# Skill: Tune KSampler settings to fix mushy, fried, or unrepeatable images

> Lets an AI agent diagnose and correct KSampler configuration (sampler, scheduler, steps, CFG, seed, denoise) in ComfyUI -- whenever a generated image is blurry, over-saturated, artifact-ridden, or not reproducible.

Source: sodigi·learn -- comfyui/semplery-shagi-cfg · https://sodigi.io/learn/comfyui/semplery-shagi-cfg

## When to use
- Generated image is blurry, mushy, or undercooked.
- Colors look acid, burnt, or edges are breaking up.
- User wants to reproduce a successful result exactly.
- User asks what sampler, scheduler, steps, CFG, or denoise do.

## Core rules
- Safe default set: sampler = dpmpp_2m, scheduler = karras, steps = 25, CFG = 7, denoise = 1.0.
- Change one knob at a time -- changing multiple simultaneously makes it impossible to isolate cause.
- Steps sweet spot is 20-30; above 40 gives almost no visible improvement and just wastes time.
- CFG sweet spot is 6-8; above 10-12 almost always produces fried artifacts and acid colors.
- Denoise must stay at 1.0 for text-to-image; only lower it deliberately for img2img (reworking an existing image).
- Same seed + same settings = identical output; lock the seed once a composition is found, then refine.
- Euler (or euler_a) is faster and more "creative"; dpmpp_2m is smoother and better for realism and detail.
- Scheduler karras pairs best with dpmpp_2m; leave it there until you have a reason to experiment.

## Procedure
1. Start with the safe default set (above) before touching anything else.
2. Generate once. Evaluate the result.
3. Image mushy / undercooked: raise steps to 30 first; if still mushy, raise CFG to 8.
4. Image fried / acid colors / broken edges: lower CFG toward 5.
5. Found a composition you like: lock the seed immediately; change only one other setting next.
6. Want a different character or feel: swap sampler between dpmpp_2m and euler; leave everything else.
7. Record the winning combination (sampler + scheduler + steps + CFG + seed) so you can reproduce it.

## Ready-to-use prompt
```
I am tuning KSampler settings in ComfyUI. Apply this default starting configuration and generate a test image:

Sampler: dpmpp_2m
Scheduler: karras
Steps: 25
CFG: 7
Seed: random (for exploration)
Denoise: 1.0

After the first result:
- If the image is mushy or lacks detail, try steps 30 and CFG 8 (one change at a time).
- If the image has acid colors or harsh edges, try CFG 5.
- Once a good result appears, lock the seed and describe what you changed.

Prompt to test with: [your prompt here]
```

## Pitfalls
- Setting steps to 60-100 "for quality" -- after 30 the gain is negligible; you only heat the GPU.
- Setting CFG to 12+ "so it definitely follows the prompt" -- usually produces artifacts, not compliance.
- Lowering denoise below 1.0 during text-to-image -- produces vague mush that is not the model's fault.
- Changing sampler, scheduler, steps, and CFG all at once -- impossible to diagnose which change helped.
- Not locking a successful seed -- the same prompt with a random seed will never produce the same image again.
- Exploring exotic samplers at the bottom of the list on day one -- spend that time on the two reliable ones first.