diff --git a/backend/internal/cfai/image.go b/backend/internal/cfai/image.go
index 264a435..bd09738 100644
--- a/backend/internal/cfai/image.go
+++ b/backend/internal/cfai/image.go
@@ -29,8 +29,11 @@ import (
 	"encoding/json"
 	"fmt"
 	"image"
+	"image/draw"
+	"image/jpeg"
 	_ "image/jpeg" // register JPEG decoder
-	_ "image/png"  // register PNG decoder
+	"image/png"
+	_ "image/png" // register PNG decoder
 	"io"
 	"net/http"
 	"time"
@@ -189,6 +192,11 @@ func (c *imageGenHTTPClient) GenerateImage(ctx context.Context, req ImageRequest
 	return c.callImageAPI(ctx, req.Model, body)
 }
 
+// refImageMaxDim is the maximum dimension (width or height) for reference images
+// sent to Cloudflare Workers AI. CF's JSON body limit is ~4 MB; a 768px JPEG
+// stays well under that while preserving enough detail for img2img guidance.
+const refImageMaxDim = 768
+
 // GenerateImageFromReference generates an image from a text prompt + reference image.
 func (c *imageGenHTTPClient) GenerateImageFromReference(ctx context.Context, req ImageRequest, refImage []byte) ([]byte, error) {
 	if len(refImage) == 0 {
@@ -196,6 +204,10 @@ func (c *imageGenHTTPClient) GenerateImageFromReference(ctx context.Context, req
 	}
 	req = applyImageDefaults(req)
 
+	// Shrink the reference image if it exceeds the safe payload size.
+	// This avoids CF's 4 MB JSON body limit and reduces latency.
+	refImage = resizeRefImage(refImage, refImageMaxDim)
+
 	var body map[string]any
 	if req.Model == ImageModelSD15Img2Img {
 		pixels, err := decodeImageToRGBA(refImage)
@@ -286,6 +298,60 @@ func applyImageDefaults(req ImageRequest) ImageRequest {
 	return req
 }
 
+// resizeRefImage down-scales an image so that its longest side is at most maxDim
+// pixels, then re-encodes it as JPEG (quality 85). If the image is already small
+// enough, or if decoding fails, the original bytes are returned unchanged.
+// This keeps the JSON payload well under Cloudflare Workers AI's 4 MB body limit.
+func resizeRefImage(data []byte, maxDim int) []byte {
+	src, format, err := image.Decode(bytes.NewReader(data))
+	if err != nil {
+		return data
+	}
+	b := src.Bounds()
+	w, h := b.Dx(), b.Dy()
+
+	longest := w
+	if h > longest {
+		longest = h
+	}
+	if longest <= maxDim {
+		return data // already fits
+	}
+
+	// Compute target dimensions preserving aspect ratio.
+	scale := float64(maxDim) / float64(longest)
+	newW := int(float64(w)*scale + 0.5)
+	newH := int(float64(h)*scale + 0.5)
+	if newW < 1 {
+		newW = 1
+	}
+	if newH < 1 {
+		newH = 1
+	}
+
+	// Nearest-neighbour downsample (no extra deps, sufficient for reference guidance).
+	dst := image.NewRGBA(image.Rect(0, 0, newW, newH))
+	for y := 0; y < newH; y++ {
+		for x := 0; x < newW; x++ {
+			srcX := b.Min.X + int(float64(x)/scale)
+			srcY := b.Min.Y + int(float64(y)/scale)
+			draw.Draw(dst, image.Rect(x, y, x+1, y+1), src, image.Pt(srcX, srcY), draw.Src)
+		}
+	}
+
+	var buf bytes.Buffer
+	if format == "jpeg" {
+		if encErr := jpeg.Encode(&buf, dst, &jpeg.Options{Quality: 85}); encErr != nil {
+			return data
+		}
+	} else {
+		if encErr := png.Encode(&buf, dst); encErr != nil {
+			return data
+		}
+	}
+	return buf.Bytes()
+}
+
 // decodeImageToRGBA decodes PNG/JPEG bytes to a flat []uint8 RGBA pixel array
 // required by the stable-diffusion-v1-5-img2img model.
 func decodeImageToRGBA(data []byte) ([]uint8, error) {