F
Frank Houbre
Guest
The first title card I generated for Lost Garden came back covered in letters that weren’t in any alphabet. Not a typo, not a font issue. The model had understood I wanted text and then drawn something that was confidently, beautifully, not a word. I regenerated it a few more times. One pass fixed a line and broke another. Eventually I gave up and did what I should have done from the start: I made the card somewhere else.
If you want readable text in an AI-generated film, the answer is almost never to get the video model to write it. You generate the image without any words in it, then add the real text as a layer afterward, in an actual editor. That one sentence is the whole workaround, and once you know it, it changes how you plan every title card, sign and on-screen note in the project.
This is not a complaint about the models. They were never trained to spell, they were trained to make convincing pixels, and convincing pixels and correct letterforms are two different skills that happen to look similar from a distance. Knowing that changes what you ask them to do.
A video diffusion model predicts what a frame should plausibly look like, not what a sentence should correctly say. Letters are small, dense, high-frequency shapes, and the model has no concept of spelling, it only has a concept of “text-shaped marks in this kind of composition.” That’s why you get something that reads as a sign from across the street and turns into static the moment you get close enough to try reading it.
Static image models handle targeted edits far better than video models handle text in motion. OpenAI’s ChatGPT Images 2.5 launched in September 2026, was built specifically around precise edits, including swapping out a single element such as a product, a background, or a piece of text, without touching the rest of the image. That’s a different job from asking a model to invent correct, legible lettering inside a moving shot, and it’s the whole reason the two-stage workaround below exists.
Add them in post, almost always. The practical reason is not aesthetic, it’s operational: baked-in text can’t be revised without regenerating the whole clip, and a regeneration costs credits and time. A typo in a title card you generated is a full re-render. A typo in a text layer sitting on top of the footage in your editor is a ten-second fix.
There’s one real exception. If the text has to feel like it’s physically part of the world, not sitting on top of it, carrying the same grain, the same light, the same slight imperfection as everything around it, generating it in-world can be worth the extra passes. That’s a deliberate choice for a handful of shots, not your default for every title card and every episode number.
On Lost Garden I now split it cleanly:
This is the harder case: a carved sign, a note someone leaves on a door, a name on a screen in the background. You can’t just slap a clean font over it, it has to look like it belongs to the shot.
The two-stage method that actually holds up:
A few things make or break this, and I learned most of them the slow way. Keep the camera slow near text: a static hold or a gentle push-in survives the animation pass, while a whip pan or fast handheld motion smears the lettering into mush within a few frames, even starting from a perfect still. Check the still before you spend any video credits on it, since image generations are cheaper than video generations by a wide margin on every platform I’ve used, so the discipline is simple: don’t move to video until the image is actually right. And once a sign or a logo reads correctly, save that exact file. Reusing a locked asset is what keeps the same prop legible across multiple shots, instead of regenerating a slightly different, slightly wrong version every time it appears.
This is the same discipline I ended up applying to character reference sheets on Lost Garden: lock the thing that has to stay identical, and stop asking the generator to reinvent it from scratch every single time.
Here’s the version I run now, after enough wasted renders to trust it:
One of the more useful habits I picked up while building ScreenWeaver’s storyboard stage was planning exactly what text needs to appear, and where, on paper, before a single frame gets generated. It cuts the number of “wait, what does that say” surprises down to almost nothing. None of this is exotic, really. It’s closer to how title design has always worked in post-production. The surprising part is that AI video forces you to rediscover the same boundary between what the camera captures and what the editor adds that film editors have respected for a century, just for a completely different reason.
Can any AI video model generate readable text directly?
Results vary and keep shifting month to month, but as of 2026, text accuracy is still inconsistent enough across general-purpose video and image models that the safer default is generating text separately and compositing it, rather than trusting any single model to spell correctly on the first try.
Does this slow down production?
It adds one extra step for text-heavy shots, but it removes the much bigger cost of regenerating a whole clip over a spelling error. Net, it’s faster.
What if the text only needs to be implied, not read?
Then don’t fight this problem at all. A blurred sign, a screen angled away from camera, or text that’s legible as “there’s writing here” without needing to be read word for word, skips the whole workaround. Save the two-stage method for text the story actually needs the audience to read.
Is this specific to anime-style projects like Lost Garden?
No. The failure mode is about how diffusion models handle small dense shapes, not about art style. Live-action AI shots hit the same wall with street signs, screens and handwriting.
If you’re planning a script or storyboard and want to flag exactly where on-screen text needs to appear before you generate a single shot, that’s the kind of pre-production step ScreenWeaver is built to catch early, alongside the rest of the shot breakdown. And if you want to see how this plays out across a full six-character episode, Lost Garden is the project all of this gets tested on first.
If you want readable text in an AI-generated film, the answer is almost never to get the video model to write it. You generate the image without any words in it, then add the real text as a layer afterward, in an actual editor. That one sentence is the whole workaround, and once you know it, it changes how you plan every title card, sign and on-screen note in the project.
This is not a complaint about the models. They were never trained to spell, they were trained to make convincing pixels, and convincing pixels and correct letterforms are two different skills that happen to look similar from a distance. Knowing that changes what you ask them to do.
Why can’t AI video models render readable text?
A video diffusion model predicts what a frame should plausibly look like, not what a sentence should correctly say. Letters are small, dense, high-frequency shapes, and the model has no concept of spelling, it only has a concept of “text-shaped marks in this kind of composition.” That’s why you get something that reads as a sign from across the street and turns into static the moment you get close enough to try reading it.
Static image models handle targeted edits far better than video models handle text in motion. OpenAI’s ChatGPT Images 2.5 launched in September 2026, was built specifically around precise edits, including swapping out a single element such as a product, a background, or a piece of text, without touching the rest of the image. That’s a different job from asking a model to invent correct, legible lettering inside a moving shot, and it’s the whole reason the two-stage workaround below exists.
Should title cards live inside the AI tool or get added in post?
Add them in post, almost always. The practical reason is not aesthetic, it’s operational: baked-in text can’t be revised without regenerating the whole clip, and a regeneration costs credits and time. A typo in a title card you generated is a full re-render. A typo in a text layer sitting on top of the footage in your editor is a ten-second fix.
There’s one real exception. If the text has to feel like it’s physically part of the world, not sitting on top of it, carrying the same grain, the same light, the same slight imperfection as everything around it, generating it in-world can be worth the extra passes. That’s a deliberate choice for a handful of shots, not your default for every title card and every episode number.
On Lost Garden I now split it cleanly:
- Opening titles and episode cards get built entirely in post, as clean typography over a generated or composited background, so a word change never touches a rendered clip.
- In-world text that has to read as part of the set goes through the two-stage method below, and only where the story genuinely needs the audience to read something inside the frame.
How do you add in-world text that doesn’t look fake?
This is the harder case: a carved sign, a note someone leaves on a door, a name on a screen in the background. You can’t just slap a clean font over it, it has to look like it belongs to the shot.
The two-stage method that actually holds up:
- Generate the text as a still image first, using a model known for clean typography rather than your general scene generator. Write the exact wording in quotes in the prompt and describe the physical object carrying it: “a carved wooden sign reading ‘TURN BACK’, weathered paint, mounted at the treeline.” Generate it large, ideally 2K or higher, because compression eats detail you don’t have spare.
- Animate the approved still instead of generating fresh video with the text baked in. Once the lettering is correct in a frame you’ve checked with your own eyes, feed that still into an image-to-video pass on a tool like Runway. You’re asking the model to add motion to something that’s already right, not asking it to invent correct spelling from a text prompt, which was the failure mode you were trying to avoid.
A few things make or break this, and I learned most of them the slow way. Keep the camera slow near text: a static hold or a gentle push-in survives the animation pass, while a whip pan or fast handheld motion smears the lettering into mush within a few frames, even starting from a perfect still. Check the still before you spend any video credits on it, since image generations are cheaper than video generations by a wide margin on every platform I’ve used, so the discipline is simple: don’t move to video until the image is actually right. And once a sign or a logo reads correctly, save that exact file. Reusing a locked asset is what keeps the same prop legible across multiple shots, instead of regenerating a slightly different, slightly wrong version every time it appears.
This is the same discipline I ended up applying to character reference sheets on Lost Garden: lock the thing that has to stay identical, and stop asking the generator to reinvent it from scratch every single time.
What’s the actual workflow, start to finish?
Here’s the version I run now, after enough wasted renders to trust it:
- Plan the text before you generate anything. Decide in the script or shot breakdown which text is informational (title cards, episode numbers, subtitles) and which is diegetic (a sign the camera sees inside the story), since that single decision sends each piece down a different path.
- Informational text goes to post, no exceptions, because it needs to stay editable for as long as the project exists.
- Diegetic text goes through the two-stage method: still image first, slow camera on the animation pass, locked asset once it’s approved.
- Don’t trust the first result, whether it’s a title card typo or a garbled sign, and check it before it goes anywhere near the timeline.
- Budget for it. Text-heavy shots cost more passes than text-free ones, and if a scene doesn’t strictly need on-screen words, it’s often worth finding another way to convey the same information.
One of the more useful habits I picked up while building ScreenWeaver’s storyboard stage was planning exactly what text needs to appear, and where, on paper, before a single frame gets generated. It cuts the number of “wait, what does that say” surprises down to almost nothing. None of this is exotic, really. It’s closer to how title design has always worked in post-production. The surprising part is that AI video forces you to rediscover the same boundary between what the camera captures and what the editor adds that film editors have respected for a century, just for a completely different reason.
FAQ
Can any AI video model generate readable text directly?
Results vary and keep shifting month to month, but as of 2026, text accuracy is still inconsistent enough across general-purpose video and image models that the safer default is generating text separately and compositing it, rather than trusting any single model to spell correctly on the first try.
Does this slow down production?
It adds one extra step for text-heavy shots, but it removes the much bigger cost of regenerating a whole clip over a spelling error. Net, it’s faster.
What if the text only needs to be implied, not read?
Then don’t fight this problem at all. A blurred sign, a screen angled away from camera, or text that’s legible as “there’s writing here” without needing to be read word for word, skips the whole workaround. Save the two-stage method for text the story actually needs the audience to read.
Is this specific to anime-style projects like Lost Garden?
No. The failure mode is about how diffusion models handle small dense shapes, not about art style. Live-action AI shots hit the same wall with street signs, screens and handwriting.
If you’re planning a script or storyboard and want to flag exactly where on-screen text needs to appear before you generate a single shot, that’s the kind of pre-production step ScreenWeaver is built to catch early, alongside the rest of the shot breakdown. And if you want to see how this plays out across a full six-character episode, Lost Garden is the project all of this gets tested on first.