| --- |
| license: other |
| license_name: ideogram-4-non-commercial |
| license_link: https://github.com/ideogram-oss/ideogram-4/model_licenses/LICENSE-IDEOGRAM-4-NON-COMMERCIAL |
| base_model: ideogram-ai/ideogram-4-fp8 |
| pipeline_tag: text-to-image |
| tags: |
| - ideogram4 |
| - mlx |
| - mflux |
| - apple-silicon |
| - distillation |
| - cfg-distillation |
| - step-distillation |
| extra_gated_prompt: >- |
| This model is a derivative of Ideogram 4 and is provided under and subject to the |
| Ideogram Non-Commercial Model Agreement. By accessing it you agree to be bound by |
| that agreement, including its non-commercial use restriction and acceptable use policy. |
| --- |
| |
| # Ideogram v4 distilled for Apple Silicon (mflux) |
|
|
| Two single-branch distillations of the Ideogram v4 conditional DiT (9.3B, fp8), |
| targeting local inference on Apple Silicon via [mflux](https://github.com/filipstrand/mflux): |
|
|
| - **`cfg-fold-12step/`** — classifier-free guidance folded into the conditional |
| transformer (LoRA self-distillation on guided dual-DiT trajectories, then fused with |
| per-row fp8 requantization). Halves every step: 24→12 forwards at 12 steps. |
| - **`step-fold-6step/`** — the CFG-fold further distilled 12→6 steps on compiled teacher |
| trajectories (6-step grid = the even nodes of the 12-step grid, proven exact). |
|
|
| **Only the derivative conditional transformer is redistributed here.** Text encoder |
| (Qwen3-VL-8B), VAE (FLUX.2), scheduler and the unused unconditional transformer come from |
| the gated upstream [ideogram-ai/ideogram-4-fp8](https://huggingface.co/ideogram-ai/ideogram-4-fp8) |
| release, which you must accept separately. An assembly script in the companion GitHub |
| repo links both into a runnable mflux model directory. |
|
|
| ## Results (M4 Pro, 48GB, 800×992, warm) |
|
|
| | model | forwards | e2e | notes | |
| |---|---|---|---| |
| | stock dual-DiT, 12 steps | 24 | ~250s | baseline | |
| | cfg-fold-12step | 12 | ~135s | seed-compatible with baseline character | |
| | step-fold-6step | 6 | ~86s | ≈ fal ideogram-v4-instant (8 fwd) wall-clock, −25% compute | |
|
|
| Character notes from blind A/Bs: the folds preserve the base model's filmic/atmospheric |
| range (dragged-shutter motion, halation, editorial grades); fal's official instant |
| distill reads crisper/more commercial on the same briefs. The 6-step variant is |
| style-faithful but recomposes at a fixed seed relative to 12-step. |
|
|
| An additional measured property of single-branch operation: the dual-DiT's |
| seed-dependent refusal placards (which trigger on many harmless prompts at production |
| resolutions) do not occur in either fold, consistent with fal's official single-branch |
| distillation. Nothing about the model's content training was modified; users remain |
| bound by Ideogram's Acceptable Use Policy. |
|
|
| ## How to use |
|
|
| See [the companion GitHub repo](https://github.com/felix-ab/ideogram4-mlx-distill)sitory for `download_assemble.py`, the guidance-1.0 cadence |
| runner, prompt-schema notes and the benchmark harness. Minimal facts: run with |
| guidance 1.0 (single branch); 6-step preset `mu 0.5, std 1.75`; dimensions divisible |
| by 16; structured-JSON prompts. |
|
|
| ## Training / distillation procedure |
|
|
| 1. **CFG-fold**: rank-32+ LoRA on the conditional DiT, trained to match guided dual-DiT |
| outputs on 56 self-generated caption briefs (text-weighted loss), then fused into the |
| fp8 weights via per-row requantization (`nscale = rowmax|W′|/448`). |
| 2. **Step-fold**: teacher trajectories captured through the compiled predictor (eager |
| capture drifts — SSIM 0.96 by step 12 vs 1.0000 compiled), student trained on |
| trajectory states at the 6-step grid nodes, encoder-resident encode-per-visit. |
| 3. Verification: bit-deterministic golden renders (SSIM 1.0000), fused-vs-adapter |
| equivalence, text-rendering checks, refusal-rate matrix, speed ladder. |
|
|
| No images beyond the base model's own outputs were used for these distillations. |
|
|
| ## Limitations & risks |
|
|
| Non-commercial only (license flows down). Baked guidance (no CFG sweeps). 6-step |
| recomposition vs 12-step. Incidental background text unreliable (as in the base model). |
| fp8 weights: Apple-Silicon-tuned; CUDA users should prefer fal's official release. |
|
|
| ## Attribution |
|
|
| “Ideogram 4 is provided under and subject to the Ideogram Non-Commercial Model Agreement |
| available at https://github.com/ideogram-oss/ideogram-4/model_licenses/LICENSE-IDEOGRAM-4-NON-COMMERCIAL. |
| All rights reserved. Copyright © Ideogram, Inc.” — see NOTICE. Distillations by Felix |
| Brener (2026); modified files are enumerated in MODIFICATIONS.md. Not an official |
| Ideogram product; not endorsed, approved or validated by Ideogram, Inc. |
| |
