Something about how the Antigravity models analyze images makes them quickly become bloated and slow, becoming unusable for any project that needs images and constant tweaking. Any project involving a back-and-forth analysis of more than 5 images over 3 turns will become a slog. Perhaps images could be pruned from the context window after a few turns, as they are probably no longer relevant. Any needed context would then be gathered by the agent by analyzing the image again of its own volition. I believe the images are being fed in a very token-inefficient format, such as Base64. I’m not sure, but there’s definitely something in there making this a problem for trying to incorporate images into my projects.
Have anybody else experienced that? Like making a game for example, and asking for Antigravity to handle the visuals. As soon as you ask it to not only get the images, but then analyze them to check quality or other stuff, it becomes very slow very fast. I’ve been asking it to use sub-agents to generate, and analyze the images and then kill them and refresh new ones, but that keeps making the subsequent batches not match the previously established rules/style. it’s very frustrating.
Hi @Luiz_Gabriel_dos_San,
Yeah, images eat up ton of tokens, so analyzing them over multiple turns bloats the context window fast — that’s what’s causing the slowdown. Your sub-agent idea is the right call, the problem is just that they lose your style rules when you kill them.
Here’s what I’d do for your game asset workflow:
- Save your rules as a Skill — drop your art rules (colors, style, etc.) into
.agents/skills/style-guide/SKILL.md. Every new sub-agent reads this automatically, so the rules stick even after a restart.
- Write results to files, not chat — have sub-agents dump their analysis into something like
review.md instead of replying in chat. Keeps the image data out of your main context.
- Shrink images first — resize to ~512px before the quality check. That’s more than enough for style checks but way cheaper on tokens.
So your loop becomes: sub-agent spins up → reads SKILL.md → analyzes the resized image → writes results to review.md → done. No bloat, no style drift.
One quick question - which model are you using for this task?
Yeah, i’ve been using skills, and i might try to instruct them also via skill to write the results in a .md file. But the problem is. For example, they are generating the images, and i’m telling them to audit the images themselves. So they have to see the final result which might have started as a small image but then it gets resized bigger later. I might see if it is possible to tell them to audit the image before running the hires fix.
And i’m currently using Gemini 3.8 Flash (Low). I was using Gemini 3.7 Flash (Low) before, but the same problem was already there too.