<h1 id="ffusion-turbo">FFusion Turbo</h1>An experimental weight-adjusted Z-Image Turbo checkpoint, retuned to lean digital / CGI instead of the default photorealistic bias. SFW-oriented.Drop-in replacement for <code>z_image_turbo_bf16</code> — same architecture, same 9-step turbo workflow, same VAE and text encoder. Just swap the diffusion model.<hr /><h3 id="experimental-notice">🧪 Experimental Notice</h3>This is a weight experiment, not a finetune on new data. The model was adjusted to shift its aesthetic prior toward rendered / CGI output. Results will vary — some prompts respond strongly, others look nearly identical to base turbo. Consider this a sandbox release.<hr /><h3 id="what-it-does">🎨 What it does</h3><ul><li>Cleaner 3D render / CGI aesthetic out of the box</li><li>Stronger digital illustration and stylized outputs</li><li>Less aggressive skin-texture / pore / wrinkle bias from the stock turbo</li><li>Still fast — 6–10 steps, same settings as base turbo</li></ul>Best for: product renders, concept art, stylized characters, abstract compositions, anything that should look made rather than photographed.<hr /><h3 id="usage">⚙️ Usage</h3>SettingValueBaseZ-Image TurboSteps8–10 (9 is the sweet spot)CFG1.0Samplerany turbo-compatible (euler, dpm++ 2m)VAEstock Z-Image <code>ae.safetensors</code>Text encoderstock <code>qwen_3_4b.safetensors</code>PrecisionBF16No trigger word needed — it's a base checkpoint, style is always on.<hr /><h3 id="tips">💡 Tips</h3><ul><li>Pairs well with CGI / 3D-style LoRAs — stacks their effect instead of fighting it</li><li>If you want to pull back toward photoreal, blend with stock turbo at 0.5 / 0.5</li><li>Works with any Z-Image Turbo ControlNet / workflow unchanged </li></ul><pre><code>---

### 🔬 Model Stats

Full BF16 checkpoint — 453 tensors, **6.155B parameters**, no NaN / no Inf. Clean build.

| Module | Tensors | Params | % of Total |
|---|---:|---:|---:|
| `layers.*` (transformer blocks) | 390 | 5.43 B | **88.2%** |
| `noise_refiner` | 26 | 361.8 M | 5.9% |
| `context_refiner` | 22 | 353.9 M | 5.8% |
| `cap_embedder` | 3 | 9.8 M | 0.2% |
| `final_layer` | 4 | 1.2 M | &lt;0.1% |
| `t_embedder` | 4 | 0.5 M | &lt;0.1% |
| `x_embedder` | 2 | 0.25 M | &lt;0.1% |
| pad tokens | 2 | — | — |
| **Total** | **453** | **6.155 B** | 100% |

### 📊 Weight Distribution

- **Global range:** min ≈ **−14.00**, max ≈ **+13.94** — in-line with typical DiT checkpoints
- **Most active modules** (highest std): the deep layers `layers.26` through `layers.29` — this is where the style adjustment is concentrated. Their `ffn_norm2` and `attention_norm2` tensors show std up to 3.2 vs. a model average of ~0.32
- **Most conservative modules:** `t_embedder` (std ~0.005–0.02) — timestep embedding is nearly untouched, as expected
- **Feed-forward `w2` weights** carry the largest absolute values (up to ±14), consistent with how Z-Image's MLP projections store learned priors

### ✅ File Integrity

| Check | Result |
|---|---|
| NaN tensors | **0** |
| Inf tensors | **0** |
| Dtype consistency | 100% BF16 |
| Architecture match vs. `z_image_turbo_bf16` | structurally identical (906/906 keys) |

---</code></pre>

FFusion Turbo

Full BF16 checkpoint — 453 tensors, 6.155B parameters, no NaN / no Inf. Clean build.▸ <code>layers.*</code> (transformer blocks) — 390 tensors · 5.43 B params · 88.2% ▸ <code>noise_refiner</code> — 26 tensors · 361.8 M params · 5.9% ▸ <code>context_refiner</code> — 22 tensors · 353.9 M params · 5.8% ▸ <code>cap_embedder</code> — 3 tensors · 9.8 M params · 0.2% ▸ <code>final_layer</code> — 4 tensors · 1.2 M params · &lt;0.1% ▸ <code>t_embedder</code> — 4 tensors · 0.5 M params · &lt;0.1% ▸ <code>x_embedder</code> — 2 tensors · 0.25 M params · &lt;0.1%Total: 453 tensors · 6.155 B parameters━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━📊 Weight Distribution▸ Global range: min ≈ −14.00, max ≈ +13.94 — in line with typical DiT checkpoints▸ Most active modules (highest std): the deep layers <code>layers.26</code> through <code>layers.29</code> — this is where the style adjustment is concentrated. Their <code>ffn_norm2</code> and <code>attention_norm2</code> tensors show std up to 3.2 vs. a model average of ~0.32▸ Most conservative modules: <code>t_embedder</code> (std ~0.005–0.02) — timestep embedding is nearly untouched, as expected from a weight-adjustment rather than a retrain▸ Feed-forward <code>w2</code> weights carry the largest absolute values (up to ±14), consistent with how Z-Image's MLP projections store learned priors━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━✅ File Integrity▸ NaN tensors: 0 ▸ Inf tensors: 0 ▸ Dtype consistency: 100% BF16 ▸ Architecture match vs. <code>z_image_turbo_bf16</code>: structurally identical (906/906 keys)

Files with SHA256: c7c25cf8196359461a35ece92f497527d761e08a8dee9c3907da70113f8869b1