FFusion Turbo

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 z_image_turbo_bf16 — same architecture, same 9-step turbo workflow, same VAE and text encoder. Just swap the diffusion model.

🧪 Experimental Notice

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.

🎨 What it does

• Cleaner 3D render / CGI aesthetic out of the box

• Stronger digital illustration and stylized outputs

• Less aggressive skin-texture / pore / wrinkle bias from the stock turbo

• Still fast — 6–10 steps, same settings as base turbo

Best for: product renders, concept art, stylized characters, abstract compositions, anything that should look made rather than photographed.

⚙️ Usage

SettingValueBaseZ-Image TurboSteps8–10 (9 is the sweet spot)CFG1.0Samplerany turbo-compatible (euler, dpm++ 2m)VAEstock Z-Image ae.safetensorsText encoderstock qwen_3_4b.safetensorsPrecisionBF16

No trigger word needed — it's a base checkpoint, style is always on.

💡 Tips

• Pairs well with CGI / 3D-style LoRAs — stacks their effect instead of fighting it

• If you want to pull back toward photoreal, blend with stock turbo at 0.5 / 0.5

• Works with any Z-Image Turbo ControlNet / workflow unchanged
  
  

---

### 🔬 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 | <0.1% |
| `t_embedder` | 4 | 0.5 M | <0.1% |
| `x_embedder` | 2 | 0.25 M | <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) |

---