Wendeline

Reformed lurker. 20 yrs NYC property management. Built a one-person AI pipeline that ships a weekly horror series (KILLIAN). Now building it for clients.

[Wendeline @wjcnyc]

I've spent 20 years in NYC property management.

On the side I built a one-person AI video pipeline that ships a horror series every week. No crew. No payroll. Just me and a locked workflow.

This is 15 seconds of last week's episode.

Now I'm building these for other people's businesses. If you want one, my DMs are open.

[Wendeline @wjcnyc]

25 years. I was there, and I still carry names with me.

At the memorial they're read aloud one at a time, because every single one was a whole life: a desk someone never came back to, a kitchen that stayed quiet, a family that kept going anyway.

For everyone who was lost, everyone who ran toward it, and everyone who's spent 25 years living with the empty chair. We remember. We never forget.

[Wendeline @wjcnyc]

Same character, two different AI systems.

Left: the hero reference that's anchored 21 episodes. Right: the same woman out of the new engine I've been testing, warmer skin, softer paint, more light in the face.

Same character, two AI systems.

Left: the reference that's anchored 21 episodes. Right: the new engine I'm testing, warmer skin, softer paint, more light in the face.

The style suffix gets rewritten once, on purpose, then locks again for the next 20. So before I lock it: left or right?

(And yes, still fighting the new engine for her amber eyes.)

[Wendeline @wjcnyc]

EP01 never got animated. Back then I didn't trust the motion tools with a face, so this frame shipped as a still and the series was built on that rule.

Last night I ran it through the current pipeline. First try, no drift, syringe steady, 21 episodes of workflow between the image and the motion.

The tools didn't just get better. The list of things I know not to ask them for got longer. That list is the pipeline.

[Wendeline @wjcnyc]

Episode 22 is on the bench. Same deal as the last 21: one person, one locked workflow, a new episode every week.

The scariest thing in this pipeline is still the render queue.

Now the image, because it's worth talking about. It's the same shop as the fitting-room ghost frame, the red tang jacket folded on the counter where the gray arm was handing it back, the shopkeeper reaching up, the old brass register. As a story beat it's lovely: the "before" of that haunting.

[Wendeline @wjcnyc]

Today the pipeline fought back. Faces drifting on a locked prompt, regenerate, compare, kill, repeat.

This frame is one of the casualties. Gorgeous, first try, and it still got cut because the shot around it wouldn't hold. That's the part nobody sells you: the folder of good images that didn't survive being a sequence.

40 calls a day about which flaws ship. Some days the employee of the month is just me.

[Wendeline @wjcnyc]

Two months ago I spent an evening on a trend edit. It did its job on Instagram and never went up anywhere else. The audio was Dua Lipa, and rights don't care how good your edit is.

Not an episode, just a one-evening experiment with the same pipeline. Lesson kept: never build on sound you don't own. The Jester survives, the track doesn't.

Same story, honest details, and "one evening, same pipeline" is quietly a better flex anyway.

[Wendeline @wjcnyc]

21 episodes in. Some honest numbers from running a weekly AI horror series completely alone:

One person. About a day of work per episode. Zero crew, zero payroll, zero freelancers. Roughly 40 shots per episode, each one a still first, and only the shots that need motion get animated. One style suffix locked since January so 21 episodes look like one show instead of a pile of clips.

The pipeline is the employee. It doesn't call in sick, it doesn't know it's a long weekend, and EP21 ships anyway.

I spent 20 years in NYC property management before this. Same lesson in both jobs: systems beat effort. If your business needs video every single week without hiring a video team, that's exactly the thing I build. DMs open.

[Wendeline @wjcnyc]

Tried to make this thing talk. It refused. The smoke moves, the face won't.

Lip sync needs a clean, well-lit mouth, and this one's buried in smoke. That's the trade with every shot: the style that makes it look good is the same thing that breaks the next tool in the chain.

He stays silent. Silence is scarier anyway.

[Wendeline @wjcnyc]

Not every shot is a monster. EP02, 4 AM outside the diner, the shot that made her a character instead of a prop.

One person, one pipeline, a new episode every week. Coffee's the only crew.

[Wendeline @wjcnyc]

Everyone asks how the series looks like one show instead of 20 random clips.

Four frames, four different episodes, weeks apart. The trick isn't a fancy model. It's one locked style suffix, pasted on every single prompt since EP01, and never touching it.

Three of these episodes are out. The fourth is from one that isn't. Same look before the episode even exists; that's the point of locking it early. Same rule when I build one of these for a client: the look comes first.

[Wendeline @wjcnyc]

@AIBizarrothe @Lang_Enthused The unglamorous part: mine is 40+ words long and hasn't changed since January. The urge to "improve" it is the trap.

[Wendeline @wjcnyc]

@kaanakz @Kling_ai This is the most useful reply I've gotten on here. The masked-denoise idea maps exactly to my drift problem. I'm in the hosted-tool lane, not ComfyUI, but "change less per generation" is the transferable principle. Appreciate you writing it out.

[Wendeline @wjcnyc]

@Sparrow8210 The last-3-seconds bug is almost universal with Kling. It commits to physics early and the errors compound at the end. My fix: generate 5s when I need 3 and always cut before the wheels come off. The trim is part of the workflow.

[Wendeline @wjcnyc]

@kaanakz @Kling_ai Same, though I’ve had better luck locking the still first and animating from that — the character lives in the image, not the prompt. Where it still breaks for me is a hard angle change. Wardrobe drifts every time. Does yours hold on profile turns?

[Wendeline @wjcnyc]

Nobody posts the folder of shots that didn't survive. Here's mine from this week.

An elevator with two 6s, two 5s, and no 13 button, under a glowing 13. Grand Central with the same woman haunting both clock faces and a schedule board in a language that doesn't exist.

Killing a shot fast is most of the skill. The episode that ships next week looks nothing like the one I started.

[Wendeline @wjcnyc]

@kaanakz @Kling_ai Same method here, hero reference plus a locked style suffix on every prompt. Where it still breaks for me is faces across scene changes; the reference holds the outfit but drifts the jaw. If you've cracked that one I'm all ears.

[Wendeline @wjcnyc]

Yesterday: a shot that stayed a still.

This morning I animated this one. Same workflow, one more step: Gemini still in, Kling does the capture. The ghost going into the jar is the shot.

Same question every time: does the frame need time to work? This one does. I make that call ~40 times per episode.

[Wendeline @wjcnyc]

@AIInfluencerHQ This list is why I skipped LoRA training entirely. Locked style suffix plus a hero reference image gets 90% of the consistency with none of the training loop. Slower models, faster pipeline.

[Wendeline @wjcnyc]

@TheRealSiesta @TheoMediaAI Speed's solved. Consistency isn't. I ship a weekly series and the bottleneck was never render time, it's getting the same face and the same room 40 shots in a row.