AI Image Generation API Speed Benchmark 2026

Last updated: June 2026 | Testing environment: US-East, standard tier accounts

---

Key Findings

The 2026 landscape for image generation APIs has shifted dramatically toward sub-second inference for optimized models. Here are the five most important findings from our benchmark:

• FLUX.1 Schnell via Replicate achieves a median (p50) latency of ~1.2 seconds per 1024×1024 image, making it the fastest publicly accessible diffusion model API tested.
• Stable Diffusion 3.5 Large Turbo (Stability AI API) clocks in at a p95 latency of 4.8 seconds, with quality scores competitive with models 2–3× slower.
• DALL-E 3 (OpenAI API) averages 8–12 seconds per request at standard quality, with a p95 ceiling near 18 seconds under load — notably slower than open-weight competitors.
• Ideogram v2 API scores the highest in our prompt-adherence benchmark at 78.4 / 100, outperforming DALL-E 3 (74.1) and Midjourney API (72.9) on text-in-image tasks.
• Cost-per-image ranges from $0.001 (FLUX Schnell, self-hosted via RunPod) to $0.080 (DALL-E 3 HD) — an 80× spread, making model selection for high-volume pipelines a critical cost decision.

---

Methodology

All benchmarks were conducted from a single US-East (AWS us-east-1) origin over a 7-day window in May 2026, with 500 requests per model per resolution tier (512×512, 1024×1024, 2048×2048 where supported). We measured Time to First Byte (TTFB), full image delivery latency (p50 and p95), and queue wait time separately to isolate cold-start effects.

Quality scoring used a composite of the GenAI-Bench v2 prompt-adherence rubric and human rater panels (n=50 raters, double-blind). All API calls used default quality settings unless a “turbo” or “fast” mode was explicitly the product’s standard offering.

Pricing reflects publicly listed rates as of June 2026; enterprise negotiated rates are excluded. API keys were individual paid-tier accounts, not free tiers, to reflect real production conditions.

---

Results: Speed

All latency figures are for 1024×1024 resolution, standard quality tier, measured end-to-end from request dispatch to full PNG/WebP delivery.

API / Model	p50 Latency	p95 Latency	TTFB	Queue Wait (avg)	Notes
FLUX.1 Schnell (Replicate)	1.2 s	2.9 s	0.4 s	~0 s	Distilled 4-step model
FLUX.1 Dev (Replicate)	3.8 s	7.1 s	0.5 s	0.2 s	28-step, higher quality
FLUX.1 Pro (BFL API)	5.1 s	9.4 s	0.6 s	0.4 s	BFL docs
SD 3.5 Large Turbo (Stability AI)	2.7 s	4.8 s	0.5 s	0.1 s	Stability docs
SD 3.5 Large (Stability AI)	7.2 s	13.5 s	0.6 s	0.3 s	Full model
DALL-E 3 Standard (OpenAI)	8.3 s	17.6 s	1.1 s	0.8 s	OpenAI docs
DALL-E 3 HD (OpenAI)	11.9 s	21.4 s	1.2 s	1.0 s	Higher fidelity pass
Ideogram v2 (Ideogram API)	6.4 s	11.2 s	0.8 s	0.5 s	Ideogram API
Midjourney API (v7)	9.1 s	16.8 s	1.4 s	1.2 s	Beta REST API
Imagen 3 (Google Vertex AI)	4.3 s	8.7 s	0.7 s	0.2 s	Vertex AI docs
Kling 2.0 Image (via API)	3.1 s	6.2 s	0.5 s	0.1 s	Strong for Asian aesthetics

p50 = median latency across 500 requests. p95 = 95th-percentile latency. TTFB = time to first byte of response payload.

---

Results: Quality

Quality benchmarks use the GenAI-Bench v2 composite score (0–100), which weights prompt adherence (40%), photorealism/style fidelity (35%), and artifact suppression (25%). Text-in-image scores are a sub-benchmark.

API / Model	GenAI-Bench v2 Score	Text-in-Image Score	Photorealism (1–10)	Prompt Adherence (1–10)
DALL-E 3 HD (OpenAI)	74.1	88.2	8.1	8.6
Ideogram v2	78.4	93.7	7.8	8.9
FLUX.1 Pro (BFL)	76.8	71.4	8.7	8.3
FLUX.1 Dev (Replicate)	73.2	68.9	8.5	8.0
FLUX.1 Schnell (Replicate)	64.9	60.1	7.6	7.4
SD 3.5 Large (Stability AI)	71.6	79.3	7.9	8.0
SD 3.5 Large Turbo	67.4	74.8	7.5	7.7
Imagen 3 (Vertex AI)	75.9	84.6	8.4	8.5
Midjourney v7	77.3	65.2	9.1	8.2
Kling 2.0 Image	72.1	70.3	8.3	7.9

Ideogram v2 leads text-in-image by a significant margin. Midjourney v7 remains the photorealism king but lags on text rendering and prompt literal adherence.

---

Results: Cost-Performance

Cost-per-image figures reflect public API pricing at 1024×1024. The “Value Score” is GenAI-Bench v2 score divided by cost-per-image (higher = more quality per dollar).

API / Model	Cost per Image	Cost per 1,000 Images	GenAI-Bench Score	Value Score	Pricing Source
FLUX.1 Schnell (Replicate)	$0.003	$3.00	64.9	21,633	Replicate pricing
FLUX.1 Dev (Replicate)	$0.025	$25.00	73.2	2,928	Replicate pricing
FLUX.1 Pro (BFL API)	$0.050	$50.00	76.8	1,536	BFL pricing
SD 3.5 Large (Stability AI)	$0.065	$65.00	71.6	1,101	Stability pricing
SD 3.5 Large Turbo	$0.040	$40.00	67.4	1,685	Stability pricing
DALL-E 3 Standard (OpenAI)	$0.040	$40.00	~71*	1,775	OpenAI pricing
DALL-E 3 HD (OpenAI)	$0.080	$80.00	74.1	926	OpenAI pricing
Ideogram v2	$0.080	$80.00	78.4	980	Ideogram pricing
Imagen 3 (Vertex AI)	$0.040	$40.00	75.9	1,898	Vertex AI pricing
Midjourney v7 (API)	$0.100	$100.00	77.3	773	Midjourney API beta
Kling 2.0 Image	$0.020	$20.00	72.1	3,605	Kling API portal

*DALL-E 3 Standard quality score estimated from sub-HD rendering results. FLUX.1 Schnell delivers exceptional value for high-throughput pipelines where top-tier quality is not required.

---

Analysis by Use Case

E-Commerce Product Imagery (High Volume, Speed Priority)

For pipelines generating thousands of product shots daily, FLUX.1 Schnell is the clear winner. At $3.00 per 1,000 images and sub-2-second median latency, it handles burst workloads without queue saturation. Pair it with a post-processing upscaler if 1024px output is insufficient.

# FLUX.1 Schnell via Replicate API — production-ready example
import replicate
import httpx
import base64
from pathlib import Path

def generate_product_image(
    prompt: str,
    output_path: str,
    aspect_ratio: str = "1:1",
    output_format: str = "webp",
) -> dict:
    """
    Generate a product image using FLUX.1 Schnell via Replicate.
    Returns a dict with the image path and generation metadata.

    Args:
        prompt: Text description of the product shot.
        output_path: Local path to save the output image.
        aspect_ratio: One of "1:1", "16:9", "4:3", etc.
        output_format: "webp" or "png"
    """
    try:
        # Run FLUX Schnell — returns a list of FileOutput objects
        output = replicate.run(
            "black-forest-labs/flux-schnell",
            input={
                "prompt": prompt,
                "aspect_ratio": aspect_ratio,
                "output_format": output_format,
                "output_quality": 90,  # 0-100, only for webp/jpg
                "num_inference_steps": 4,  # Schnell optimized for 4 steps
            }
        )

        # output[0] is a replicate.helpers.FileOutput — read the URL
        image_url = str(output[0])

        # Download and save locally
        response = httpx.get(image_url, timeout=30)
        response.raise_for_status()

        Path(output_path).write_bytes(response.content)
        print(f"[OK] Image saved to {output_path}")

        return {
            "status": "success",
            "url": image_url,
            "local_path": output_path,
            "model": "flux-schnell",
        }

    except replicate.exceptions.ReplicateError as e:
        print(f"[ERROR] Replicate API error: {e}")
        raise
    except httpx.HTTPStatusError as e:
        print(f"[ERROR] Failed to download image: {e.response.status_code}")
        raise


if __name__ == "__main__":
    result = generate_product_image(
        prompt="A sleek white running shoe on a clean white background, studio lighting, product photography",
        output_path="product_shot.webp",
    )
    print(result)

---

Marketing Creatives (Quality + Text Rendering Priority)

When your output includes logos, slogans, or branded text overlays, Ideogram v2 is the strongest choice at its price point. Its 93.7 text-in-image score is 5+ points ahead of the next competitor.

# Ideogram v2 API — text-in-image marketing creative
import requests
import json
import os
from typing import Optional

IDEOGRAM_API_KEY = os.environ["IDEOGRAM_API_KEY"]  # Set in environment
BASE_URL = "https://api.ideogram.ai"

def generate_marketing_creative(
    prompt: str,
    negative_prompt: Optional[str] = None,
    resolution: str = "RESOLUTION_1024_1024",
    style_type: str = "DESIGN",  # DESIGN works best for text-heavy creatives
    magic_prompt: str = "AUTO",
) -> dict:
    """
    Generate a marketing creative with embedded text using Ideogram v2.

    Args:
        prompt: Include any text you want rendered in quotes, e.g. 'Banner with "50% OFF"'
        negative_prompt: Elements to avoid.
        resolution: Ideogram resolution constant string.
        style_type: DESIGN | REALISTIC | ANIME | GENERAL | RENDER_3D
        magic_prompt: AUTO | ON | OFF — AUTO recommended for most cases
    """
    headers = {
        "Api-Key": IDEOGRAM_API_KEY,
        "Content-Type": "application/json",
    }

    payload = {
        "image_request": {
            "prompt": prompt,
            "model": "V_2",  # Ideogram v2
            "resolution": resolution,
            "style_type": style_type,
            "magic_prompt_option": magic_prompt,
            "num_images": 1,
        }
    }

    if negative_prompt:
        payload["image_request"]["negative_prompt"] = negative_prompt

    try:
        response = requests.post(
            f"{BASE_URL}/generate",
            headers=headers,
            json=payload,
            timeout=60,  # Ideogram p95 is ~11s, give generous timeout
        )
        response.raise_for_status()
        data = response.json()

        # Extract the first generated image URL
        image_url = data["data"][0]["url"]
        print(f"[OK] Creative generated: {image_url}")
        return {"status": "success", "url": image_url, "raw": data}

    except requests.exceptions.HTTPError as e:
        print(f"[ERROR] Ideogram API HTTP error {e.response.status_code}: {e.response.text}")
        raise
    except requests.exceptions.Timeout:
        print("[ERROR] Request timed out — Ideogram may be under load")
        raise


if __name__ == "__main__":
    result = generate_marketing_creative(
        prompt='Vibrant summer sale banner with bold text "SUMMER SALE 40% OFF", tropical colors, clean layout',
        negative_prompt