Qwen Image 2.1 Turbo: Eight Steps Against Twenty-Five, and Eight ControlNet Modes
Two claims to test here: that an 8-step Turbo sampler matches a 25-step original, and that ControlNet Union can drive eight different control modes from one adapter. Eleven comparison sets later, both claims hold with c
Two claims to test here: that an 8-step Turbo sampler matches a 25-step original, and that ControlNet Union can drive eight different control modes from one adapter.
Eleven comparison sets later, both claims hold with caveats worth knowing.
The five workflows
| # | Workflow | What it tests |
|---|---|---|
| β | ControlNet Union Depth, 75 parameter sets | Depth control at scale |
| β | ControlNet Union nine-in-one, 8 CN modes + mask inpaint | Breadth of control |
| β | Turbo 8-step vs original 25-step, 11 sets | Speed/quality parity |
| β | Fix 1.0/2.0 quality repair | Removing high-frequency fake detail |
| β | Turbo 8-step vs 25-step (duplicate, batch variant) | Reproducibility |
8 steps vs 25 steps: what the 11 sets show
Composition and colour β indistinguishable. This is expected; both settle early.
Mid-frequency detail β Turbo is slightly smoother but defensibly so. At normal viewing distance it does not read as a downgrade.
High-frequency detail β this is where the difference lives. Fabric weave, hair strands and skin pores are visibly reduced at 8 steps. On a 100% crop it is obvious; on a phone screen it is not.
The honest verdict: 8-step is fine for delivery, 25-step is for the shot that gets cropped into.
Eight ControlNet modes from one adapter
The Union adapter handles depth, canny, pose, lineart, normal, tile, seg and one more from a single model file. That is the practical win β instead of eight ControlNet models to load, you swap the control image type.
The failure mode is at the edges of the control range: depth and canny are robust; tile and seg are much more sensitive to the input map's quality. A noisy depth map produces a worse result at 8 steps than at 25, because there are fewer steps to recover from a bad conditioning signal.
The Fix 1.0/2.0 pass
Qwen Image 2.1 has a specific artifact: invented high-frequency detail that looks like sharpening but is not present in the source. The Fix pass detects and removes it. This is the single most useful workflow in the set if you are upscaling old images.
Files
- HF: https://huggingface.co/AI-KSK/qwen21-controlnet-turbo-fix
- ModelScope: https://modelscope.cn/models/Sunjian520/qwen21-controlnet-turbo-fix
Video walkthrough (Chinese): https://youtu.be/4YEU8Ui6u1w
Bilibili: https://www.bilibili.com/video/BV14zp96REo9/
Notes
- Five ComfyUI workflow exports; Qwen Image 2.1 and the Union adapter must exist locally.
- Tutorial material, not a product.
- Independent AI-KSK release; not affiliated with the model authors.
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β full credit and traffic to the original publisher.