| --- |
| license: other |
| license_name: ideogram-4-non-commercial |
| license_link: https://huggingface.co/ideogram-ai/ideogram-4-fp8 |
| base_model: ideogram-ai/ideogram-4-fp8 |
| pipeline_tag: text-to-image |
| tags: [text-to-image, diffusion, flow-matching, quantization, int8, w8a8, ideogram] |
| --- |
| |
| # Ideogram 4 — INT8 W8A8 (Transformer Lab) |
|
|
| An **INT8 W8A8** post-training quantization of the Ideogram 4 DiT (per-channel int8 |
| weights + per-token dynamic int8 activations). |
|
|
| ℹ️ **Quantized DiT only.** This checkpoint is the DiT (both CFG branches). To generate you |
| also need the **Qwen3-VL text encoder and VAE** from the base repo [`ideogram-ai/ideogram-4-fp8`](https://huggingface.co/ideogram-ai/ideogram-4-fp8) |
| and the custom inference code at [`github.com/ideogram-oss/ideogram4`](https://github.com/ideogram-oss/ideogram4). |
| The quantization recipe and loader are included **in this repo** (`recipe.json`, `safetensors_loader.py`). |
|
|
| ## Why INT8 |
| INT8 **holds the FP8 quality ceiling**: on a 200-prompt benchmark the paired same-seed |
| bootstrap CI for INT8−FP8 includes zero on both Pick and CLIP (statistically |
| indistinguishable at this sample size), and it beats NF4 by **+1.9 CLIP** (CI excludes |
| zero). Text rendering stays legible (OCR NED 0.704 vs NF4 0.760). |
|
|
| ## Samples |
|  |
|
|
| ## Method |
| Per-channel int8 weights + per-token dynamic int8 activations + SmoothQuant (α=0.5) + |
| mixed-precision protection of the **top-17 fragility-prone layers** (the FFN |
| down-projections, ~8% of linears), kept in bf16. See `recipe.json` for the exact module |
| list and tensor layout. |
|
|
| ## Notes |
| - On-disk **~20.4 GB** — at 8-bit weights this is **FP8-class in size, not smaller than |
| NF4** (10.4 GB). Its win is *quality*, not memory. |
| - Without a fused Ampere INT8 GEMM it runs ~184 s/img (no speed win yet); a custom-kernel |
| build for the speedup is planned. |
|
|
| ## How to run (self-contained) |
|
|
| Everything you need is in this repo. The safetensors is the **quantized DiT only**, so |
| step 1 fetches the text encoder + VAE + the inference package. |
|
|
| ```bash |
| # 1) one-time: install the ideogram4 package + download the base components |
| # (needs your own access to the GATED base repo ideogram-ai/ideogram-4-fp8) |
| python download_deps.py |
| |
| # 2) generate |
| python usage.py "a poster that says HELLO" |
| ``` |
|
|
| Files here: |
| - `ideogram4-int8-w8a8.safetensors` — the INT8 W8A8 DiT (both CFG branches). |
| - `safetensors_loader.py` — reconstructs the W8A8 layers + loads them (reference impl). |
| - `download_deps.py`, `usage.py` — setup + a minimal generation example. |
| - `recipe.json` — the exact recipe (protected-layer list, tensor layout). |
|
|
| > `safetensors_loader.py` is a **reference**: the math is validated, but the standalone |
| > loader hasn't been GPU-tested end to end yet — verify before production use. This INT8 |
| > build runs eager (no fused INT8 kernel yet), so it holds FP8 quality but isn't faster. |
| |
| ## License |
| Derived from Ideogram 4 under its **non-commercial, research-only** license. See `LICENSE`. |
| |
