KO | EN

GitLite

Minimalist Open Source Repository Search Engine

gitlite — search

#dsh #dsh-plugin #typescript #developer-tools #claude-code #ai-agent #codex #deepseek-harness #ai-agents #cli #react #mcp

ComfyUI-ZImageTurbo-FlowSampler

★ 10 Open GitHub ↗

Rectified-flow samplers for Z-Image-Turbo. Clean interpolation formula without sigma division, linear schedulers, higher-order variants. Fixes built-in euler/euler_ancestral instability issues.

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# ComfyUI-ZImageTurbo-FlowSampler

Custom ComfyUI nodes for Z-Image-Turbo implementing verified rectified-flow samplers with clean interpolation formula, linear sigma schedulers, and higher-order variants.

## The Discovery

Z-Image-Turbo was distilled using:
- Linear sigma schedule (1.0 → 0.0)
- Deterministic Euler steps via rectified-flow interpolation
- No intermediate noise injection
- Explicit final step handling when σ_next = 0

After extensive code analysis of ComfyUI's sampling implementation, a critical mismatch was identified between the built-in samplers and Z-Image-Turbo's distillation process.

## Built-in Sampler Issues

### `sample_euler` (default)

```python
# Karras ODE formulation
d = (x - denoised) / sigma  # Division by sigma → unstable near σ≈0
x = x + d * dt              # Implicit final step
```

- Contains churn machinery (`s_churn`, `s_tmin`, `s_tmax`)
- Division by sigma causes numerical instability
- No explicit σ_next = 0 handling
- Designed for traditional diffusion, not rectified flow

### `sample_euler_ancestral`

Contains a rectified-flow branch (`sample_euler_ancestral_RF`), but activation requires:
- Sampler must be `euler_ancestral`
- Model sampling type must be `CONST`
- eta > 0 (introduces ancestral noise)

For deterministic Z-Image-Turbo workflows (CFG=1.0, eta=0), this clean path is never reached.

## Custom Implementation

```python
# Core update from sample_euler_flow
if sigma_next == 0:
    x = denoised
else:
    ratio = sigma_next / sigma
    x = ratio * x + (1.0 - ratio) * denoised
```

## Comparison

| Feature | Built-in euler | Built-in euler_ancestral | Custom euler_flow |
|---------|---------------|-------------------------|-------------------|
| Update formula | Karras ODE derivative | RF interpolation (gated) | RF interpolation (always) |
| Division by sigma | Yes | No (when RF branch active) | No |
| Final step handling | Implicit | Explicit | Explicit |
| Churn machinery | Present | Present (outside RF branch) | Removed |
| Ancestral noise | Not required | Required (eta > 0) for RF | Optional (eta=0 default) |
| Z-Image turbo alignment | Partial | Conditional | Complete |

## Nodes Provided

### Samplers
- **euler_flow**: 1st-order rectified-flow Euler (recommended)
- **euler_flow_alt**: Derivative formulation (for comparison)
- **midpoint_flow**: 2nd-order midpoint method
- **heun_flow**: 2nd-order Heun method
- **euler_flow_verbose**: Debug mode with step-by-step output

### Schedulers
- **ZImageTurbo Scheduler**: Linear sigma schedule (1.0 → 0.0)
- **ZImageTurbo Scheduler (Advanced)**: With denoise control for img2img
- **FlowMatch Scheduler**: Generic flow-matching with shift parameter
- **FlowMatch Scheduler (Dynamic)**: Resolution-aware shift calculation
- **ZImage Presets**: One-click configurations for Z-Image-Turbo, Flux, SD3

### Utilities
- **Sigma Viewer**: Debug node to visualize sigma schedules
- **ZImage Sampler**: Unified sampler node with type selection and eta control

## Installation

### Via ComfyUI Manager
1. Search for "ComfyUI-ZImageTurbo-FlowSampler"
2. Install and restart ComfyUI

### Manual Installation
```bash
cd ComfyUI/custom_nodes
git clone https://github.com/capitan01R/ComfyUI-ZImageTurbo-FlowSampler.git
```
Restart ComfyUI.

## Nodes
![](https://raw.githubusercontent.com/capitan01R/ComfyUI-ZImageTurbo-FlowSampler/main/images/nodes.png)
### Recommended Settings for Z-Image-Turbo
- **Steps**: 8-9
- **CFG**: 1.0
- **Sampler**: euler_flow
- **Scheduler**: ZImageTurbo Scheduler (linear)
- **eta**: 0.0 (deterministic, sharpest)

### eta Parameter Guide
- **0.0**: Deterministic, sharpest results
- **0.1-0.2**: Subtle variation, can reduce artifacts
- **0.3-0.5**: Balanced diversity with slight softening
- **0.5-1.0**: More creative variation, softer outputs

### Example Workflow
1. Add **ZImage Sampler** node, select `euler_flow`, eta=0.0
2. Add **ZImageTurbo Scheduler** node, steps=8
3. Connect to **SamplerCustomAdvanced**
4. Set CFG=1.0

## Testing

All nodes tested on Z-Image-Turbo at:
- 8-9 steps
- CFG 1.0
- Linear sigma schedule
- Various eta values (0.0-1.0)

Observations on bf16 generations:
- Sharper detail preservation
- Improved facial coherence
- Reduced numerical artifacts in final steps

## Technical Details

### Why Interpolation > Derivative

Both formulations are algebraically equivalent when σ ≠ 0:

**Karras ODE:**
```
d = (x - x̂) / σ
x' = x + d * (σ_next - σ)
```

**RF Interpolation:**
```
ratio = σ_next / σ  
x' = ratio * x + (1 - ratio) * x̂
```

The interpolation form provides:
- No division by potentially small sigma values
- Explicit final-step handling
- Cleaner code path without legacy diffusion machinery
- Direct alignment with Z-Image-Turbo distillation process

## Compatibility

**Primary target**: Z-Image-Turbo

Also compatible with other flow-matching models via appropriate schedulers:
- Flux (use dynamic scheduler or presets)
- SD3/SD3.5 (use shifted scheduler with shift=3.0)

---

**Author**: Capitan01R  
**Version**: 1.0.0

Based on analysis of:
- Tongyi-MAI/Z-Image official implementation
- ComfyUI k_diffusion sampling code
- Rectified flow mathematical framework

---

**Note**: Linear sigma schedule is critical for Z-Image-Turbo. Non-linear schedules break the distillation and produce degraded results.

ComfyUI-ZImageTurbo-FlowSampler

Explore Similar Repositories

Was this content helpful?

Recent Feedback

Download README