Flow-DPO on Z-Image Turbo: Cinematic Lighting Without Losing the Speed
Z-Image Turbo is a fast, four-to-nine-step model. Flow-DPO is a preference-tuned LoRA aimed at cinematic lighting and composition. Putting them together is the obvious experiment, and this release does it with a control network and segmented rendering on top.
The two graphs
| Workflow | Nodes | Base latent |
|---|---|---|
| Full comparison (CN + Flow-DPO + UltraReal + Engineer) | 180 | 512 × 472 |
| Flow-DPO comparison only | 47 | 1024 × 1024 |
Both load the same model stack:
| Node | Value |
|---|---|
UNETLoader | z_image_turbo_bf16.safetensors |
CLIPLoader | qwen_3_4b.safetensors, type qwen_image |
VAELoader | ae.safetensors |
LoraLoaderModelOnly | zit_fdpo_v1.safetensors at 1.0 |
The encoder type differs from other Z-Image workflows
qwen_3_4b.safetensors is loaded here with type qwen_image.
In other Z-Image Turbo workflows (the PiD release, for example), the same encoder file is loaded with type lumina2.
Same encoder, same base model family, two different type strings. That is the same soft-failure pattern that runs through the Anima family: get the type wrong and nothing errors, the image just responds weakly to the prompt.
If you are merging nodes from these graphs into your own, copy the loader node rather than retyping it. That rule applies across every model family covered in this series.
Flow-DPO at full weight
zit_fdpo_v1.safetensors at weight 1.0.
Flow-DPO is a preference-tuned LoRA — it was trained on pairs of images where one was preferred, so it shifts the model toward whatever the preference data considered better. The claim here is cinematic lighting.
At full weight the preference is applied completely. That means you get the LoRA's aesthetic, not a mix of it and the base model's. For a lighting-focused LoRA that is usually what you want — a half-weight cinematic lighting LoRA tends to produce flat images with occasional dramatic patches.
The trade: preference LoRAs at full weight also carry the biases in their training data. If your prompt wants a look the preference data did not cover, the LoRA will fight the prompt.
Why 512×472 for the big graph
The 180-node graph generates at 512 × 472 — under half a megapixel, and an unusual aspect (1.085, close to square).
That is a segmented-rendering setup. ControlNet-guided segmented rendering generates regions at low resolution and composites, rather than generating one large image. Low base resolution is the point: each segment is cheap, and the control network supplies the structure.
47-node graph at 1024×1024 is the straightforward one-pass comparison.
What the comparison is measuring
Four things are being layered:
- Z-Image Turbo — the fast base
- Flow-DPO — the lighting/composition preference
- UltraReal — real-texture fine-tune
- Engineer V4 — presumably a prompt-engineering preset or a workflow variant
With four variables and two graphs, this is not a controlled experiment. It is a demonstration of a production stack. The useful output is "what does this combination produce", not "which component contributes what".
To attribute results you would need one graph per variable, which is four graphs, not two.
Where this breaks
- Encoder type inconsistency with other Z-Image workflows (above).
- 512×472 base means the 180-node graph alone does not produce a deliverable image — it needs a composite or an upscale step.
- Four variables, two graphs makes attribution impossible without rebuilding.
- Flow-DPO at 1.0 will fight prompts that ask for a non-cinematic look, and there is no example of a lower weight.
- 180 nodes is among the largest graphs in this series; debugging it is expensive.
Practical approach
Run the 47-node graph first at 1024×1024. It is a clean Z-Image Turbo + Flow-DPO render with nothing else layered on. That tells you what the LoRA does on its own. Only then open the 180-node graph, and when you do, change one thing at a time.
If Flow-DPO is fighting your prompt, the first experiment is the weight — try 0.6 or 0.7 and compare. The workflow gives you no example of that, but it is one number.
Workflows
Two graphs, published on RunningHub with invite code rh-v1111:
- Z Image Turbo + CN + Flow-DPO + UltraReal + Engineer V4 comparison
- Z Image Turbo + Flow-DPO + UltraReal + Engineer V4 comparison test
Files
https://huggingface.co/AI-KSK/ksk-video-200-comfyui-workflows
Notes
- Workflow files only. Model weights are not included; each keeps its own licence.
- Tutorial exports, not a product.
- Node counts and loader values quoted above are read from the published JSON files.
- Independent AI-KSK release. Not affiliated with the model authors or RunningHub.
