Building the Future of Image Generation with Ideogram's CEO

Key points

• The headline build: a 9.3B-parameter open-weights model, ~9× smaller than the ~80B 'SOTA', that runs on a single/consumer GPU and reaches the level of 'Nanobanana or GPT Image' on text and photorealism — chosen because 'there's still so much to do' on innovation rather than scale.
• The explicit wedge — 'we can't win on scaling': as an ex-Googler, Mohamad says even a 10× raise can't beat Google on chips, so Ideogram competes on focus (graphic design / editable text), differentiation, and open weights to be 'at least another option for people who care about design.'
• Taste as a measured moat, not a vibe: 'we really want our models to have taste' = 'going outside of the norm... not conforming to the average opinion, which is a little against being on top of the leaderboard'; he ran deliberately little RL to keep the model a 'raw' multi-style generator instead of the same RL-flattened look frontier models repeat — and evaluates taste with human designers because 'AI is not very good at doing the actual taste evaluation yet.'
• Original differentiation was accurate text rendering when image gen 'was synonymous with garbled text' (DALL·E 2 travel-poster memes) — which unexpectedly was 'the whole graphic design and storytelling industry,' making typography/logo/t-shirt design Ideogram's brand.
• The training recipe: go image→text with detailed bounding-box and per-element descriptions (via VLMs, far richer than the short/inaccurate internet alt-text the early models used), then train text→image backwards — so quality tracks how precisely the data describes every element.
• JSON prompting as an intermediate representation: the model is trained ONLY on ~4,000-token JSON, an LLM expands a vague idea into that structure ('Magic Prompt'), and — unlike OpenAI/Google — Ideogram shows users the actual model input for control and consistency; the representation may shift from custom JSON to HTML since LLMs are already trained on it.
• 'It's not about how good a model is in the general sense. It's about how good is this model for my use case' — the customization thesis: artists with ~15-50 pieces fine-tune to their own style (an artist-in-residence got '3x faster' on a comic book), and enterprises whose brand guidelines generic models miss get custom models that 'understand my brand DNA.'
• Three customization tiers map to budget: open-source quantized weights (free/low), Ideogram's self-serve training app ($60 for 2 trainings/month, ≥15 images), and white-glove enterprise work where their annotation team curates brand-specific prompts/keywords/mascots.
• Open weights is the go-to-market, not charity: it lets Ideogram partner with chip-makers, inference providers and enterprises for on-prem, on-device and data-sovereignty use — 'we are very serious about building the foundation model and we would like to work with you, whoever you are.'
• The build workflow is already agentic: Ideogram runs its own MCP internally — 'ask it to connect to the API and generate a bunch of images... in a couple hours you have your landing page up and running' — with the open problem being automated evaluation in the loop so a human needn't inspect every output.

Notable quotes

Themes

• David-vs-Goliath wedge (focus over scale)
• taste as a competitive moat
• open-weights distribution strategy
• customization for enterprise brand
• intermediate-representation prompting