Updated: 2026-05

1. What You’ll Learn on This Page

Adjust the K-Sampler parameters (Step, CFG, Sampler, Seed) one by one to see how the results change. Use the same workflow you used on the previous page (Minimum Workflow).

The goal here isn’t to “memorize the correct combinations,” but to get a feel for which parameters affect what.

2. Seed

Even with the same prompt and the same parameters, a different seed will produce a different image. Conversely, if the seed is the same, you will get exactly the same result no matter how many times you run it.

2.1 Give it a try

1. Change the “Post-generation control” setting in K-Sampler from randomize to fixed
2. Enter any number in the seed field and run the program
3. Run it again → The same image appears
4. Change the seed to a different number and run the program → A different image appears

2.2 When to Use Seeds

• Reproducibility: If you want to regenerate “that image” → You can reproduce it if you record the seed
• Comparative Experiments: When changing parameters to compare results, you must fix the seed for the comparison to be valid
• Variations: If you want to explore different color variations using a prompt you like, just change the seed

When conducting comparative experiments in class, be sure to secure the seeds.

3. Number of steps

How many times should the noise be reduced?

3.1 Try it out

Run the program with the seed fixed, varying only the number of steps.

Number of steps	Result
5	Blurriness, graininess, and distortion are noticeable
10	Shapes are visible but rough
20	Standard quality
30–40	More detailed, though the difference from 20 may be subtle
50 or more	Cost-effectiveness decreases; the difference is particularly hard to see on standard models

3.2 Exceptions: Turbo / LCM Models

Z Image Turbo, SDXL Turbo, Flux schnell, SD Turbo, and similar models are designed with a “few-step” approach in mind. These models achieve satisfactory results within 4 to 8 steps. Increasing the number of steps beyond that does not significantly improve the results.

4. CFG (Classifier-Free Guidance)

Fidelity to the prompt.

4.1 Let’s Try It

Keep the seed and number of steps fixed, and vary only the CFG.

CFG	Result
1	Virtually unrestricted; the AI draws freely
3	Vaguely follows the prompt
7–8	Typical usage range
12	Excessively faithful to the prompt; colors begin to appear oversaturated
15 or higher	Artifacts become noticeable; results break down

4.2 Model Compatibility

• SD 1.5 series: CFG 7–8 is standard
• Flux dev: Designed to operate at CFG 1–3 (higher CFG values cause instability)
• Turbo / LCM series: CFG 1–2 is recommended

Recommended node values vary by model, so be sure to check the model card (model description) when using a new model.

5. Sampler (sampler_name)

An algorithm for reducing noise.

5.1 Major Samplers

Sampler	Features	Recommended Use
euler	Classic and stable	Use this if you’re unsure about SD 1.5
euler_a	A variant of euler; highly sensitive to the seed	When you want to generate diversity
dpmpp_2m	Newer, high quality	When aiming for medium to high quality
dpmpp_2m_sde	Stochastic version, high quality even with fewer steps	For good results in a short time
lcm	For LCM-based models only	When using LCM-based models

5.2 Try it out

Keep the seed fixed and change only the sampler. Even with the same prompt, the atmosphere changes depending on the sampler (this is particularly noticeable when comparing euler and euler_a).

6. Scheduler

A strategy for determining how much noise to reduce at each step. Works in conjunction with the sampler.

Scheduler	Characteristics
normal	Standard, nearly linear
karras	Fine-grained in the latter half, quality-oriented
exponential	Fast in the early stages
simple	Simple, lightweight

If you’re unsure, use normal. If you want to improve the quality at SD 1.5, use karras.

7. The Effect of Resolution

The width × height of the “Empty Latent Image” node.

• Small (512×512): Fast, low credit consumption, but poor detail rendering
• Medium (768×768–1024×1024): Good balance
• Large (1536×1536 or higher): High quality, but computationally intensive with high credit consumption

Since SD 1.5 was trained at 512×512, it tends to produce distorted results at other sizes. For SDXL and Flux, 1024×1024 is the standard. Results are most stable at resolutions close to the training resolution.

8. Comparison Exercises (For Class Use)

To save credits, the comparison test will be conducted using SD 1.5 (0.3–0.5 credits per disc).

Exercise A: The Effect of the Number of Steps

• Same prompt, same seed, CFG 7, euler/normal
• Run the model four times with 5, 10, 20, and 40 steps
• Compare the results side by side to see at what point the differences become virtually indistinguishable

Exercise B: The Impact of CFG

• Same prompt, same seed, 20 steps, euler/normal
• Run CFG 5 times with values of 1, 3, 7, 12, and 20
• Observe where oversaturation and image degradation occur

Exercise C: Seed Variation

• Same prompt, 20 steps, CFG 7, euler/normal
• Run the model using four different seeds
• Experience firsthand how much variation can occur even with the same prompt

Exercise D (Optional): Sampler Comparison

• Same prompt, same seed, 20 steps, CFG 7
• Comparison using euler / euler_a / dpmpp_2m / dpmpp_2m_sde

Depending on the available class time, Exercises A through C will be required, while Exercise D will be an advanced assignment.

9. Saving and Organizing Results

When comparing images generated with different parameters, you need to keep track of each image’s metadata (prompt, seed, step, and CFG).

Comfy Cloud embeds the metadata of generated images into the image files. If you drag and drop the saved PNG file directly onto the workflow screen, the entire workflow from that point will be restored.

Tip: If you get a result you like, save the image first (right-click → Save Image). That way, you’ll be able to recreate it perfectly later if you forget how you made it.

10. What’s Next

• img2img / inpaint — Uses an existing image as a starting point
• ControlNet — Specifies composition and pose using a separate image
• LoRA — Narrows down the art style or subject