A prompt with structure and depth.
- Text features are computed once and reused during sampling.
- Optional Magic Prompt expansion is upstream of the denoiser.
structured caption → text-only encoder → 13 feature tapsIMAGE / IDEOGRAM 4.0Your imagination.
Your imagination.
In production.
An image model for typography, composition and visual expression. Explore the text encoder, joint transformer stream and flow-matching path behind its generation process.
IDEOGRAM 4.0↗
OPEN THE 3D ARCHITECTURE ATLAS DiT PARAMETERS9.3B
TRANSFORMER34-layer DiT
SAMPLINGFlow matching
WORKFLOWAsynchronous jobs
INSIDE THE MODEL / 3D ARCHITECTURE ATLASA shared stream.
A shared stream.
A visible idea.
Structured language and image latents meet inside a flow-matching transformer.
Drag to rotate · Arrow keys supported
MODULE 01 / MULTI-LAYER CONDITIONING
13
TEXT ENCODER HIDDEN-STATE TAPSA prompt with structure and depth.
Structured captions describe objects, literal text and composition. A frozen Qwen3-VL-8B-Instruct encoder runs in text-only mode; thirteen hidden-state taps are combined for conditioning.
structured caption → text-only encoder → 13 feature tapsREADING THE DIAGRAM
The same latent canvas is updated repeatedly. Text features condition the joint stream; the VAE decodes the final image. The animation is explanatory; no model inference is executed.
BACKBONE MAP
34 layers. A deliberate pattern.
34 Joint DiT
The 34 blocks belong to the joint DiT. The text encoder, guidance branch and VAE are separate components.
A CLOSER LOOK / IDEOGRAMThe details behind
The details behind
the diagram.
Configuration, memory and computation—each with a specific role in the model.
Text and image share the room.
- 9.3B describes the DiT, not every component in the pipeline.
- The transformer width is 4,608, with 18 attention heads.
text + image tokens → attention → SwiGLU → next blockMove the canvas through a velocity field.
- Guidance uses distinct conditional and unconditional components.
- Q, K and V are recomputed as the denoising state changes.
zₙ₊₁ = zₙ + v_guided × (tₙ₊₁ − tₙ)Pixels become a compact token grid.
- Each packed token contains 128 latent values.
- After sampling, unpatching and VAE decoding recover RGB pixels.
image H×W → latent H/8×W/8 → tokens H/16×W/16FOLLOW THE RESEARCHGo straight
Go straight
to the source.
Developer papers, released configurations and reference implementations. Reviewed 10 September 2026.
I1
OFFICIAL DOCUMENTATION
I2Ideogram 4.0 pipeline guide
Text encoder, denoiser and VAE decoder
SOURCE CODE
I3Ideogram architecture & model implementation
Configuration, attention and transformer blocks
OFFICIAL DOCUMENTATION
I4Sampler presets and inference guide
Turbo 12, Default 20 and Quality 48 presets
SOURCE CODE
Reference flow and guidance pipeline
Text features, guidance branches and Euler update
WHAT THIS GUIDE COVERS
The released architecture is documented in code and pipeline guides; no undisclosed training-corpus claims are made here.
Open weights retain their model-specific licence. The code licence is not a substitute for weight-use rights.
MADE MEANINGFUL IN YOUR BUSINESSFrom capability
From capability
to possibility.
Product imagery
Explore art direction, compositions and campaign assets around your product.
Typography-led creative
Generate concepts that put words and visual composition in the same frame.
Creative iteration
Bring repeatable settings and predictable workflow integration to the creative process.
Know your inference service scope.
Perchy provides managed AI inference through authenticated APIs. Customers do not receive SSH, root, operating-system or bare-metal access. The service does not provide model-training or fine-tuning functionality. Server administration remains with Perchy.
Managed API availability requires appropriate model and commercial-use rights. Licence costs are included when assessing service economics. Outputs require review before publication. Image generation uses a latent denoising workflow; it is not an autoregressive chat KV-cache workload.
Read the model developer’s documentationKEEP EXPLORINGAnother model.
All models Another model.
Another way to compute.
One conversation.
A whole new trajectory.