How I Automated My Social Media Pipeline with n8n, part 2 - Cloudinary and Mistral

In the first part, we set up a n8n pipeline to sync photos from my phone to a local Samba server, use a tiny CLIP-based microservice to filter out non-flower pictures, identify the flowers via the Pl@ntNet API, and store the results in Postgres.

This time, let’s build the second n8n workflow that picks up where the first one left off:

• Enhance the photos using Cloudinary
• Generate a short poetic description using Mistral AI
• Handle errors without everything falling apart

Get sources from my GitLab.

Both APIs have a generous free tier. I’m only posting on my personal Bluesky account, , so I’m not throwing money at this. The point is to stay focused on what matters: growing things in my garden, tracking blooming dates, and snapping the occasional decent shot.

Here’s the whole workflow — it might look like spaghetti at first, but don’t worry, we’ll untangle it. Under the hood, it’s all pretty chill:

The trigger is a simple cronjob. From there, an SQL query fetches all flower photos that were successfully identified but haven’t yet made it into the “post_draft” table. The graph then splits into two totally independent subflows:

• one for Cloudinary
• one for Mistral

Image Enhancement via Cloudinary

Cloudinary free tier is surprisingly generous:

25,000 transformations per month? I’ll take it. The catch: AI-based transformations are way more expensive — like 100x more expensive for example with the enhance function which is interesting.

So the first part of this branch is just me trying to avoid burning through my quota. I don’t want to transform the same image twice for no reason.

In plain English: if we need to generate a draft, we first check if the Cloudinary enhanced image already exists. If it does, we stop right there. If not, we grab the original — sitting on the shared Samba drive — and start the enhancement process.

Now comes the fun part: we get to mess with Cloudinary’s API, which is honestly great. Their image transformations are super flexible.

First step: upload the image. Unfortunately, I couldn’t find a way to apply transformations directly on the fly without uploading first.

To upload properly, you’ll need to create an Upload Preset in their dashboard:

The game with n8n is preserving the required data from node to node. In my case, it’s the primary key, the original name of the image being processed. If I lose it after a node, I’m screwed for linking the node’s result with my information system. So here, by configuring Cloudinary to preserve my image name, I make things easier for myself in n8n. After upload, Cloudinary returns my image name, which I can then use to do a DB insert or whatever

Note: Cloudinary doesn’t have a built-in credential type in n8n. Just use the Generic Credential Type → Basic Auth → enter your API keys manually. Done.

Error handling

n8n has a built-in error system at the workflow level, which is cool if you want Slack alerts or emails when something breaks.

But that’s not what I’m aiming for here. I want to catch HTTP errors, log them, and move on. No full-stop crashes. Let the other items keep flowing.

To do that, tweak the settings on your HTTP Request node. Change “On Error” to:

Now it’ll split into two branches: one for success, one for errors. That way you can log failures but keep processing the rest.

Once the image is uploaded, we move on to the transformation step.

Cloudinary image transform

I use an Edit Fields (Set) node to stash both the original upload info and the final transformation URL::

https://res.cloudinary.com/{{ cloud name }}/image/upload/c_fill,g_auto:subject,ar_1:1,w_500,q_auto/e_improve:60/{{ public_id }}.jpg

Here I’m chaining 2 transformations sequentially (see the / between the two):

• first c_fill,g_auto:subject,ar_1:1,w_500,q_auto,f_auto
• then e_improve:60

You crop first, then enhance. Doing it the other way around would mess with the lighting and contrast. Applying auto contrast to a whole image won’t have the same result as applying the same filter to part of that image.

The crop step:

• c_fill, ar_1:1, w_500 create a 500px wide square without distortion
• g_auto finds the “subject” of the image - this is why we use Cloudinary, to create auto-cropping on the photographed flower (which is hard, ask your favorite llm…). I pass it the subject parameter which activates deep learning; I could have passed classic to work with “simple” heuristics.
• q_auto to reduce the image size a bit.

Then e_improve:60 adjusts color, contrast and brightness, blending the original and filtered image for more natural results. The “60” is just the intensity

These are the parameters I chose initially, they’ll evolve - I think you need to play with the API and test the numerous settings.

And that’s the end of the branch: either the transformation fails (log it), or we save the new image.

Sometimes the result is… chef’s kiss:

Sometimes it’s an absolute disaster. Here we have a double failure: the photographer chose an aphid-eaten bluebell as his model, and the cropping software failed miserably:

But most of the time, it’s actually very good.

Draft Text Generation via Mistral

Here too, Mistral AI gives you a decent free tier. So far I’ve only hit one limit: one request per second.

n8n doesn’t care though — it blasts all items in parallel unless told otherwise. So if you’re processing a bunch of images, say goodbye to that rate limit.

Solution: throttle. Run the calls sequentially. Here’s what that branch looks like:

The prompt looks like this:

{
  "model": "mistral-small-latest",
  "messages": [
    {
      "role": "system",
      "content": "You're a botanical poet."
    },
    {
      "role": "user",
      "content": "Here is an identified plant: {{ scientificName }} It can also be called {{ commonNames[0] }}. The photo was taken on {{ date_exif }} and the filename is {{ filename }}. Describe it in a poetic sentence, specifying the Latin name, the date the photo was taken, and the filename following this format: {{ filename }} | {{ date_exif }} - {{ scientificName }} Here a poetic sentence."
    }
  ],
  "temperature": 0.7,
  "max_tokens": 1024,
  "stream": false
}

Yes, I’m relying on the LLM to follow a strict output format. Dangerous game. LLMs can hallucinate formats anytime. But here we are.

To respect the rate-limit, I use a “Loop over items” node. But here’s the catch: this node doesn’t merge outputs cleanly. So at the loop’s exit, I use a regex check to see if the filename is in the response:

If not, I merge and log everything: filename, prompt, Mistral’s answer.

Honestly, the mistral-small model performed like a champ — out of about 100 test calls, not a single one messed up the format.

Example:

2025/05/25 - Campanula persicifolia L. Under May’s sky, its white bells sing the ephemeral beauty of days.

Draft Recording

Ideally, I wanted to go a step further: save the result as a scheduled draft somewhere. Spread out the posts. But I didn’t find a proper tool for that.

Social media tools are overpriced. Hootsuite? €99/month and doesn’t support Bluesky. Postpone offers a free tier, but limits to 5 posts per month, which is a bit low. Found not one but two Bluesky Schedulers and this other one, neither worked.

I could try Zapier or other no-code toys, but the issue isn’t automation — it’s editing. I want to preview, tweak the image, rewrite a line. You know, be human.

Maybe time to try Firebase Studio? With some vibe coding?

To be continued.