Updated March 2026. The original version of this cheat sheet was written for SD 1.5 in May 2023. Almost everything has changed since then -- new architectures (SDXL, SD 3.5, Flux), new UIs (ComfyUI), new hardware (RTX 5090), and a complete reversal on negative prompt philosophy. This is the current version.
This is my working reference for Stable Diffusion parameters. Not a tutorial -- just the settings I reach for when things aren't working or when I want to push quality.
Which Model to Use
This is the first decision now, and it matters more than any parameter tweak.
| Model | Best For | Resolution | Notes |
|---|---|---|---|
| Flux 2 | Photorealism, prompt adherence | 1024x1024+ | Best open-weight model for photorealism in 2026. Integrated into Adobe Photoshop [1] |
| SDXL | General use | 1024x1024 | Massive ecosystem of fine-tunes. Juggernaut XL, Realistic Vision, DreamShaper |
| SD 3.5 Large | Top quality (Stability's flagship) | 1024x1024 | MMDiT architecture. SD 3.0 was deprecated April 2025 [2] |
| SDXL Lightning | Speed | 1024x1024 | 2-8 step generation. Better quality than Turbo at higher resolution [3] |
| SD 1.5 | Legacy workflows | 512x512 | Huge fine-tune library but being phased out. SD 2.0/2.1 officially deprecated |
If you're starting fresh: Flux 2 for photorealism, SDXL for everything else. SD 3.5 is good but the ecosystem is smaller.
Which UI to Use
| UI | Best For |
|---|---|
| ComfyUI | Power users. Node-based, better VRAM management, 15% faster, best Flux support. Industry standard for serious work as of 2025 [4] |
| Automatic1111 | Beginners. Simpler interface, huge extension library. Still works fine for SDXL |
| Fooocus | One-click generation. Minimal configuration. Good for quick results |
I use ComfyUI. The learning curve is steeper (expect 10-20 hours to get comfortable), but the VRAM management alone is worth it -- it runs SDXL on 8GB where A1111 crashes.
Samplers
The sampler debate is mostly settled.
Go-to choices:
- DPM++ 2M Karras -- best speed-to-quality ratio. This is my default for almost everything.
- DPM++ SDE Karras -- slightly better at low step counts. Good when you're iterating fast.
- Euler a -- still reliable. More variety in outputs, good for exploration.
When to switch:
- Lack of diversity in outputs? Try DPM++ SDE or Euler a.
- Artifacts or oversaturation? Try DPM++ 2M Karras or plain Euler.
- Need speed above all? Euler a or DPM++ 2M (non-Karras).
- Want maximum quality? DPM++ 3M SDE Karras or UniPC.
Step counts: 20-30 steps for most samplers. Lightning models need only 2-8.
CFG (Classifier Free Guidance)
How strictly the model follows your prompt vs. its own interpretation.
| Range | Effect |
|---|---|
| 1-4 | Very creative, loose interpretation. Often incoherent |
| 5-7 | Good balance for most work |
| 7-10 | Strong prompt adherence. Sweet spot for SDXL photorealism |
| 10-15 | Risk of artifacts and overcooked colors |
| 15+ | Almost always too much. Artifacts guaranteed |
Note: SD 3.5 uses a different guidance mechanism. The CFG concept still applies but the scale behaves differently -- start lower (3-5) and adjust.
Resolution
The days of 512x512 are over.
| Model | Native Resolution | Practical Range |
|---|---|---|
| SD 1.5 | 512x512 | 512x512 to 768x768 |
| SDXL | 1024x1024 | 1024x1024 (standard), 1024x768, 768x1024 |
| SD 3.5 | 1024x1024 | 1024x1024+ |
| Flux | 1024x1024 | 1024x1024+, 4K possible on high-end GPUs |
Going above the native resolution risks artifacts and composition issues. Use hi-res fix or upscaling instead of generating at 2048x2048 directly.
Clip Skip
Less relevant than it used to be.
- SD 1.5: Clip skip 1-2 matters a lot. Anime models often use clip skip 2.
- SDXL: Uses dual text encoders (CLIP + OpenCLIP). Clip skip is mostly ignored -- the architecture handles it differently.
- SD 3.5 / Flux: Not applicable in the same way. These models use transformer-based text encoding.
If you're on SDXL or newer: don't worry about clip skip. If you're on SD 1.5: keep it at 1 for photorealism, 2 for anime.
Negative Prompts
The philosophy has flipped. In 2023, the advice was to use long negative prompt lists. In 2026, the consensus is: start with nothing and add only what you need to fix.
Why the change:
- SDXL and Flux understand natural language much better than SD 1.5
- Long negative prompts can actually restrict creativity and produce worse results
- "bad anatomy" is too vague to be useful. "ugly" doesn't work because SD wasn't trained on labeled "ugly" images
- Some models perform demonstrably worse with long negatives [5]
Current approach:
- Generate without any negative prompt first.
- If you see a specific problem (extra fingers, blurry background), add a targeted negative for that.
- Use emphasis weighting:
(blurry:1.3)instead of justblurry. - Keep it short -- 5-10 terms max.
GPU Quick Reference
| GPU | VRAM | Good For |
|---|---|---|
| RTX 3060 12GB | 12GB | SD 1.5, basic SDXL |
| RTX 4070 Ti | 12GB | SDXL, some Flux |
| RTX 4090 | 24GB | Everything. The workhorse |
| RTX 5090 | 32GB | Everything including 4K and batch generation |
| 8GB cards | 8GB | Minimum viable. ComfyUI helps with VRAM management |
The 24GB mark is where things get comfortable for SDXL and Flux without constant VRAM juggling.
Troubleshooting Quick Fixes
| Problem | Try |
|---|---|
| Blurry output | Increase steps. Check resolution matches model's native res |
| Extra fingers/limbs | Add extra fingers, extra limbs to negative prompt. Or use ControlNet |
| Oversaturated colors | Lower CFG. Switch to DPM++ 2M Karras |
| Composition is wrong | Use ControlNet (depth, canny, pose) instead of fighting the prompt |
| Generation is slow | Use Lightning model, reduce steps, use ComfyUI for better VRAM |
| Out of VRAM | Switch to ComfyUI, reduce batch size, use fp16 |
References
1. Flux 2 and NVIDIA RTX AI Integration -- NVIDIA's coverage of Flux 2 with ComfyUI.
2. Stability AI Release Notes -- SD 3.0 deprecation and 3.5 release.
3. SDXL-Lightning by ByteDance -- 2-8 step generation at 1024px.
4. ComfyUI vs Automatic1111 2026 Comparison -- Performance and feature comparison.
5. How to Use Negative Prompts Effectively -- Updated guide on minimal negative prompt philosophy.
6. Understanding Stable Diffusion Samplers -- Sampler comparison and selection guide.
7. Best Stable Diffusion Models for 2026 -- Current model landscape.
Related Posts
- Stable Diffusion Photorealism: Settings & GPU Limits Guide -- deep dive into achieving photorealistic results with current models.
Comments