Hey guys! Here is an invitation for all of you to imagine this: in a static photo, a glass of wine on a marble countertop suddenly got to swirl, light shifted, and condensation on the glass actually moved like condensation should.
After going through the creator`s Discord, I was surprised to find a model path in passing: fal-ai/minimax/hailuo-2.3/pro/image-to-video, worte it down and immediately rushed to take the test.
So, if you are curious, or even skeptical about this magic, take the journey with me to get closer to the mystery as the following article unfolds in front of us.
What Is Hailuo 2.3 Pro?
Hailuo is MiniMax‘s video generation model line — and 2.3 is their sharpest version yet. The release delivered a major leap in video realism, camera control, and motion physics, making it MiniMax’s most capable generative video model to date.
What sets it apart from most of the competition is where it focuses its energy: physics and human performance. In comparisons across the major 2026 video models, Hailuo 2.3 consistently ranks best for humans and micro-expressions — body movement, subtle gestures, and emotionally readable moments come through clearly.
The model comes in four variants, which matters for how you use it:
| Variant | Resolution | Mode | Speed | Cost (fal.ai) |
| Standard | 768p | T2V + I2V | Normal | $0.28 / 6s video |
| Pro | 1080p | T2V + I2V | Normal | $0.49 / video |
| Fast Standard | 768p | I2V only | ~50% faster | $0.19 / 6s video |
| Fast Pro | 1080p | I2V only | ~50% faster | $0.32 / 10s video |
Source: fal.ai model pricing page, verified March 2026
The Pro variant we’re covering here is the 1080p full-quality tier — the one you’d actually want for anything you plan to publish. The Fast variant is currently image-to-video only and optimized for speed with up to 50% cost reduction, making it ideal when you’re iterating quickly or testing concepts before committing to the final output.
How to Access It via Fal.ai
Account and Credits Required
Fal.ai is a serverless AI inference platform — think of it as a single API that gives you access to dozens of video and image models without building your own GPU infrastructure. It raised $140M at a $4.5B valuation in December 2025, with Sequoia, NVIDIA, and a16z as investors, and provides API access with a focus on inference speed thanks to custom CUDA kernels.
To run Hailuo 2.3 Pro, you need:
- A fal.ai account — sign up at the home page of fal (no credit card required to create an account)
- Credits loaded in your billing dashboard
On the free credits question: fal.ai offers promotional credits for new users, but they expire — there is no permanent free tier. Free credits and coupons have variable expiration depending on the specific grant, ranging from 1 week to 1 year — check the billing dashboard for your current credit balance and expiration dates.
Practically speaking: sign up, add a small amount ($5–10) to test with, and you’ll have enough budget for 10–20 Pro generations before committing further. Purchased credits last 365 days.
One thing I appreciated: server errors (HTTP 500+) are never charged — if a runner fails to process your request due to an infrastructure issue, you pay nothing. That’s a meaningful protection when you’re testing expensive video generations.
Finding the Model on Fal.ai
There are two ways to access Hailuo 2.3 Pro on fal.ai: the playground (no-code, browser-based) and the API.
Playground route: Navigate directly to:
https://fal.ai/models/fal-ai/minimax/hailuo-2.3/pro/image-to-videoThe interface accepts drag-and-drop image files, images from web pages, clipboard paste (Ctrl/Cmd+V), or a direct URL. Accepted file types are jpg, jpeg, png, webp, gif, and avif.
API route: Your model path for all API calls is:
fal-ai/minimax/hailuo-2.3/pro/image-to-videoGet your API key from the fal.ai dashboard under Settings → API Keys. Store it as an environment variable — never hardcode it in client-side code.
Step-by-Step: Image to Video with Hailuo 2.3 Pro
Using the Playground (No Code)
- Go to
fal.ai/models/fal-ai/minimax/hailuo-2.3/pro/image-to-video - Upload your image (drag-and-drop or paste from clipboard)
- Write your motion prompt in the text field
- Click Run — cost is $0.49, deducted from your credit balance
- Wait roughly 60–120 seconds for generation
- Download the output video from the result URL (files are stored on fal’s CDN for at least 7 days)
Prompt structure that works well for I2V:
[What moves] + [how it moves] + [camera behavior] + [atmosphere/quality]Example that gave me great results with a portrait photo:
“Subject slowly turns head to face camera. Hair moves gently in breeze. Soft bokeh background. Cinematic, shallow depth of field. 1080p.”
Example for a product shot:
“Liquid swirls gently inside the glass. Condensation forms on the surface. Camera pushes in slowly. Warm, moody bar lighting.”
Keep prompts under 200 words. I noticed that overloaded prompts (trying to specify five simultaneous movements) confused the model and produced worse results than clear, single-focus instructions.
Using the API (JavaScript)
The @fal-ai/client package is the current recommended SDK — the older @fal-ai/serverless-client has been deprecated. Here’s a working implementation:
import { fal } from "@fal-ai/client";
// Set your API key
fal.config({
credentials: process.env.FAL_KEY
});
const result = await fal.subscribe(
"fal-ai/minimax/hailuo-2.3/pro/image-to-video",
{
input: {
prompt: "Camera slowly pushes in. Subject turns toward light. Natural micro-expressions. Cinematic depth of field.",
image_url: "https://your-image-url.com/photo.jpg"
},
logs: true,
onQueueUpdate: (update) => {
if (update.status === "IN_PROGRESS") {
update.logs.map((log) => log.message).forEach(console.log);
}
},
}
);
console.log(result.data.video.url); // Download URL for your generated video
console.log(result.requestId);For async/queue-based workflows (better for production use where you don’t want to hold an open connection):
import { fal } from "@fal-ai/client";
// Submit job to queue
const { request_id } = await fal.queue.submit(
"fal-ai/minimax/hailuo-2.3/pro/image-to-video",
{
input: {
prompt: "Product rotates slowly on pedestal. Studio lighting. Clean white background.",
image_url: "https://your-image-url.com/product.jpg"
},
webhookUrl: "https://your-server.com/webhook/fal-results"
}
);
// Poll for status
const status = await fal.queue.status(
"fal-ai/minimax/hailuo-2.3/pro/image-to-video",
{ requestId: request_id, logs: true }
);
// Fetch result when complete
const result = await fal.queue.result(
"fal-ai/minimax/hailuo-2.3/pro/image-to-video",
{ requestId: request_id }
);
console.log(result.data.video.url);Python equivalent:
import fal_client
import os
def on_queue_update(update):
if isinstance(update, fal_client.InProgress):
for log in update.logs:
print(log["message"])
result = fal_client.subscribe(
"fal-ai/minimax/hailuo-2.3/pro/image-to-video",
arguments={
"prompt": "Camera slowly pushes in. Soft ambient light. Subject blinks naturally.",
"image_url": "https://your-image-url.com/photo.jpg"
},
with_logs=True,
on_queue_update=on_queue_update,
)
print(result["video"]["url"])Install with: pip install fal-client
Important: Generated video URLs (https://v3.fal.media/...) are publicly accessible to anyone with the link until expiration. If you need private storage, download files to your own infrastructure immediately after generation.
Output Quality and Speed
After running dozens of generations, here’s my honest read.
Where it genuinely impressed me:
In lighting-heavy scenes, Hailuo 2.3 shows mastery of exposure balance and light behavior — headlights, reflections, and motion blur all integrate naturally without overexposure or color banding. The geometry of the scene remains stable even under complex lighting, a clear improvement over earlier versions that struggled with temporal consistency.
The physics on fabric, liquids, and hair is where this model visibly pulls ahead of most alternatives. Portrait animation with realistic micro-expressions is where it truly shines — the wine glass test I started with? Exactly as good as it looked in that Discord screenshot.
The model maintains stable, high-quality visuals across different styles including anime and illustrative work — you won’t see the flicker that sometimes appears with AI when working outside of photorealism.
Where it frustrated me:
Output clip length is capped. At 1080p, Hailuo is limited to 6 seconds maximum in the Pro tier. For social content that’s usually fine, but if you need longer clips without stitching, you’ll need a different tool.
No audio. Every clip comes out silent. If your workflow requires sound, you’re adding a separate step — either MMAudio via fal.ai or a dedicated audio tool.
Generation speed: In my testing, the standard Pro endpoint averaged 90–120 seconds per generation during off-peak hours (early morning EST), stretching to 3–4 minutes during peak load periods. The Fast Pro variant at $0.32 reliably came in under 60 seconds — worth considering if speed is important to your workflow.
Free Credits and Pricing on Fal.ai
Let me be specific here because “free credits” on fal.ai requires some nuance.
New users often receive promotional free credits to get started, but these typically expire in 90 days, while purchased credits last for 365 days. There is no ongoing permanent free tier for video generation — once your intro credits are gone, you pay per generation.
Here’s the actual cost breakdown for Hailuo 2.3 on fal.ai as of March 2026:
| Endpoint | Price | Good for |
| hailuo-2.3/pro/image-to-video | $0.49 / video | Final quality outputs |
| hailuo-2.3/standard/image-to-video | $0.28 / 6s, $0.56 / 10s | Lower-res iteration |
| hailuo-2.3-fast/pro/image-to-video | $0.32 / 10s | Speed-priority Pro work |
| hailuo-2.3-fast/standard/image-to-video | $0.19 / 6s, $0.32 / 10s | Fastest + cheapest option |
Source: fal.ai official pricing, March 2026
For context on how this compares across the platform: fal.ai uses true pay-as-you-go pricing with no subscriptions or minimum commitments — you pay for generation and computing, with no hidden fees for API calls, storage, or CDN delivery.
My practical recommendation: load $20–30 to start. That gives you 40–60 Pro generations — enough to build a real sense of the model’s strengths and limitations before deciding whether to integrate it into a regular workflow. For developers evaluating the fal.ai API documentation before committing, their FAQ covers concurrency limits, error charging policies, and data retention in detail.
Hailuo 2.3 Pro vs Other Image-to-Video Models
Here’s how it stacks up against the models you’re most likely choosing between:
| Model (via fal.ai) | Cost / video | Resolution | Max length | Audio | Physics |
| Hailuo 2.3 Pro | $0.49 | 1080p | 6s | ❌ | ⭐⭐⭐⭐⭐ |
| Hailuo 2.3 Fast Pro | $0.32 | 1080p | 10s | ❌ | ⭐⭐⭐⭐ |
| Kling 2.5 Turbo Pro | $0.07/s | 1080p | 15s | ✅ | ⭐⭐⭐⭐ |
| Wan 2.2 A14B | $0.10/s | 1080p | 6s | ❌ | ⭐⭐⭐ |
| Wan Pro | $0.80/5 vids | 1080p | 6s | ❌ | ⭐⭐⭐⭐ |
Source: fal.ai pricing page, verified March 2026
Hailuo 2.3 Pro vs Kling 2.5 Turbo Pro: This comes down to what you’re filming. Hailuo AI is essentially a physics engine dressed as a video generator — it simulates how materials actually behave, like water surface tension or how silk moves differently from cotton. Kling’s strength is cinematic motion: tracking shots, multi-shot storyboards, and native synchronized audio. If your image has a human subject or physical material behavior, Hailuo often wins. If you need audio baked in or a longer clip, Kling wins cleanly.
Hailuo 2.3 Pro vs Wan 2.2: Wan 2.2 is the open-source option — teams with their own GPU infrastructure can self-host it for nearly zero marginal cost. Wan 2.2 punches well above its weight at its price point, but at the fal.ai API level, Hailuo 2.3 Pro’s physics simulation and temporal consistency are noticeably stronger for human subjects and product material work.
The honest summary: Hailuo 2.3 Pro is the specialist. It wins on physics and character performance. It loses on clip length, audio, and cinematic flexibility.
Who Should Use This
Use Hailuo 2.3 Pro via fal.ai if you’re:
- A developer building a product that needs high-quality I2V via API — the fal.ai integration is clean, well-documented, and production-ready
- An e-commerce creator animating product photos where material accuracy matters (fabric, liquid, glass, metal)
- A content creator who works with portrait photography and wants micro-expression animation that doesn’t look robotic
- Someone who generates high volume content and wants to control exactly what they pay per clip with no subscription minimum
- Testing concepts quickly using the Fast Pro variant before committing Pro credits to finals
Look elsewhere if you’re:
- A creator who needs audio in the same generation pass — to Kling 2.6 or check Wan 2.5 on fal.ai
- Working on clips longer than 6–10 seconds without wanting to stitch — Kling’s 3-minute support is unmatched
- Completely non-technical and want a GUI-only experience — fal.ai’s playground works, but MiniMax’s own Hailuo AI platform is more beginner-friendly
- Prioritizing cinematic camera language over physics realism — Veo 3.1 or Runway are stronger picks there
Conclusion
Hailuo 2.3 Pro on fal.ai is the clearest answer to one specific question: I have a great still image with a human subject or physical material, and I need it animated at 1080p with realistic motion — what do I use?
The model path fal-ai/minimax/hailuo-2.3/pro/image-to-video is straightforward, the API is clean, $0.49 per generation is competitive for the quality tier, and the physics simulation is genuinely best-in-class for character and product work.
The tradeoffs are real — 6 seconds max, no audio, a credit system with no permanent free tier. But for the use case it’s built for, I haven’t found anything that touches it at this price point.
FAQ
Q: What image formats does Hailuo 2.3 Pro accept on fal.ai?
JPG, JPEG, PNG, WebP, GIF, and AVIF. You can upload directly, paste from clipboard, drag-and-drop, or provide a public URL.
Q: Can I generate audio with Hailuo 2.3 Pro?
No. Hailuo 2.3 is a visual-only model — output clips are silent. For audio-native generation, consider Kling 2.6 via fal.ai or the Hailuo AI native platform.
Q: Can I use fal.ai-generated videos commercially?
Generally yes — fal.ai’s pay-as-you-go API grants commercial use rights for generated content. Always verify the specific terms for your use case in fal.ai’s documentation, particularly if you’re building a commercial product on top of the API.
Previous Posts:
