diff --git a/internal/deepdoc/parser/pdf/inference/cache.go b/internal/deepdoc/parser/pdf/inference/cache.go
index f683f72c83..9d130e254d 100644
--- a/internal/deepdoc/parser/pdf/inference/cache.go
+++ b/internal/deepdoc/parser/pdf/inference/cache.go
@@ -123,16 +123,16 @@ func newDocAnalyzerCacheWithStore(inner doctype.DocAnalyzer, s cacheStore, ttl t
 }
 
 // cacheKey returns the cache key used for the given method +
-// image. It encodes the image to JPEG (matching what the
+// image. It encodes the image to PNG (matching what the
 // inference HTTP client actually sends) and hashes the
-// resulting bytes. JPEG encoding is deterministic for a given
-// image and quality, so two DocAnalyzer calls with the same
+// resulting bytes. PNG encoding is deterministic for a given
+// image, so two DocAnalyzer calls with the same
 // image content produce the same key.
 func cacheKey(method string, img image.Image) (string, error) {
 	if img == nil {
 		return "", fmt.Errorf("%s: nil image", method)
 	}
-	data, err := util.EncodeJPEG(img)
+	data, err := util.EncodePNG(img)
 	if err != nil {
 		return "", fmt.Errorf("%s: encode: %w", method, err)
 	}
diff --git a/internal/deepdoc/parser/pdf/inference/client.go b/internal/deepdoc/parser/pdf/inference/client.go
index 0298720a18..c2b8fb95bf 100644
--- a/internal/deepdoc/parser/pdf/inference/client.go
+++ b/internal/deepdoc/parser/pdf/inference/client.go
@@ -78,12 +78,12 @@ type bboxesResponse struct {
 
 // DLA analyzes a full page image and returns labeled regions.
 func (c *Client) DLA(ctx context.Context, pageImage image.Image) ([]pdf.DLARegion, error) {
-	data, err := util.EncodeJPEG(pageImage)
+	data, err := util.EncodePNG(pageImage)
 	if err != nil {
 		return nil, fmt.Errorf("dla: encode: %w", err)
 	}
 	var resp bboxesResponse
-	if err := c.post(ctx, "/predict/dla", data, "dla.jpeg", &resp); err != nil {
+	if err := c.post(ctx, "/predict/dla", data, "dla.png", &resp); err != nil {
 		return nil, fmt.Errorf("dla: %w", err)
 	}
 	regions := make([]pdf.DLARegion, 0, len(resp.BBoxes))
@@ -107,12 +107,12 @@ func (c *Client) DLA(ctx context.Context, pageImage image.Image) ([]pdf.DLARegio
 
 // TSR recognises table structure from a cropped image.
 func (c *Client) TSR(ctx context.Context, cropped image.Image) ([]pdf.TSRCell, error) {
-	data, err := util.EncodeJPEG(cropped)
+	data, err := util.EncodePNG(cropped)
 	if err != nil {
 		return nil, fmt.Errorf("tsr: encode: %w", err)
 	}
 	var resp bboxesResponse
-	if err := c.post(ctx, "/predict/tsr", data, "tsr.jpeg", &resp); err != nil {
+	if err := c.post(ctx, "/predict/tsr", data, "tsr.png", &resp); err != nil {
 		return nil, fmt.Errorf("tsr: %w", err)
 	}
 	cells := make([]pdf.TSRCell, 0, len(resp.BBoxes))
@@ -152,7 +152,7 @@ type ocrRecognizeResponse struct {
 // OCRDetect detects text regions (bounding boxes) in an image.
 // DeepDoc /predict/ocr with operator=det returns quad boxes: [[[x0,y0],[x1,y1],[x2,y2],[x3,y3]], ...]
 func (c *Client) OCRDetect(ctx context.Context, cropped image.Image) ([]pdf.OCRBox, error) {
-	data, err := util.EncodeJPEG(cropped)
+	data, err := util.EncodePNG(cropped)
 	if err != nil {
 		return nil, fmt.Errorf("ocr detect: encode: %w", err)
 	}
@@ -161,7 +161,7 @@ func (c *Client) OCRDetect(ctx context.Context, cropped image.Image) ([]pdf.OCRB
 	var rawEnvelope struct {
 		Output json.RawMessage `json:"output"`
 	}
-	if err := c.post(ctx, "/predict/ocr", data, "ocr_detect.jpeg", &rawEnvelope, "operator", "det"); err != nil {
+	if err := c.post(ctx, "/predict/ocr", data, "ocr_detect.png", &rawEnvelope, "operator", "det"); err != nil {
 		return nil, fmt.Errorf("ocr detect: %w", err)
 	}
 
@@ -197,12 +197,12 @@ func (c *Client) OCRDetect(ctx context.Context, cropped image.Image) ([]pdf.OCRB
 // OCRRecognize recognizes text in a cropped image region.
 // DeepDoc /predict/ocr with operator=rec returns [[["text", confidence], ...]]
 func (c *Client) OCRRecognize(ctx context.Context, cropped image.Image) ([]pdf.OCRText, error) {
-	data, err := util.EncodeJPEG(cropped)
+	data, err := util.EncodePNG(cropped)
 	if err != nil {
 		return nil, fmt.Errorf("ocr rec: encode: %w", err)
 	}
 	var result ocrRecognizeResponse
-	if err := c.post(ctx, "/predict/ocr", data, "ocr_rec.jpeg", &result, "operator", "rec"); err != nil {
+	if err := c.post(ctx, "/predict/ocr", data, "ocr_rec.png", &result, "operator", "rec"); err != nil {
 		return nil, fmt.Errorf("ocr rec: %w", err)
 	}
 	var texts []pdf.OCRText
diff --git a/internal/deepdoc/parser/pdf/inference/client_test.go b/internal/deepdoc/parser/pdf/inference/client_test.go
index 6f8e8bc016..6eaf16a2e9 100644
--- a/internal/deepdoc/parser/pdf/inference/client_test.go
+++ b/internal/deepdoc/parser/pdf/inference/client_test.go
@@ -50,8 +50,8 @@ func TestDeepDocHTTP_DLA(t *testing.T) {
 			t.Fatalf("missing 'request' multipart field: %v", err)
 		}
 		defer file.Close()
-		if !strings.HasSuffix(header.Filename, ".jpeg") {
-			t.Errorf("filename = %q, want *.jpeg", header.Filename)
+		if !strings.HasSuffix(header.Filename, ".png") {
+			t.Errorf("filename = %q, want *.png", header.Filename)
 		}
 
 		// Return canned DLA response: one table region (classId=5).
@@ -124,11 +124,11 @@ func TestDeepDocHTTP_OCRDetect(t *testing.T) {
 		if op := r.FormValue("operator"); op != "det" {
 			t.Errorf("operator = %q, want 'det'", op)
 		}
-		// Verify image is JPEG (not PNG).
+		// Verify image is PNG (not JPEG).
 		file, header, _ := r.FormFile("request")
 		defer file.Close()
-		if !strings.HasSuffix(header.Filename, ".jpeg") {
-			t.Errorf("filename = %q, want *.jpeg", header.Filename)
+		if !strings.HasSuffix(header.Filename, ".png") {
+			t.Errorf("filename = %q, want *.png", header.Filename)
 		}
 
 		// Return canned OCR detect response: 1 quad box.