DEV Community

AI-KSK
AI-KSK

Posted on Edited on Originally published at huggingface.co

High-Noise Locks Structure, Low-Noise Explodes Style: The Split That Brings Back the Fun

One-pass generation forces a trade: the sampler either spends its budget on structure or on style. Splitting high-noise and low-noise passes lets each do one job, and the result feels different in a way that is hard to describe but easy to see.

The two graphs

Workflow Nodes Base
2512 two-step 21 qwen_image_2512_bf16
Base + Turbo segmented 39 z_image_turbo_bf16

The two-step 2512 graph

Node Value
UNETLoader qwen_image_2512_bf16.safetensors
CLIPLoader qwen_2.5_vl_7b_fp8_scaled.safetensors, type qwen_image
VAELoader qwen_image_vae.safetensors
LoraLoaderModelOnly Wuli-Qwen-Image-2512-Turbo-LoRA-2steps-V1.0-bf16 at 1.0

Two steps. That is the extreme end of distillation, and it works because the LoRA was trained specifically for two-step generation on this base.

Two steps means the model makes two decisions total. Everything in the output is one of those two decisions, so prompt wording matters disproportionately — there is no sampling budget to recover from a vague instruction.

The segmented graph

Node Value
UNETLoader z_image_turbo_bf16
CLIPLoader qwen_3_4b, type lumina2
VAELoader ae.safetensors

39 nodes for a two-pass structure with a Turbo low-noise pass.

What the split actually does

High-noise pass — the sampler starts from near-pure noise and makes large structural decisions. Composition, subject placement, large shapes. Because the noise level is high, small changes are cheap and the model can commit to a layout.

Low-noise pass — starts from the high-noise result with only a little noise added. It cannot restructure; it can only refine. Style, texture, and detail happen here.

The reason this produces a different result from one pass: the low-noise pass has a coherent structure to stylise. A single pass has to trade structure against style at every step, which produces a compromise. Two passes let the style pass be aggressive without risking the layout.

That aggressiveness is what the title means by "the fun comes back" — a Turbo low-noise pass can push style hard because it cannot break the composition.

Where this breaks

  1. Two steps offers no correction room. A bad prompt produces a bad image with no intermediate stage to fix it.
  2. High-noise must be right. If the structure is wrong, the low-noise pass will stylise a wrong image beautifully.
  3. CLIP type differs between graphs — qwen_image for 2512, lumina2 for Z-Image. Same encoder family, different projection.
  4. Two model families means two sets of models on disk.
  5. 21 nodes has minimal validation.

Practical approach

Write the prompt for the high-noise pass as if it were the only prompt — that is where layout comes from. Then treat the low-noise pass as a style treatment. If the result looks good but wrong, the problem is upstream in step one.

Workflows

Two graphs, published on RunningHub with invite code rh-v1111:

  • 2512 two-step
  • Z-Image Base + Turbo segmented rendering

Files

https://huggingface.co/AI-KSK/ksk-video-169-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 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)