Ideogram releases its flagship image model openly
Ideogram has released its newest flagship text-to-image model as open weights, an unusual move for a closed image lab that gives builders JSON-structured prompting with direct control over layout and color.
Ideogram 4.0 is a 9.3B-parameter open-weight text-to-image model, the company said in a technical write-up. Like other recent open releases, it uses a single-stream Diffusion Transformer (DiT). Text and image tokens share the same self-attention sequence and projections across all 34 layers.
Ideogram highlights two design choices. The first is the text encoder: the model uses Qwen3-VL-8B-Instruct, a vision-language model, in text-only mode. The DiT takes hidden states from 13 of the encoder's intermediate layers, concatenated together, instead of a single hidden state. The second is training data: the model was trained only on structured JSON captions. These captions describe each element in the image and include a style block, with optional bounding boxes and color palettes. The reference inference pipeline checks every prompt against the JSON schema and rejects any prompt that does not parse.
This gives users direct controls that plain-text prompts do not. They can specify up to 16 hex colors per image and five per element, and they can place elements with bounding boxes.
Other specs: 2,048 max text tokens, resolutions from 256 to 2,048 pixels per side, flexible aspect ratios, and nf4 and fp8 quantization. Sampling uses a flow-matching Euler sampler with asymmetric classifier-free guidance, where the unconditional pass drops text tokens entirely. The presets end with a few low-guidance "polish" steps. For example, V4_QUALITY_48 runs 45 steps at guidance 7, then three steps at guidance 3.
The material provided does not include license terms or comparative benchmark figures.