Ideogram

I'm so excited to share that we are launching p-image-ideogram! Co-built with our friends Ideogram The model starts at 0.003$ per image generated and sits comfortably as the obvious choice on the cost and latency pareto frontiers 🚀 More info here 👇

Today we're introducing P-Image-Ideogram: a family of Pareto-optimal image models with the best quality-speed-cost trade-off, co-developed with Pruna AI. Any model rated higher in quality costs more or runs slower. This launch is possible because of a decision we made in June: releasing Ideogram 4.0 with open weights. Open weights let partners go deeper than just an API, and Pruna AI took the 4.0 weights and, working closely with our team, optimized them for entirely new speed-cost operating points. What shipped today: - Four quality modes (Very Low, Low, Medium, High), native 1K and 2K generation - From $0.003 per image and latency from 3s - On the Pareto frontier for both preference vs. price and preference vs. speed on Design Arena's blind, head-to-head leaderboard Live now on the API and across partner platforms including ComfyUI, Replicate, Cloudflare, Together AI, Leonardo.Ai, and more. We hope this enables broader access to frontier-quality image generation and shows what an open model ecosystem makes possible. Try it here: https://lnkd.in/gXNWkVcp

Today, we’re introducing Ideogram Object Remover as the state-of-the-art object removal method. Removing an object from an image is not too hard, but removing every trace of it is challenging. Most tools erase the object but leave some evidence behind: a shadow on the pavement, a reflection in the glass, or a subtle shift in other pixels that’s annoying for professional users. An important piece of the problem is deciding which pixels to change. We believe a user-provided mask remains the best interface because it keeps the user in control of what needs to be removed. That said, the mask expresses the intent, and it is not a hard constraint on the edit. The user should not have to trace and mask shadows, reflections, or lighting changes. We’ve worked hard to enable the Object Remover to identify and remove the object's full footprint even outside of the mask. Importantly, the model should be able to see through the mask. Masks are rarely pixel-perfect, so a loose selection should not destroy the background it overlaps. Transparent objects like glass carry background information the model should use rather than discard. And text is the extreme case: a practical text selection covers the holes inside letters and the spaces between them, and all of those pixels in the background are worth preserving. We use diffusion models, the state of the art at modeling natural image priors, to fill the removed region with sensible material. But we designed the generative component to be conservative, because the goal of object removal is to restore the scene rather than to imagine new content. The model reconstructs what is plausibly behind the object and leaves everything outside the removal untouched: nothing invented, nothing else changed. After careful evaluation and several iterations, on RemovalBench, Ideogram Object Remover outperforms leading image editing models, including Nano Banana 2, FLUX Erase, and GPT Image-2. Not only does it rank #1 on removal quality, it also offers the lowest price per request. Ideogram Object Remover is available today at https://lnkd.in/gawTGb6M and through the Ideogram API. You can try it for free and integrate it into your workflow or product. Congrats Steven Caro and Gian Favero on the launch!

Ideogram just released the #1 open weight image model in the world. It punches way above its weight - it's small enough to run on a consumer GPU, but goes head-to-head with Nano Banana and GPT Image for design. Yoko Li and I went deep with founder and CEO Mohammad Norouzi on how they trained it and his hot takes on image research. My summary: 1) LLMs should act as “compilers” for diffusion models. Ideogram treats image generation as a two-step pipeline. Users often give a vague natural-language prompt (e.g. "the meaning of life") — the diffusion model shouldn't be the thing deciding what that means. Instead an LLM does the reasoning, expanding the prompt into a long, highly structured JSON schema that plots out exact coordinates, color palettes, and every element in the scene. The diffusion model is then just a pure execution engine that renders the spec. 2) Heavy RL may hurt artistic diversity. Top frontier models often default to the same recognizable, generic "AI look" because they rely too heavily on reinforcement learning. Ideogram deliberately used very little RL on this model. The trade-off is that their "raw" model requires much more precise prompting, but it preserves a vastly wider range of artistic and minimalist styles that heavily RL-tuned models naturally collapse away from. 3) Training data can be structured for better control in generations. The standard industry practice is to use a VLM to replace junky HTML alt-text with dense synthetic captions. Ideogram does that too, but the captions go further: each one is structured JSON that lays out every element in the image with bounding boxes (in normalized 0–1000 coordinates), per-element color palettes, and typed text fields. Because the same structured format is used at training and inference, layout, color, and text placement become things you can specify directly, which is what makes the precise design control possible. 4) You can’t train a highly capable model on “narrow” data. If you don’t have a massive amount of compute and are trying to compete in the image model wars, it’s tempting to try to narrow your scope. But a quality image model requires a general, foundational understanding of the physical world first — it can then be fine-tuned or customized for specific narrow use cases like illustration or typography. 5) Agents scale exploration, but the last 10% requires a canvas. With the rise of autonomous agents and protocols like MCP, you can hit an API and generate thousands of design iterations in an hour. But chatbots and pure agentic loops fail at the final mile. Agents are perfect for high-level creative exploration, but closing the final 10% of a design inherently requires a spatial UI canvas where a human can point, click, and edit specific regions.