Diffusion Parameters

Choose model-specific settings and compare one controlled change at a time

Last updated: 2026-10-27

Start with the selected model's defaults. Steps, guidance, strength, scheduler, precision and available controls differ between families and between ordinary and distilled/turbo variants. A preset for SDXL is not a universal recipe for FLUX, Qwen or a specialized Create source.

Controls that matter first

Control What to check
Model / pipeline Installed compatible weights and the active production/Alpha label. Changing family can remove unsupported modules or controls.
Steps How much sampling work to request. More steps take longer and do not always improve distilled/few-step models. Start with that model's recommendation.
Guidance / CFG Prompt influence where supported. Some models fix it or use a different guidance path; a hidden/disabled control is not a missing feature to force on.
Denoising strength How far processing may change the source. Compare a lower-strength test when preserving structure; high values can replace more of the input.
Seed Fix it for comparisons, randomize for exploration. The same seed alone does not guarantee identical output across changed models, software, settings or hardware.
Template / tile size Sphere coverage and per-tile work. Use a modest first task and inspect boundaries/poles. More faces do not by themselves prove lower memory use.
Batch / placement / precision Workload and memory tradeoffs, only where the selected model exposes them. Fit estimates do not reserve VRAM; offload also needs system RAM.

A useful comparison

  1. Confirm model readiness and choose a small input/template.
  2. Write a short global prompt. Use negative prompts only on models that support them.
  3. Render a baseline with the model's defaults and a fixed seed.
  4. Change one relevant control and compare the resulting graph steps. Keep the better version, or return to the earlier one.
  5. Save the working Workflow before making a large change.

Model-specific differences

  • SD 1.5, SD 2.1 and SDXL commonly expose ordinary steps/CFG and negative prompts; even within SDXL, distilled variants need different schedules.
  • SD 3.5, FLUX, Qwen, Chroma and Z-Image have their own supported guidance/scheduling, negative-prompt and precision rules. Trust the selected model's visible controls and recommendations rather than transplanting numeric ranges.
  • 360 LoRA generation settings belong to its selected base/adapter. Outpaint currently fixes its steps and guidance. These create starting images; later tile processing has its own configuration.
  • Edit, projection conversion and procedural creation do not need diffusion sampling controls merely because they appear alongside diffusion pipelines.

See Pipelines for supported families, Model Management for setup, Prompts for prompt scope and Multipass for several distinct processing passes.

Esc
Searching...
No results found.
Type to search the documentation
↑↓ Navigate ↵ Select Esc Close

We use cookies for analytics to understand how visitors use our site. Privacy Policy