Your prompt asked for a sign that reads OPEN, or a phone screen showing MESSAGE. What came back is hieroglyphic letterforms that jitter across frames, spelled differently in almost every frame. This is the single most common AI-video defect, and it is not a prompt mistake: most video models treat letters as texture, not language, so they paint something that looks like writing without ever “knowing” the word.
Fastest fix (works every time): generate the clip with a blank surface where the text should go, then add the text in post with CapCut, Premiere Pro, or DaVinci Resolve. Compositing real text on top is the only method that is 100% legible and on-brand.
If you must keep text in-frame: regenerate on a text-capable model. As of June 2026 the order is roughly Veo 3.1 (sharpest single-frame text) and Kling 3.0 (most consistent short text across frames), then Sora 2; everything else mangles it.
Which bucket are you in?
Match your symptom to find the cause and the right step below.
| Symptom | Most likely cause | Go to |
|---|---|---|
Letters look like writing but spell nonsense (H3LL0) | Model renders text as texture, not language | Step 1 or 2 |
| Frame 1 is correct, later frames wobble/respell | Per-frame regeneration, no object tracking | Step 1 (Kling/Veo) or Step 2 |
| Big headline OK, small labels garbled | Text too small (too few pixels) | Step 2 or 3 |
| One sign fine, a street full of signs is a mess | Too many text regions at once | Step 2 |
| Plain text works, “neon cursive” doesn’t | Stylized font pushed past model limits | Step 1 with plain font, then Step 2 |
| Already have the bad clip, can’t re-render | N/A | Step 4 |
Common causes
Ordered by how often they are the real problem.
1. Model represents text as texture, not language
Most video models — Runway, Pika, Luma, Hailuo, and older Kling/Veo builds — were trained to render visual scenes, not to spell. They produce shapes that read as “text-like” but carry no actual word.
How to spot it: generate any clip with the word HELLO on a sign. If the output reads H3LL0, HEILO, or changes every frame, the model cannot do reliable text.
2. Text exists per-frame but not across frames
Even on a model that gets frame 1 right, each frame is regenerated independently. By frame 30 the same word has different kerning, color, or shape because nothing is tracking it as a fixed object. Kling 3.0 and Veo 3.1 are the most stable here as of June 2026, but none are perfect.
How to spot it: scrub to frame 1 and frame 30 and compare. If the text “breathes” or respells, the model is regenerating it rather than tracking a static object.
3. Text is small in the frame
Smaller text means fewer pixels, which means less capacity to spell correctly. A big headline on a wall is far easier than a button label on a UI.
How to spot it: estimate the text height in pixels. Under about 40 px tall almost always garbles; over ~200 px tall, the better models can sometimes manage.
4. Multiple text instances in one clip
A street scene with three signs, two posters, and a license plate asks the model to spell in five places at once. Expect it to fail in at least four.
How to spot it: count distinct text regions. More than one and you should plan to add them in post.
5. Stylized fonts requested
“Cursive neon,” “graffiti tag,” “1920s film title card” all push text into a stylized space where even Veo 3.1 and Sora 2 slip.
How to spot it: reduce the prompt to plain uppercase sans-serif. If plain works and stylized doesn’t, the style was the problem.
Shortest path to fix
Step 1: Regenerate on a text-capable model (if re-rendering is feasible)
As of June 2026, three models handle in-frame text well enough to ship for short words and signage. Keep the text to 1-3 words, large, and the only text in the scene.
# Veo 3.1 (Google) — sharpest single-frame text, 4K, native audio
- Best for crisp signage and short slogans ("SUMMER SALE 50% OFF" stays legible)
- 8-second clips; chain generations for longer sequences
- Prompt:
"A wooden shop sign with the text OPEN in clear black block letters,
sharp focus, no other text anywhere in the scene."
# Kling 3.0 — most CONSISTENT short text across frames
- Best when the camera moves and you need 1-3 words to hold steady the whole clip
- Multi-shot storyboard mode keeps text stable across cuts
- Prompt:
"Close-up of a red neon DINER sign glowing steadily, no flicker,
no other text in the frame."
# Sora 2 / Sora 2 Pro (OpenAI) — strong for text on physical objects
- Good for text that sits on a sign or screen as a real object
- Storyboard tool lets you specify exact text per shot
- OpenAI has been shifting Sora 2 access and pricing repeatedly in 2026, so check current availability before you plan a job around it
- Prompt:
"Vintage diner sign with the word DINER in red neon, glowing steadily,
no flicker, no other text."Prompt rules that materially help: spell the word in CAPS, say the text "WORD", add no other text in scene (every extra word region competes for the model’s spelling budget), and keep the typeface plain.
Step 2: Generate without text, then add it in post (production-grade)
This is the reliable, repeatable method. Generate the clip with a blank surface where text belongs, then composite real text on top.
# Prompt the AI video for a blank surface
"A wooden shop sign hanging from a chain, blank surface, no text, no markings,
clean weathered wood ready for signage."
# CapCut (free, desktop / web / mobile)
- Text -> Add text -> place over the blank sign
- Animate position with keyframes to track the sign's apparent motion
- For a tilted surface, use the 3D rotation controls to match perspective
- Export
# Premiere Pro
- Window -> Essential Graphics -> New Layer -> Text
- Track with manual keyframes, or use Mocha (Boris FX) for planar tracking
- Apply Drop Shadow to ground the text into the scene
# DaVinci Resolve (Fusion page)
- Text+ node for the wording
- Tracker node bound to the surface, connect to the Text+ transform
- Merge the Text+ over the AI footageStep 3: Static graphic overlay for fixed-camera shots
If the AI clip has no camera motion and the text region is static, drop a PNG on top.
# Create the text in Photoshop, Figma, or Affinity Designer
- Match the implied lighting and color temperature of the scene
- Add slight noise / grain to match the footage's sensor response
- Export as PNG with transparency
# Composite in any editor
- Place the PNG on a track above the AI video
- Align it to the intended surface
- Add a subtle 5-10% opacity grain layer over both to unify the lookStep 4: Mask and patch garbled text in post (when you can’t re-render)
You already have the clip with bad text and regenerating is off the table.
# DaVinci Resolve (Fusion)
- Mask the garbled text region
- Patch Replacer node samples adjacent clean surface to cover it
- Add a Text+ node with the correct word on top, tracked to the surface
# Adobe After Effects
- Mask the garbled region -> Edit -> Content-Aware Fill to remove it
- Add a Text layer with the correct wording
- Track with Mocha to match the original motion
# CapCut (Pro)
- Remove objects -> brush over the garbled text to erase it
- Add a new Text element with the desired wording
- Position it over the cleaned regionStep 5: Last resort — accept stylized illegibility
For purely decorative background signage (alley shots, distant billboards), illegible text reads as “atmospheric foreign language” and audiences accept it. Only do this when the text carries no information.
# Decision rule:
- Hero text (logo, headline, dialogue card)? Fix with Step 2 or 3.
- Background atmosphere text (neon, distant posters)? Acceptable as-is.
- UI / button text (phone screens, monitors)? Always add in post.How to confirm it’s fixed
- Scrub frame-by-frame across the whole clip (not just the first frame). The word must spell identically and hold position in every frame.
- Watch at 100% / full resolution, not a thumbnail — garbling often hides in scaled-down previews.
- Export and play the final file. Compression can smear thin lettering; if it does, increase text size or stroke weight and re-export.
- For composited text, confirm it tracks the surface with no slip or float when the camera or object moves.
Prevention
- Plan video and graphics as separate layers from the very start.
- Generate AI clips with deliberately blank surfaces reserved for text overlay.
- Build a library of branded text templates in your editor so post is fast.
- Reserve Veo 3.1 / Kling 3.0 / Sora 2 for shots where in-frame text is unavoidable.
- For UI / app-demo videos, use real screen recordings or mockups composited in post — never ask the model to render a working interface.
FAQ
Which AI video model is best at text in June 2026? For a single sharp frame, Veo 3.1. For 1-3 words that stay consistent while the camera moves, Kling 3.0. Sora 2 is strong when text sits on a real object (a sign or a screen). For anything longer than a few words, add it in post — no model is reliable there yet.
Why does the text change spelling every frame? Most models regenerate each frame independently and have no concept of a persistent “word” object, so the lettering drifts. This is per-frame regeneration, not a bad seed. Either move to a model with better temporal stability (Kling 3.0 / Veo 3.1) or composite the text in post so it is a fixed layer.
Can I fix garbled text without re-rendering the clip? Yes. Mask the bad region and remove it with After Effects Content-Aware Fill, DaVinci Resolve’s Patch Replacer, or CapCut’s Remove objects, then place a correct text layer on top and track it to the surface. See Step 4.
Why does small text garble but big headlines come out fine? Smaller text gets fewer pixels, and below roughly 40 px tall the model lacks the resolution to spell. Make the text larger in the shot, or add it in post where pixel budget is irrelevant.
Does prompting the text "OPEN" actually help?
It helps marginally on text-capable models: caps, quoting the exact word, and adding no other text in scene reduce competing regions. It will not rescue a model that renders text as texture — for those, post is the only fix.