The sampler's settings

<!-- lesson: sampler-fields -->

Lesson "sampler-fields", in the group Pictures: first steps of the lesson list. Steps: 10.

#What you will be able to do

You set the fields of the sampler the way ComfyUI's own SDXL template sets them, and learn what each one does.

#What you need

• It goes on from A first picture from text. The card says so; it is a hint, not a lock.
• Nodes: Load Checkpoint, Empty Latent Image, KSampler, Save Image / Preview Image.
• The app downloads and installs nothing. The step texts below say which files and folders, and which ComfyUI template names them.

#The steps

The words of each step, as the lesson shows them. Node names are in bold.

1. Fields everybody meets (reading)

   The sampler is where the picture is made, and its fields are the ones you will change most. This lesson goes through them on the graph of "A first picture from text".

   The values are those of ComfyUI's own SDXL template. For SD 1.5 they work too. Newer families (Flux, Z-Image, Qwen-Image, Wan) want other numbers: their templates say which.

2. The picture graph is needed

   The canvas must have the graph of "A first picture from text": Load Checkpoint, Empty Latent Image, KSampler and Save Image.

   If it is not there, do that lesson first.

3. The seed (reading)

   The seed field is the number the starting noise is made from. The same seed, the same settings and the same prompt give the same picture.

   Under it in ComfyUI there is control_after_generate: what happens to the number after each run. randomize takes a new random number, increment adds one, fixed keeps it.

   Take randomize while you look for ideas. Take fixed while you tune one picture: then you see what each change did.

4. Steps

   Set steps to 25.

   The sampler removes the noise in this many steps. More steps, a cleaner picture and a longer wait, but after about 30 the picture hardly changes. Distilled and turbo models need only 4 to 8 steps: their pages say so.

5. CFG

   Set cfg to 7.

   It says how strictly the model follows the prompt. Higher follows more closely but burns the colours and hardens the picture; lower is softer and freer. For SD 1.5 and SDXL keep 5 to 8.

   Distilled models and speed-up LoRAs want cfg 1, which also halves the time.

6. Sampler and scheduler (reading)

   Two fields choose how the noise is removed.

   • sampler_name, the method. euler is simple and reliable. dpmpp_2m gives more detail in the same steps. Names with ancestral or sde add fresh noise at every step: livelier, but the picture never settles and changes with the number of steps.
   • scheduler, how the noise is spread over the steps. karras spends more steps on fine detail and suits the dpmpp methods; normal and simple spread them evenly and suit most new families.

   For a start take a pair from a template and change one thing at a time.

7. Choose the pair

   Set sampler_name to dpmpp_2m and scheduler to karras.

   This is the pair of ComfyUI's SDXL template and of many guides for SD 1.5 and SDXL.

8. Denoise

   Leave denoise at 1.

   It says how much of the starting latent is replaced with noise. With an empty blank everything is drawn anew, so 1. Below 1 is for redrawing a picture that comes in: image to image, a second pass.

9. Size and batch (reading)

   The size is set in Empty Latent Image: width and height. Make the picture near the size the model was trained at: 512 by 512 for SD 1.5, about 1024 by 1024 for SDXL. Far from it the model repeats things or breaks the composition.

   Other shapes with the same number of pixels work: 832 by 1216 for a portrait, 1216 by 832 for a landscape.

   batch_size makes several pictures in one run. It needs that many times the memory; 1 is the safe start.

10. Try it (reading)

Save the graph with "File ▸ Save workflow for ComfyUI…", open the file in ComfyUI and run it there. This app runs nothing in ComfyUI itself.

In ComfyUI set control_after_generate to fixed, run once, then change only steps or only cfg and run again. Seeing one change at a time is the fastest way to learn what each field does.

#What the lesson checks

A reading step is done when you press Got it, next. A doing step ticks itself off when the canvas holds:

• Step 2, The picture graph is needed:
  • a Load Checkpoint node.
  • a Empty Latent Image node.
  • a KSampler node.
  • a Save Image or Preview Image node.
  • MODEL of Load Checkpoint into model of KSampler (a Load LoRA or LoraLoaderModelOnly in between is accepted).
  • LATENT of Empty Latent Image into latent_image of KSampler.
• Step 4, Steps:
  • steps of KSampler is 25.
• Step 5, CFG:
  • cfg of KSampler is 7.
• Step 7, Choose the pair:
  • sampler_name of KSampler is dpmpp_2m.
  • scheduler of KSampler is karras.
• Step 8, Denoise:
  • denoise of KSampler is 1.

#See also

• Sampler and sampler names
• Scheduler
• Steps
• CFG and guidance
• Seed
• Denoise
• How lessons work