Chroma QuickSilver – v1.2

Fast | Accurate toggle

Accurate – raw Chroma1-HD

Fast – flash delta-weight model applied (≈ low-step weights similar to LoRA, ~50 % fewer steps)

Sampler toggle

KSampler – standard

NAG – “negative-guidance”; add tokens like “hands” or “cold lighting” to suppress them. OOTB configured with a regular negative prompt to be used with Fast mode (fast mode disables the model's negative prompt)

What’s new since v1.1

Several changes to mimic settings in the recent example workflow provided by Lodestones

• Default sampling config now a custom beta scheduler setup

• Default CFG to 3.8

• Applied model shift of 1.0

ControlNet via FLUX.1-dev-ControlNet-Union-Pro

The workflow has an example DepthAnything node allowing you to quickly upload a reference image.

You still need a solid prompt or you'll get garbage.

What’s new since v1.0

• default steps / cfg adjusted:

  • Accurate: 40 steps / 3.5 cfg

  • Fast: 16 steps / 1 cfg

• resolution selector node – one-click aspect-ratio swap

• mode switch cleaned up – single selector drives model / clip

• placeholder LoRA slot – drop in your own when ready

• Save Image with Metadata node – Civitai reads prompt, seed, etc. automatically on upload

Use pattern basics (detailed recommendations inside workflow):

• Fast/NAG to create and refine prompt.

• Accurate+KSampler for final gens.

• Accurate/NAG if you need stronger model guidance and negative attention on image features. (Try turning on both to compare!)

Inside the workflow you’ll find notes on swapping the fast config (large delta-weight) for a flash LoRA or any fine-tuned flash model.

Tip: Chroma1-HD is not distilled, so generation is slower. Turn on ComfyUI’s “Preview” mode to catch bad anatomy or prompt drift early and cancel before the full render.