Sampler_Name Options

TL;DR: The sampler_name defines how the denoising process interprets and walks through the noise space during diffusion. Each algorithm has its own way of dealing with noise, speed, coherence, prompt alignment, and quirks. Think of these like different chefs following the same recipe—some are minimalist Michelin stars, others are heavy-metal grillmasters. Each affects your results.

🤖 Table of Samplers

Sampler	Best For	Recommended Scheduler	Notes	Strengths	Weaknesses
euler	Fast deterministic generations, previews	simple	Very predictable, sharp at high steps	Very fast and consistent	Can be harsh or noisy at high steps
euler_cfg	Same as euler with better prompt control	simple	Slightly better at following prompts	Adds better control to classic Euler	Slightly more resource usage
euler_ancestral	Textured, slightly chaotic results	simple	More randomness, good for art	Creates more textured, artistic outputs	Less predictable
euler_ancestral_cfg_pp	Textured creative results with better prompt adherence	simple	Adds CFG handling	Balanced creativity and prompt guidance	Complexity may increase render time
heun	Balanced generations, alternative to Euler	karras	Stable but slower	Stable and smooth results	Slower than Euler
heunpp2	Better quality than Heun, improved CFG	karras	Photorealistic scenes	Improved quality with prompt fidelity	Still underused and less tested
dpm_fast	Quick drafts and prototyping	exponential	Fastest among DPMs	Extremely fast rendering	Sacrifices image quality
dpm_adaptive	Quality-aware fast sampling	exponential	Adjusts internally for better output	Auto-balances speed and quality	Unpredictable output fidelity
dpmpp_2s_ancestral	Highly varied textures, creative images	karras	Good for expressive scenes	Great for varied creative imagery	May over-randomize details
dpmpp_2s_ancestral_cfg_pp	Creative + prompt fidelity	karras	More guided variation	Balances texture and prompt control	Slower due to CFG
dpmpp_sde	Clean gradients, realism	karras	Smooth transitions	Realistic gradients and transitions	Higher VRAM usage
dpmpp_sde_gpu	Same as above but faster on GPU	karras	GPU optimized	GPU-accelerated smooth rendering	Needs compatible hardware
dpmpp_2m	Balanced realism and speed	karras	Versatile for many styles	Balanced, great for realism	More steps needed than ancestral samplers
dpmpp_2m_cfg_pp	Detailed, prompt-loyal realism	karras	Most recommended general-purpose sampler	Best for realism with tight prompt control	Increased processing time
dpmpp_2m_sde	Even smoother realism	karras	Great for portraits	Extremely smooth, clean output	Slower with high step counts
dpmpp_2m_sde_gpu	Smooth realism, GPU-friendly	karras	High batch efficiency	Smooth realism, GPU-friendly	Needs VRAM headroom
dpmpp_3m_sde	Highest fidelity, complex detail	karras	Slow but premium output	Ultimate detail and fidelity	Very slow
dpmpp_3m_sde_gpu	High-detail GPU accelerated	karras	For batch high-end generation	High-detail GPU accelerated	Heavy on GPU resources
ddpm	Legacy and experimentation	ddim_uniform	Slow and stable	Stable and accurate to source diffusion	Very slow and outdated
lcm	Ultra-fast low-step generation	karras	Requires LCM-tuned models	Lightning-fast generation	Requires special models and low steps
ipndm	Experimental, high coherence	sgm_uniform	Use with caution	High detail and structure	Experimental and unstable
ipndm_v	Variation of IPNDM, smoother results	sgm_uniform	Experimental	More stable than ipndm	Still experimental
deis	Fast, lightweight quality	exponential	Compact generation	Quick generation with decent quality	Occasional prompt drift
res_multistep	Artistic, surreal images	normal	Needs higher steps	Dreamlike stylized art	Unstable and slow
res_multistep_ancestral	Dreamlike, unstable beauty	normal	Chaos-driven	Hyper-stylized art	Chaos-prone and unpredictable
re_multistep_ancestral_cfg_pp	Prompt-driven surrealism	normal	Slow and expressive	Controlled surrealism	High computational cost
gradient_estimation	Precision edge-case work	linear_quadratic	Very slow, niche	Precision where others fail	Exceptionally slow
gradient_estimation_cfg_pp	Same as above with prompt fidelity	linear_quadratic	Ultra-niche	Prompt-sensitive precision	Same slowness plus complexity
er_sde	Stable, smooth realism	karras	Balance of all factors	Balanced realism	Slower generation time
seeds_2	Internal, unknown use	normal	Undocumented	Possibly internal for seed processing	Undocumented use
seeds_3	Internal, unknown use	normal	Undocumented	Possibly internal for seed control	Undocumented use
ddim	All-purpose generation	ddim_uniform	Balanced across most needs	Fast, versatile, prompt-sensitive	Can lack texture or detail
uni_pc	High-quality, stable outputs	karras	Modern, robust sampler	High stability, great realism	Moderately slower

🧠 Detailed Sampler Breakdown

📌 Euler & Friends

• euler: The OG. Fast, low-memory, deterministic.
  • Use for: Fast previews, consistent outputs.
  • Avoid if: You want dreamy aesthetics.
• euler_cfg: Euler with better CFG control.
  • Use if: You find euler too rigid.
• euler_ancestral: Adds randomness for richer textures.
  • Use for: More creative, slightly less predictable results.
• euler_ancestral_cfg_pp: Post-prompt processing; blends chaotic charm with CFG wizardry.

---

⚙️ Heun Variants

• heun: Like Euler but tries to be smarter. A compromise between speed and precision.
• heunpp2: Heun, but updated with second-order CFG handling.
  • Pro tip: Works well for photorealism when Euler feels too harsh.

---

🚀 DPM Family (Euler on Steroids)

• dpm_fast: Speed demon. Sacrifices some quality for rapid generation.
• dpm_adaptive: Dynamically adjusts for better quality mid-run.
• dpmpp_2s_ancestral: Two-stage, good for varied textures. More artistic.
• dpmpp_2s_ancestral_cfg_pp: Same as above but with better prompt adherence.
• dpmpp_sde: Introduces SDE smoothing—great for clean gradients and realism.
• dpmpp_sde_gpu: GPU-optimized for large batches.
• dpmpp_2m: Two-mode version. Think “midpoint-aware” sampler.
• dpmpp_2m_cfg_pp: CFG-enhanced version. Best used with complex prompts.
• dpmpp_2m_sde: Even smoother. Mixes SDE and midpoint sampling.
• dpmpp_2m_sde_gpu: GPU-tuned for smoother multi-image workflows.
• dpmpp_3m_sde: Three-mode version. Slower but higher fidelity.
• dpmpp_3m_sde_gpu: GPU-optimized flavor for serious jobs.

🧪 Tip: The dpmpp samplers are your best bet for realism, complexity, and prompt fidelity. Try dpmpp_2m_cfg_pp with karras scheduler for S-tier output.

---

🧱 Basics and Benchmarks

• ddpm: The original reverse diffusion. Good for understanding the roots of it all but slow for production.
• ddim: A happy medium. Fast, decent quality, and widely supported.

---

🧠 Neural Wizardry

• lcm: Latent Consistency Models.
  • Use for: Insanely fast generation (think < 6 steps).
  • Note: Needs LCM-tuned models/checkpoints. Use low steps (4–6).
• ipndm, ipndm_v: Implicit noise prediction. High quality but needs babysitting.
  • Experimental: Try if you enjoy edge-case debugging.
• deis: High-speed lightweight solver. Not always accurate but shockingly fast.
• uni_pc: State-of-the-art. Combines stability with high detail.
  • Highly recommended for any polished workflow.

---

🎨 Artistic Samplers

• res_multistep: Applies restarts during denoising for richer style changes.
• res_multistep_ancestral: More chaotic cousin. Better for surrealism.
• re_multistep_ancestral_cfg_pp: If you must marry surrealism and prompt obedience.

---

🧮 Math Nerd Specials

• gradient_estimation: Gradient-based logic. Use when nothing else aligns.
• gradient_estimation_cfg_pp: Adds prompt handling, but slow. Niche.

---

🤖 Oddballs

• er_sde: SDE (Stochastic Differential Equation) method. Balanced but slow.
• seeds_2, seeds_3: Basically undocumented, possibly used internally or experimentally. Avoid unless testing.

---

🔥 What-Not-To-Do-Unless-You-Want-a-Fire

1. Using lcm with high steps.

   • Unless you're doing AI archaeology, LCM should be used at 4–6 steps. Higher values just waste time and return mushy nonsense.

2. Combining samplers and schedulers randomly.

   • Yes, technically you can combine euler with karras, but also technically, you can eat soup with a fork. Stick to the scheduler designed for the sampler, or expect... unpredictable spaghetti.

3. Assuming GPU samplers work on CPU.

   • *_gpu versions are not friendly with CPUs. Expect crashes, freezes, or your PC trying to roast marshmallows with its fan.

4. Trying gradient_estimation for casual generations.

   • If you like waiting 30 minutes for an image that looks like the same image every other sampler does in 30 seconds… be my guest.

5. Using experimental samplers in client work.

   • ipndm, ipndm_v, and seeds_* are not production-safe unless you enjoy gambling with broken renders mid-batch.

6. Using high-CFG samplers (*_cfg_pp) with lazy prompts.

   • These samplers expect detailed, descriptive prompts. If you feed them “portrait of woman,” expect an AI-generated meme instead of a masterpiece.

7. Assuming “ancestral” means “better.”

   • It means “more chaotic.” Sometimes that works. Sometimes it gives you spider limbs.

8. Ignoring your VRAM budget.

   • Samplers like dpmpp_3m_sde_gpu will happily devour 12GB+ of VRAM like it’s brunch. Don’t say we didn’t warn you.

---

🧪 Best Pairings

Sampler	Scheduler	Suggested Steps	Notes
dpmpp_2m_cfg_pp	karras	25–35	Excellent balance of speed/detail
uni_pc	karras	20–30	Works well across styles
lcm	lcm	4–6	Ultra fast, only with LCM-ready models
ddim	exponential	20–30	Great for soft, coherent results
euler_ancestral	simple	20–40	Artistic, textured output

Want to learn more about schedulers? Check out even more documentation here!

💥 Summary

Choosing the right sampler_name is like picking the right brush for your AI-generated masterpiece. If you’re running basic txt2img? Euler or DDIM. Want hyper-realism? Go dpmpp_2m_cfg_pp. Want it now and fast? lcm is your guy.

Just don’t pick gradient_estimation_cfg_pp and wonder why your render time rivals the development of Elden Ring.