Ideogram 4¶
| Model | Source | Revision | Weight license | Parity (CPU, fp32) |
|---|---|---|---|---|
ideogram4-fp8 |
ideogram-ai/ideogram-4-fp8 | ee79a72 |
Ideogram 4 Non-Commercial License | Verified |
Catalog: ideogram-4 ·
Models: ideogram4-fp8 · Task: t2i
Architecture. A Qwen3-VL 8B text model uses fp8 linears and supplies 53,248-wide features
formed by stacking selected hidden states. Two separate 9.3B single-stream diffusion transformers
provide the conditional and unconditional classifier-free-guidance passes. Each transformer has 34
layers, 18 attention heads of width 256, adaLN and 24/20/20 multimodal RoPE. The official
32-channel autoencoder applies latent normalization. The official sampler supplies the
V4_QUALITY_48 (default), V4_DEFAULT_20 and V4_TURBO_12 presets, including each preset's
per-step guidance schedule. The native image size is 1024×1024.
The published checkpoint contains only quantized weights: large linear weights are float8 e4m3
with per-row float32 scales. imagehub keeps those weights in fp8 and dequantizes each forward pass
as weight.to(dtype) * weight_scale.to(dtype).unsqueeze(1), exactly as the official
implementation does. The result is bit-identical when the compute dtype is float32.
Usage. The repository is gated. Accept its terms and provide HF_TOKEN when downloading:
from imagehub import pipeline
with pipeline(
"t2i", backend="ideogram4-fp8", model="checkpoints/ideogram4-fp8", device="mps"
) as pipe:
result = pipe("a cat, oil painting", seed=42, preset="V4_QUALITY_48")
result.images[0].save("cat.png")
Settings. preset defaults to V4_QUALITY_48, and width/height default to 1024 and must
be multiples of 16. V4_DEFAULT_20 and V4_TURBO_12 select the official 20-step and 12-step
sampler configurations. The other supported options are steps, guidance, guidance_schedule,
seed and latents. Unrecognized settings raise an error.
Seed compatibility. On CPU in float32, the same seed and settings give the same image as the
official ideogram-oss/ideogram4 implementation at commit 990fe1c; this validation is not
against diffusers (validation). MPS and CUDA runs are smoke tests only; there is
no claim of exact equality on these devices.
Out of scope. The CUDA-only nf4 releases, magic prompt, safety moderation and fp16/bf16 parity are not supported. Magic prompt and safety moderation call external paid APIs.
License. The weights use the Ideogram 4 Non-Commercial License; see the license. imagehub does not distribute weights.
Reference and imagehub, side by side¶
Drag the slider. Left is the recorded reference, right is imagehub, from the same seed, on CPU in float32.

referenceimagehub