You paste a screenshot of a broken UI into Antigravity's chat, hit send, and the AI replies with a text-only answer that never once mentions the image. It's one of the more disorienting failure modes, because nothing looks broken — the thumbnail is right there in the chat window, and the conversation appears to be flowing normally. I hit this myself when I first dropped a Figma frame into a chat asking for a layout tweak, and the AI opened its reply with "While I couldn't see the image…" That sentence is the giveaway, and it has quietly wasted thousands of tokens for countless developers.
The key insight is this: an image appearing in chat is not the same as the model actually reading it. The image can be visually attached to your message yet silently ignored by the model you're talking to. Once you internalize that gap, the troubleshooting becomes systematic rather than frustrating. Here's the six-point checklist I run through whenever this happens, in the order I actually check them.
Start here — is the selected model even multimodal?
Before blaming the paste path or the image itself, look at which model is currently active. Antigravity lets you swap models freely, and selecting a text-only model means your image gets attached but never reaches the vision pipeline. That thumbnail you see? It was never decoded.
Check the model selector at the bottom of the chat. Gemini 3 Pro and Gemini 2.5 Pro will happily read images. Local LLMs routed through LM Link — for example a text-only Gemma 2 2B build, or older quantized models without vision support — and many lightweight models won't. I usually keep Gemini 3 Pro as my default, but I've burned plenty of sessions by switching to a Flash variant for token economy and then forgetting to switch back when I needed vision. If you're doing anything image-heavy, consciously verify the model before your first paste. A good habit is to name your chat sessions — "UI review with Pro" vs. "quick refactor with Flash" — so the intent stays visible in the sidebar.
Is the image format and size in the supported range?
Next, consider the image itself. The Gemini API underneath Antigravity supports PNG, JPEG, WebP, HEIC, and HEIF, but real-world workflows expose some gotchas that the spec sheet doesn't advertise:
- HEIC handling: Screenshots and photos from iPhones often land on macOS as HEIC. When pasted into Antigravity, conversion to a web-friendly format isn't always reliable, and the model may receive something it can't decode. Convert to JPEG or PNG first and the problem disappears in most cases. A simple way to do this is to open the HEIC in Preview and use File → Export → JPEG.
- Oversized images: Ultra-wide captures (above ~4K) sometimes fail to upload silently — the request times out or the binary is truncated, and you get no error message. Resize to roughly 1920px wide before pasting. For 4K monitor screenshots, this alone resolves roughly half the "ignored image" reports I've seen.
- Transparent PNGs: A transparent background can be interpreted by the model as empty space surrounding your content. The UI element you care about gets lost in what the model perceives as "mostly nothing." Crop tightly around the meaningful region before sending, or flatten the transparency against a solid background first.
When you want the AI to focus on a specific UI element, crop to that region in Preview (macOS) or Snipping Tool (Windows) first. It removes ambiguity and often produces a sharper, more actionable response. The rule I've come to follow: if a human would struggle to spot the issue in the image, so will the model.
Clipboard paste, drag-and-drop, or the attachment button?
How you get the image into the chat matters more than you might expect. Antigravity accepts three paths, and each has different reliability characteristics:
- Clipboard paste (⌘V / Ctrl+V) — fastest, but depends entirely on the OS clipboard state and whatever format a source application decided to put there
- Drag-and-drop — drop an image file directly into the chat area; bypasses the clipboard entirely
- Attachment button (the clip icon) — slowest but the most reliable; it reads the file from disk in a known format
When a clipboard paste fails to trigger vision, the clipboard often holds a file reference, a screenshot placeholder, or a converted format rather than raw image bytes. This is especially common on Windows when a capture tool like ShareX or Snagit is in the pipeline — those tools frequently rewrite the clipboard after you copy. If paste isn't working, save the image to disk and drag-and-drop it or use the attachment button instead. That single switch resolves the issue the majority of the time in my experience.
OS-specific gotchas worth knowing
Antigravity behaves slightly differently depending on your operating system, and these platform-specific details catch people who've moved between machines:
- macOS: Check the screenshot format with
defaults read com.apple.screencapture type. If it readsheic, switch it withdefaults write com.apple.screencapture type png; killall SystemUIServerfor fewer surprises across the entire chain - Windows: Make sure a clipboard manager like Microsoft PowerToys isn't intercepting the paste. PowerToys "Advanced Paste" can silently change the payload Antigravity receives. Temporarily disable it and retry to confirm whether it's the culprit
- Linux / WSL2: X11 vs. Wayland differences can drop the image portion of a clipboard payload, leaving only text fragments. On Wayland, prefer the native build and the attachment button over clipboard paste. Under WSL2, always save to disk first — clipboard sharing across the Windows/Linux boundary is unreliable for binary content
Confirming the AI actually sees your image
Here's the sanity-check prompt I reach for whenever I'm unsure:
List the top three dominant colors in this image in hex.
If you cannot see the image, reply only with: "Image not visible."
If the model returns hex codes, you know vision is live. If it responds with "Image not visible," the issue is on the model side rather than anywhere else in your setup — and you've just cut your search space in half. This prompt separates "the model can't see the image" from "my instructions were too vague," which is the single most common confusion point I see when developers try to diagnose this themselves.
Another fast check: ask the model to read a short piece of visible text in the image — a button label, the first line of an error message, or the app name in a title bar. OCR working is proof that image bytes are flowing through the pipeline and the vision encoder ran. If the text comes back correctly, you can trust that more nuanced visual analysis is also happening.
When all else fails, stop fighting the image
If paste still doesn't work after you've checked everything above, consider skipping the image entirely. It's an option more people should reach for:
- For UI issues, paste the relevant HTML and CSS instead. Structure is often more useful to the model than pixels, especially for layout and accessibility questions where the DOM tells a clearer story than a rendered screenshot
- For error messages, copy the text, not a screenshot. OCR introduces noise that hurts precision, and a clean stack trace or error string routes directly to the model's training on similar errors
- For diagrams or design mocks, upload to Gyazo or Imgur and paste the public URL. When Antigravity can fetch the image server-side, it's often more stable than clipboard transfer, and you get a shareable reference at the same time
My personal heuristic: if a piece of information can reasonably be expressed as text, text will usually produce a faster and more accurate reply. Treat screenshots as the last resort, not the first. The best AI conversations I've had about code rarely involved screenshots at all.
Next step
Next time a paste gets ignored, switch to Gemini 3 Pro and run the three-color prompt above. Within thirty seconds, you'll know whether the model is the problem or the paste path is. That single experiment narrows the fix dramatically and saves you from running through the rest of this checklist blind.
If you want to go deeper on multimodal workflows, the Antigravity × Gemma 4 multimodal implementation guide and the LM Studio connection troubleshooting guide pair well with this article and round out the picture.