In partnership with
Hey Creator,
If you've ever redone the same image three times because one small edit broke everything else, you're not alone.
OpenAI just released a fix for it. Let’s check it out!
Why Your Next Thumbnail Might Come From a Doodle
You know that moment when you can see the visual in your head, but describing it in a prompt turns it into something totally different? OpenAI's latest image release is built for exactly that frustration.
ChatGPT Images 2.5 went live on September 8, and the focus this time isn't flashier images — it's control. Generation runs up to 50% faster than before, which you'll actually feel when you're cycling through five thumbnail options instead of waiting one out at a time.
The real fix, though, is how edits behave now. If you've ever asked for one small tweak and watched the model quietly redo half the image, that's the exact problem this version targets — it changes only what you ask for and leaves everything else alone, even a few rounds into a conversation.
Two additions make this genuinely useful for non-designers:
Sketch — type "@Sketch," draw your rough idea by hand, and let that guide the final image. Great for layouts you can picture but can't quite put into words.
Comment edits — tap a spot on the image, drop a note like "make this darker," and only that part updates.
Where this earns a spot in your workflow: quick social visuals, thumbnails, product mockups, early concepts before a final pass in your design tool.
Where it still falls short: dense text, tightly structured layouts, anything needing exact brand precision.
It's rolling out now to everyone on ChatGPT, ChatGPT Work, and Codex — no extra signup.
If you build with the API, you get two models to choose from: Flare for speed, Sunburst for precision editing, both priced at $30 per million tokens for image output, supporting images up to 3,840px.
For anyone already writing inside ChatGPT, this is the first version that feels like it belongs in the same workflow — not a separate tool you have to remember to open.
Is Your Training Data Actually Model-Ready?
If you're fine-tuning a speech model, you've probably hit this wall: DNSMOS gives you a score, but it doesn't tell you whether the data behind that score is actually right for your model.
Treat it as a pass/fail gate and you'll end up training on audio that looks clean on paper but drags down real-world performance—while good source data gets tossed for no reason.
Voices' CTO DJ Jalali (with the team's senior audio and voice data engineers) just published a free white paper that breaks down the four-step calibration framework they use internally to set model-specific quality thresholds instead of trusting the raw DNSMOS number. It also covers where DNSMOS breaks down and how Voices validates audio for custom datasets at scale.
