Four Z-Image workflows, three different ControlNet and Turbo configurations, and one recurring configuration problem that you can see just by reading the loader nodes.
The graphs
| Workflow | Nodes | Geometry | CLIP type |
|---|---|---|---|
| Z-Image Base + ControlNet | 14 | 720×1280 | qwen_image |
| Z-Image Base 8-step & 4-step acceleration | 53 | 1280×720 | qwen_image |
| Z-Image Base + ControlNet + Turbo high/low noise | 44 | 512×472 | lumina2 |
The encoder type problem
qwen_3_4b.safetensors appears in all three, and the type differs:
- Two graphs:
qwen_image - One graph:
lumina2
Same encoder file, same Z-Image base, two different type strings, in one release.
This is the third time this pattern has appeared in this series — it shows up in the Anima family and in the Z-Image family. The consequence is always the same: wrong type does not error, it produces an image that responds weakly to the prompt, and the natural reaction is to blame the prompt.
Rule that saves time: copy the loader node from the graph you are extending. Never retype it.
The high/low noise split
The 44-node graph is named for a two-stage high/low noise split: the high-noise pass locks structure and composition, the low-noise pass refines detail.
That split exists because one sampler pass cannot optimise both. A high-noise pass decides layout using large structural changes; a low-noise pass can only refine what is already there. Running them separately lets each do its job.
The two-pass structure also explains the rest of the release: the 8-step and 4-step acceleration models are being tested for the low-noise pass specifically, because that is where step count matters most for detail.
The 8-step versus 4-step question
4-step distillation gets you speed. 8-step gets you more of the original model's behaviour at twice the cost.
For the low-noise refinement pass, 4 steps often suffices: the structure is already set, and the pass is only adding detail. For the high-noise pass, 4 steps is riskier — layout decisions benefit from more sampling.
If you are splitting high/low noise, the sensible configuration is more steps on the high-noise pass and fewer on the low-noise pass, not the same count on both.
Where this breaks
- Encoder type inconsistency across three graphs in one release.
-
720×1280versus1280×720— the two large graphs are opposites in aspect. Easy to run the wrong one for your content. -
512×472in the Turbo graph means it starts much lower and relies on the two-pass structure. - Four-step on both passes will produce weak layout; step count should differ per pass.
- 14-node ControlNet graph has no validation — a wrong loader type or missing model produces output silently.
Practical order
Start with the 14-node ControlNet graph — it is the simplest and it will immediately reveal whether your encoder type is right (weak prompt adherence is obvious with a control signal constraining composition).
Then the 53-node acceleration comparison. The two-pass Turbo graph last, because it has the most configuration surface.
Workflows
Three graphs, published on RunningHub with invite code rh-v1111:
- Z Image Base + ControlNet
- Z Image Base 8-step & 4-step acceleration
- Z Image Base + ControlNet + Turbo high/low noise
Files
https://huggingface.co/AI-KSK/ksk-video-178-comfyui-workflows
Notes
- Workflow files only. Model weights are not included; each keeps its own licence.
- Tutorial exports, not a product.
- Node counts, geometries and CLIP types quoted above are read from the published JSON files.
- Independent AI-KSK release. Not affiliated with Tongyi or RunningHub.
Optional API platform - 88API (affiliate link)
Explore 88API for access to multiple AI model APIs. Available models, routes and prices are listed on the platform; Alipay top-ups are supported.
https://88api.ai/sign-up?aff=usx3
Disclosure: I may earn a commission from eligible purchases after you register through this referral link.
Top comments (0)