text" {
+		t.Fatalf("attribution = %q, want %q", string(attr), "attribution text")
+	}
+}
+
+func TestDownloadBestMatchWithOptions(t *testing.T) {
+	t.Parallel()
+
+	provider := &mockDownloaderProvider{
+		results: []SearchResult{
+			{
+				ID:     "1",
+				URL:    "https://example.com/image1.jpg",
+				Source: Gemini,
+			},
+		},
+		payload:     "image-bytes",
+		attribution: "attribution text",
+	}
+	d := NewDownloader(provider, &DownloadOptions{
+		OutputDir:         t.TempDir(),
+		CreateDir:         true,
+		OverwriteExisting: true,
+		FileNamePattern:   "{word}_{source}",
+		MaxSizeBytes:      10 * 1024 * 1024,
+	})
+
+	result, path, err := d.DownloadBestMatchWithOptions(context.Background(), &SearchOptions{Query: "ябълка"})
+	if err != nil {
+		t.Fatalf("DownloadBestMatchWithOptions() error = %v", err)
+	}
+	if result == nil || result.ID != "1" {
+		t.Fatalf("DownloadBestMatchWithOptions() result = %+v, want ID 1", result)
+	}
+	if !strings.HasSuffix(path, ".jpg") {
+		t.Fatalf("DownloadBestMatchWithOptions() path = %q, want jpg suffix", path)
+	}
+	if _, err := os.Stat(path); err != nil {
+		t.Fatalf("downloaded file missing: %v", err)
+	}
+}
+
+func TestDownloadImageRejectsNilResult(t *testing.T) {
+	t.Parallel()
+
+	d := NewDownloader(&mockDownloaderProvider{}, nil)
+	if err := d.DownloadImage(context.Background(), nil, filepath.Join(t.TempDir(), "out.png")); err == nil {
+		t.Fatal("expected error for nil result")
+	}
+}
diff --git a/internal/image/gemini.go b/internal/image/gemini.go
new file mode 100644
index 0000000..2ccd6fa
--- /dev/null
+++ b/internal/image/gemini.go
@@ -0,0 +1,599 @@
+package image
+
+import (
+	"bytes"
+	"context"
+	"crypto/md5"
+	"encoding/base64"
+	"encoding/hex"
+	"fmt"
+	"image"
+	_ "image/jpeg"
+	"image/png"
+	"io"
+	"net/http"
+	"strings"
+	"time"
+
+	"google.golang.org/genai"
+
+	"codeberg.org/snonux/comicforge/internal/apicircuit"
+	"codeberg.org/snonux/comicforge/internal/httpctx"
+)
+
+const (
+	// DefaultGeminiImageModel is the Gemini image model used for image generation.
+	DefaultGeminiImageModel = "gemini-3.1-flash-image-preview"
+
+	// DefaultGeminiTextModel is the Gemini text model used for scene generation.
+	DefaultGeminiTextModel = "gemini-2.5-flash"
+
+	geminiAspectRatio = "4:3"
+	geminiDataPrefix  = "data:image/png;base64,"
+	geminiSource      = Gemini
+	maxCustomPrompt   = 4000
+)
+
+// GeminiConfig holds the settings needed to build a Gemini-backed image provider.
+type GeminiConfig struct {
+	APIKey    string
+	Model     string
+	TextModel string
+}
+
+// GeminiProvider implements ImageProvider for Google Gemini image generation.
+type GeminiProvider struct {
+	client     *genai.Client
+	initErr    error
+	config     *GeminiConfig
+	lastPrompt string
+
+	// PromptCallback runs after the prompt is generated and before image creation.
+	PromptCallback func(prompt string)
+}
+
+var _ ImageProvider = (*GeminiProvider)(nil)
+
+var imageHTTPClient = httpctx.ImageDownloadHTTPClient()
+var newGeminiClient = httpctx.NewGenAIClient
+var geminiGenerateText = func(ctx context.Context, c *GeminiProvider, model, systemPrompt, userPrompt string, temperature float32, maxOutputTokens int32) (string, error) {
+	return c.generateText(ctx, model, systemPrompt, userPrompt, temperature, maxOutputTokens)
+}
+var geminiGenerateImage = func(ctx context.Context, c *GeminiProvider, prompt, aspectRatio string) ([]byte, string, error) {
+	return c.generateImage(ctx, prompt, aspectRatio)
+}
+
+// NewGeminiProvider creates a new Gemini image provider.
+func NewGeminiProvider(config *GeminiConfig) *GeminiProvider {
+	normalized := normalizeGeminiConfig(config)
+	client := &GeminiProvider{config: normalized}
+
+	if normalized.APIKey == "" {
+		return client
+	}
+
+	genaiClient, err := newGeminiClient(context.Background(), &genai.ClientConfig{
+		APIKey:  normalized.APIKey,
+		Backend: genai.BackendGeminiAPI,
+	})
+	if err != nil {
+		client.initErr = err
+		return client
+	}
+
+	client.client = genaiClient
+	return client
+}
+
+// Search generates an educational image for the requested word or phrase.
+func (c *GeminiProvider) Search(ctx context.Context, opts *SearchOptions) ([]SearchResult, error) {
+	ctx, cancel := httpctx.WithTimeoutUnlessSet(ctx, httpctx.OperationTimeoutDefault)
+	defer cancel()
+
+	if err := c.ensureReady(); err != nil {
+		return nil, err
+	}
+	if opts == nil {
+		return nil, &SearchError{
+			Provider: geminiSource,
+			Code:     "INVALID_OPTIONS",
+			Message:  "search options are required",
+		}
+	}
+
+	prompt, translatedWord, err := c.buildPrompt(ctx, opts)
+	if err != nil {
+		return nil, err
+	}
+
+	c.lastPrompt = prompt
+	if c.PromptCallback != nil {
+		c.PromptCallback(prompt)
+	}
+
+	aspectRatio := geminiAspectRatio
+	if opts.AspectRatio != "" {
+		aspectRatio = opts.AspectRatio
+	}
+
+	fmt.Printf("Gemini Image Generation Prompt (%d chars): %s\n", len(prompt), prompt)
+	fmt.Printf("Gemini Image Generation: Using model %q with aspect ratio %q\n", c.modelName(), aspectRatio)
+
+	var imageBytes []byte
+	var mimeType string
+	if len(opts.ReferenceImages) > 0 {
+		imageBytes, mimeType, err = c.generateImageWithRefs(ctx, prompt, aspectRatio, opts.ReferenceImages)
+	} else {
+		imageBytes, mimeType, err = geminiGenerateImage(ctx, c, prompt, aspectRatio)
+	}
+	if err != nil {
+		if searchErr, ok := err.(*SearchError); ok {
+			return nil, searchErr
+		}
+		return nil, &SearchError{
+			Provider: geminiSource,
+			Code:     "API_ERROR",
+			Message:  fmt.Sprintf("failed to generate image: %v", err),
+		}
+	}
+
+	dataURL, err := encodeDataURL(imageBytes, mimeType)
+	if err != nil {
+		return nil, err
+	}
+	width, height, err := decodedImageDimensions(imageBytes)
+	if err != nil {
+		return nil, err
+	}
+
+	description := fmt.Sprintf("Generated educational image for %s", opts.Query)
+	if translatedWord != "" {
+		description = fmt.Sprintf("%s (%s)", description, translatedWord)
+	}
+
+	result := SearchResult{
+		ID:           c.generateImageID(opts.Query),
+		URL:          dataURL,
+		ThumbnailURL: dataURL,
+		Width:        width,
+		Height:       height,
+		Description:  description,
+		Attribution:  "Generated by Google Gemini Nano Banana",
+		Source:       geminiSource,
+	}
+
+	return []SearchResult{result}, nil
+}
+
+// Download returns the image bytes for a data URI or a remote URL.
+func (c *GeminiProvider) Download(ctx context.Context, url string) (io.ReadCloser, error) {
+	if strings.HasPrefix(url, geminiDataPrefix) {
+		return decodeDataURL(url)
+	}
+
+	req, err := http.NewRequestWithContext(ctx, http.MethodGet, url, nil)
+	if err != nil {
+		return nil, err
+	}
+
+	resp, err := imageHTTPClient.Do(req)
+	if err != nil {
+		return nil, err
+	}
+
+	if resp.StatusCode != http.StatusOK {
+		if closeErr := resp.Body.Close(); closeErr != nil {
+			return nil, fmt.Errorf("HTTP %d: %s (failed to close response body: %v)", resp.StatusCode, resp.Status, closeErr)
+		}
+		return nil, fmt.Errorf("HTTP %d: %s", resp.StatusCode, resp.Status)
+	}
+
+	return resp.Body, nil
+}
+
+// GetAttribution returns attribution text for the generated image.
+func (c *GeminiProvider) GetAttribution(result *SearchResult) string {
+	width, height := 0, 0
+	if result != nil {
+		width = result.Width
+		height = result.Height
+	}
+
+	var attribution strings.Builder
+	attribution.WriteString("Image generated by Google Gemini Nano Banana\n\n")
+	fmt.Fprintf(&attribution, "Model: %s\n", c.modelName())
+	fmt.Fprintf(&attribution, "Text model: %s\n", c.textModelName())
+	fmt.Fprintf(&attribution, "Aspect ratio: %s\n", geminiAspectRatio)
+	fmt.Fprintf(&attribution, "Size: %dx%d\n", width, height)
+	if result != nil && result.Description != "" {
+		fmt.Fprintf(&attribution, "Result: %s\n", result.Description)
+	}
+	fmt.Fprintf(&attribution, "\nPrompt used:\n%s\n", c.lastPrompt)
+	fmt.Fprintf(&attribution, "\nGenerated at: %s\n", time.Now().Format("2006-01-02 15:04:05"))
+	return attribution.String()
+}
+
+// Name returns the provider name.
+func (c *GeminiProvider) Name() string {
+	return geminiSource
+}
+
+// LastPrompt returns the most recent image prompt.
+func (c *GeminiProvider) LastPrompt() string {
+	return c.lastPrompt
+}
+
+// SetPromptCallback registers a callback that runs after prompt generation.
+func (c *GeminiProvider) SetPromptCallback(callback func(prompt string)) {
+	c.PromptCallback = callback
+}
+
+func (c *GeminiProvider) ensureReady() error {
+	if c == nil || c.config == nil {
+		return &SearchError{
+			Provider: geminiSource,
+			Code:     "NO_CONFIG",
+			Message:  "Gemini client not initialized",
+		}
+	}
+	if c.config.APIKey == "" {
+		return &SearchError{
+			Provider: geminiSource,
+			Code:     "NO_API_KEY",
+			Message:  "Google API key not configured",
+		}
+	}
+	if c.initErr != nil {
+		return &SearchError{
+			Provider: geminiSource,
+			Code:     "CLIENT_INIT_FAILED",
+			Message:  fmt.Sprintf("failed to initialize client: %v", c.initErr),
+		}
+	}
+	if c.client == nil {
+		return &SearchError{
+			Provider: geminiSource,
+			Code:     "CLIENT_NOT_READY",
+			Message:  "Gemini client not initialized",
+		}
+	}
+
+	return nil
+}
+
+func (c *GeminiProvider) resolveTranslation(_ context.Context, opts *SearchOptions, translation string) (string, error) {
+	if translation != "" {
+		fmt.Printf("Using provided translation: %s -> %s\n", opts.Query, translation)
+		return translation, nil
+	}
+
+	return opts.Query, nil
+}
+
+func (c *GeminiProvider) resolvePrompt(ctx context.Context, opts *SearchOptions, translatedWord string) (string, error) {
+	if customPrompt := strings.TrimSpace(opts.CustomPrompt); customPrompt != "" {
+		if len(customPrompt) > maxCustomPrompt {
+			customPrompt = customPrompt[:maxCustomPrompt-3] + "..."
+		}
+		fmt.Printf("Using custom prompt: %s\n", customPrompt)
+		return customPrompt, nil
+	}
+
+	return c.createEducationalPrompt(ctx, opts.Query, translatedWord), nil
+}
+
+func (c *GeminiProvider) buildPrompt(ctx context.Context, opts *SearchOptions) (string, string, error) {
+	if opts == nil {
+		return "", "", &SearchError{
+			Provider: geminiSource,
+			Code:     "INVALID_OPTIONS",
+			Message:  "search options are required",
+		}
+	}
+
+	translation := strings.TrimSpace(opts.Translation)
+	if customPrompt := strings.TrimSpace(opts.CustomPrompt); customPrompt != "" {
+		if len(customPrompt) > maxCustomPrompt {
+			customPrompt = customPrompt[:maxCustomPrompt-3] + "..."
+		}
+		fmt.Printf("Using custom prompt: %s\n", customPrompt)
+		return customPrompt, translation, nil
+	}
+
+	translatedWord, err := c.resolveTranslation(ctx, opts, translation)
+	if err != nil {
+		return "", "", err
+	}
+
+	prompt, err := c.resolvePrompt(ctx, opts, translatedWord)
+	if err != nil {
+		return "", "", err
+	}
+
+	return prompt, translatedWord, nil
+}
+
+// createEducationalPrompt generates a prompt optimized for image generation.
+func (c *GeminiProvider) createEducationalPrompt(ctx context.Context, query, translation string) string {
+	subject := promptSubject(translation, query)
+
+	scene, err := c.generateSceneDescription(ctx, query, translation)
+	if err != nil {
+		fmt.Printf("  Failed to generate scene: %v, using basic prompt\n", err)
+		scene = ""
+	}
+	if scene != "" {
+		scene = sanitizeSceneDescription(scene)
+		if !usableSceneDescription(scene) {
+			fmt.Printf("  Scene response was too short or generic, using basic prompt\n")
+			scene = ""
+		}
+	}
+
+	selectedStyle := chooseArtisticStyle()
+	if selectedStyle == defaultArtisticStyle {
+		fmt.Printf("  No artistic styles available, using generic prompt\n")
+	}
+	fmt.Printf("  Using image style: %s\n", selectedStyle)
+
+	return buildEducationalPrompt(selectedStyle, scene, subject)
+}
+
+func (c *GeminiProvider) generateSceneDescription(ctx context.Context, query, translation string) (string, error) {
+	fmt.Printf("Gemini Scene Generation: Creating scene for %q (%s)\n", query, translation)
+
+	scene, err := geminiGenerateText(
+		ctx,
+		c,
+		c.textModelName(),
+		"You are helping create educational flashcards for language learning. Generate a brief, vivid scene description that incorporates the given English word in a memorable, contextual way. The scene should be visually interesting and help with memory retention. Keep it to 1-2 sentences, focusing on visual elements that can be illustrated. The subject (the English word) should be the clear focal point of the image, prominent and centered.",
+		fmt.Sprintf("Create a scene description for the English word %q that would make a memorable flashcard image. Make sure %q is the main focus and most prominent element in the scene.", translation, translation),
+		0.7,
+		100,
+	)
+	if err != nil {
+		return "", fmt.Errorf("scene generation failed: %w", err)
+	}
+	scene = sanitizeSceneDescription(scene)
+	if !usableSceneDescription(scene) {
+		return "", fmt.Errorf("scene generation returned unusable content")
+	}
+
+	fmt.Printf("Generated scene: %s\n", scene)
+	return scene, nil
+}
+
+func (c *GeminiProvider) generateText(ctx context.Context, model, systemPrompt, userPrompt string, temperature float32, maxOutputTokens int32) (string, error) {
+	temp := temperature
+	resp, err := apicircuit.Execute(nil, c.Name(), apicircuit.CapabilityText, func() (*genai.GenerateContentResponse, error) {
+		return c.client.Models.GenerateContent(ctx, model, []*genai.Content{
+			genai.NewContentFromText(userPrompt, genai.RoleUser),
+		}, &genai.GenerateContentConfig{
+			SystemInstruction: genai.NewContentFromText(systemPrompt, genai.RoleUser),
+			Temperature:       &temp,
+			MaxOutputTokens:   maxOutputTokens,
+		})
+	})
+	if err != nil {
+		return "", fmt.Errorf("gemini API error: %w", err)
+	}
+
+	text := strings.TrimSpace(resp.Text())
+	if text == "" {
+		return "", fmt.Errorf("no response received")
+	}
+
+	return text, nil
+}
+
+func (c *GeminiProvider) generateImage(ctx context.Context, prompt, aspectRatio string) ([]byte, string, error) {
+	if aspectRatio == "" {
+		aspectRatio = geminiAspectRatio
+	}
+
+	cfg := &genai.GenerateContentConfig{
+		ResponseModalities: []string{string(genai.ModalityImage)},
+		ImageConfig: &genai.ImageConfig{
+			AspectRatio: aspectRatio,
+		},
+	}
+
+	resp, err := apicircuit.Execute(nil, c.Name(), apicircuit.CapabilityImage, func() (*genai.GenerateContentResponse, error) {
+		return c.client.Models.GenerateContent(ctx, c.modelName(), []*genai.Content{
+			genai.NewContentFromText(prompt, genai.RoleUser),
+		}, cfg)
+	})
+	if err != nil {
+		return nil, "", &SearchError{
+			Provider: geminiSource,
+			Code:     "API_ERROR",
+			Message:  fmt.Sprintf("failed to generate image: %v", err),
+		}
+	}
+
+	imageBytes, mimeType, err := extractGeneratedImage(resp)
+	if err != nil {
+		return nil, "", &SearchError{
+			Provider: geminiSource,
+			Code:     "NO_RESULTS",
+			Message:  err.Error(),
+		}
+	}
+
+	return imageBytes, mimeType, nil
+}
+
+func (c *GeminiProvider) generateImageWithRefs(ctx context.Context, prompt, aspectRatio string, refs [][]byte) ([]byte, string, error) {
+	if aspectRatio == "" {
+		aspectRatio = geminiAspectRatio
+	}
+
+	cfg := &genai.GenerateContentConfig{
+		ResponseModalities: []string{string(genai.ModalityImage)},
+		ImageConfig:        &genai.ImageConfig{AspectRatio: aspectRatio},
+	}
+
+	parts := make([]*genai.Part, 0, len(refs)+1)
+	for _, ref := range refs {
+		if len(ref) > 0 {
+			parts = append(parts, &genai.Part{
+				InlineData: &genai.Blob{MIMEType: "image/png", Data: ref},
+			})
+		}
+	}
+	refNote := fmt.Sprintf(
+		"The %d reference image(s) above show the exact character appearance that must be preserved. "+
+			"Every character, animal, or object must look identical in the new image. Now generate:\n\n",
+		len(refs),
+	)
+	parts = append(parts, &genai.Part{Text: refNote + prompt})
+
+	resp, err := apicircuit.Execute(nil, c.Name(), apicircuit.CapabilityImage, func() (*genai.GenerateContentResponse, error) {
+		return c.client.Models.GenerateContent(ctx, c.modelName(), []*genai.Content{
+			{
+				Role:  string(genai.RoleUser),
+				Parts: parts,
+			},
+		}, cfg)
+	})
+	if err != nil {
+		return nil, "", &SearchError{
+			Provider: geminiSource,
+			Code:     "API_ERROR",
+			Message:  fmt.Sprintf("failed to generate image with refs: %v", err),
+		}
+	}
+
+	imageBytes, mimeType, err := extractGeneratedImage(resp)
+	if err != nil {
+		return nil, "", &SearchError{
+			Provider: geminiSource,
+			Code:     "NO_RESULTS",
+			Message:  err.Error(),
+		}
+	}
+
+	return imageBytes, mimeType, nil
+}
+
+func extractGeneratedImage(response *genai.GenerateContentResponse) ([]byte, string, error) {
+	if response == nil {
+		return nil, "", fmt.Errorf("no response from Gemini")
+	}
+
+	for _, candidate := range response.Candidates {
+		if candidate == nil || candidate.Content == nil {
+			continue
+		}
+
+		for _, part := range candidate.Content.Parts {
+			if part == nil || part.InlineData == nil || len(part.InlineData.Data) == 0 {
+				continue
+			}
+
+			mimeType := part.InlineData.MIMEType
+			if mimeType == "" {
+				mimeType = "image/png"
+			}
+
+			return append([]byte(nil), part.InlineData.Data...), mimeType, nil
+		}
+	}
+
+	return nil, "", fmt.Errorf("no image data returned from Gemini")
+}
+
+func encodeDataURL(imageBytes []byte, mimeType string) (string, error) {
+	if len(imageBytes) == 0 {
+		return "", fmt.Errorf("no image bytes returned")
+	}
+
+	normalizedBytes, err := normalizePNG(imageBytes, mimeType)
+	if err != nil {
+		return "", err
+	}
+
+	return geminiDataPrefix + base64.StdEncoding.EncodeToString(normalizedBytes), nil
+}
+
+func decodeDataURL(url string) (io.ReadCloser, error) {
+	header, payload, ok := strings.Cut(url, ",")
+	if !ok || !strings.HasPrefix(header, "data:") || !strings.Contains(header, ";base64") {
+		return nil, fmt.Errorf("unsupported data URI: %s", url)
+	}
+
+	data, err := base64.StdEncoding.DecodeString(payload)
+	if err != nil {
+		return nil, fmt.Errorf("decode data URI: %w", err)
+	}
+
+	return io.NopCloser(bytes.NewReader(data)), nil
+}
+
+func normalizePNG(imageBytes []byte, mimeType string) ([]byte, error) {
+	if strings.EqualFold(strings.TrimSpace(mimeType), "image/png") {
+		return append([]byte(nil), imageBytes...), nil
+	}
+
+	img, _, err := image.Decode(bytes.NewReader(imageBytes))
+	if err != nil {
+		return nil, fmt.Errorf("decode generated image: %w", err)
+	}
+
+	var buffer bytes.Buffer
+	if err := png.Encode(&buffer, img); err != nil {
+		return nil, fmt.Errorf("encode generated image as png: %w", err)
+	}
+
+	return buffer.Bytes(), nil
+}
+
+func decodedImageDimensions(imageBytes []byte) (int, int, error) {
+	cfg, _, err := image.DecodeConfig(bytes.NewReader(imageBytes))
+	if err != nil {
+		return 0, 0, fmt.Errorf("decode generated image dimensions: %w", err)
+	}
+
+	return cfg.Width, cfg.Height, nil
+}
+
+func (c *GeminiProvider) generateImageID(word string) string {
+	hash := md5.Sum([]byte(word))
+	return hex.EncodeToString(hash[:])[:8]
+}
+
+func (c *GeminiProvider) modelName() string {
+	if c == nil || c.config == nil || strings.TrimSpace(c.config.Model) == "" {
+		return DefaultGeminiImageModel
+	}
+
+	return c.config.Model
+}
+
+func (c *GeminiProvider) textModelName() string {
+	if c == nil || c.config == nil || strings.TrimSpace(c.config.TextModel) == "" {
+		return DefaultGeminiTextModel
+	}
+
+	return c.config.TextModel
+}
+
+func normalizeGeminiConfig(config *GeminiConfig) *GeminiConfig {
+	normalized := &GeminiConfig{}
+	if config != nil {
+		*normalized = *config
+	}
+
+	normalized.APIKey = strings.TrimSpace(normalized.APIKey)
+	normalized.Model = strings.TrimSpace(normalized.Model)
+	normalized.TextModel = strings.TrimSpace(normalized.TextModel)
+
+	if normalized.Model == "" {
+		normalized.Model = DefaultGeminiImageModel
+	}
+	if normalized.TextModel == "" {
+		normalized.TextModel = DefaultGeminiTextModel
+	}
+
+	return normalized
+}
diff --git a/internal/image/gemini_test.go b/internal/image/gemini_test.go
new file mode 100644
index 0000000..8c34c6e
--- /dev/null
+++ b/internal/image/gemini_test.go
@@ -0,0 +1,333 @@
+package image
+
+import (
+	"bytes"
+	"context"
+	"encoding/base64"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"google.golang.org/genai"
+)
+
+func TestNewGeminiProvider(t *testing.T) {
+	t.Parallel()
+
+	client := NewGeminiProvider(&GeminiConfig{APIKey: "test-key"})
+	if client == nil {
+		t.Fatal("expected client")
+	}
+	if client.config == nil {
+		t.Fatal("expected normalized config")
+	}
+	if client.config.Model != DefaultGeminiImageModel {
+		t.Fatalf("expected default model %q, got %q", DefaultGeminiImageModel, client.config.Model)
+	}
+	if client.config.TextModel != DefaultGeminiTextModel {
+		t.Fatalf("expected default text model %q, got %q", DefaultGeminiTextModel, client.config.TextModel)
+	}
+	if client.Name() != Gemini {
+		t.Fatalf("Name() = %q, want %q", client.Name(), Gemini)
+	}
+}
+
+func TestGeminiProvider_NoAPIKey(t *testing.T) {
+	client := NewGeminiProvider(&GeminiConfig{})
+
+	_, err := client.Search(context.Background(), DefaultSearchOptions("ябълка"))
+	if err == nil {
+		t.Fatal("expected error for missing API key")
+	}
+
+	searchErr, ok := err.(*SearchError)
+	if !ok {
+		t.Fatalf("expected SearchError, got %T", err)
+	}
+	if searchErr.Code != "NO_API_KEY" {
+		t.Fatalf("expected NO_API_KEY error, got %s", searchErr.Code)
+	}
+}
+
+func TestGeminiProvider_Search_CustomPromptSkipsTextGeneration(t *testing.T) {
+	originalText := geminiGenerateText
+	originalImage := geminiGenerateImage
+	t.Cleanup(func() {
+		geminiGenerateText = originalText
+		geminiGenerateImage = or