Fast SDXL: Turbo, Lightning & Hybrid Checkpoints
SDXL Turbo and Lightning use 1–8 steps with CFG=0. Here is what each distillation method imposes and what you trade away.
On this page
Fast SDXL: Turbo, Lightning & Hybrid Checkpoints
TL;DR
- SDXL Turbo uses Adversarial Diffusion Distillation (ADD) to produce usable images in 1 step and reaches teacher-quality (SDXL Base) at 4 steps — but it requires guidance_scale=0.0 and cannot use negative prompts.
- SDXL Lightning distills SDXL Base into 1-, 2-, 4-, and 8-step variants; all require CFG=0 and an Euler scheduler with trailing timesteps. The 1-step variant is experimental and unstable.
- Hybrid checkpoints like Sleipnir TLH merge Turbo and Lightning distillation layers onto an SDXL HEL base — they handle 8+ steps with a high CFG range, making them more forgiving for ComfyUI workflows that need traditional CFG control.
- FameGrid is not a distilled variant at all — it is a standard SDXL fine-tune that needs 20–30+ steps with CFG 2–7, and its "instant" quality comes from curated training data, not few-step distillation.
How Few Steps Can You Really Use?
SDXL Turbo was trained to generate in a single network evaluation, meaning 1 step is the floor. Stability AI's model card states that "a single step is enough to generate high quality images" and that four steps further improves consistency and detail (SDXL Turbo). The ADD research paper confirms this: ADD-XL matches its teacher SDXL Base at 4 steps on 512×512 resolution, and 1-step output already outperforms prior few-step methods like LCM-XL (ADD paper).
SDXL Lightning extends the range. ByteDance provides checkpoints for 1, 2, 4, and 8 steps, with the 2-step, 4-step, and 8-step models described as having "amazing" generation quality. The 1-step model is explicitly marked "experimental" and "much less stable" — the authors recommend the 2-step variant for production work (SDXL Lightning).
The step count trade-off is not a cliff you fall off at step 4. ADD paper Table 1 shows FID scores (lower is better) of 20.6 at 1 step, 20.3 at 2 steps, and 20.3 at 4 steps — the quality gains from 2→4 steps are marginal, not dramatic (ADD paper). The main benefit of using more steps is deterministic consistency: the same seed produces more reproducible layouts and finer detail as steps increase.
Sleipnir [TLH] (Turbo, Lightning, Hyper) occupies a different niche. It is a community checkpoint merge on SDXL HEL (Ratatoskr Helheim) base. Its creator states it runs "very fast from min 8 Steps" and "can handle any step count and has a high cfg range" (Sleipnir TLH). The "Hyper" component's exact architecture is undocumented — treat it as a third distillation flavor distinct from ADD and progressive adversarial distillation.
The CFG Constraint Each Model Imposes
| Model | Steps | CFG | Scheduler | Notes | Source |
|---|---|---|---|---|---|
| SDXL Turbo | 1–4 | 0.0 | EulerDiscreteScheduler (default) | No negative_prompt support | HF |
| SDXL Lightning (2/4/8-step) | 2/4/8 | 0.0 | EulerDiscreteScheduler, trailing timesteps | LoRA or full UNet | HF |
| SDXL Lightning (1-step) | 1 | 0.0 | EulerDiscreteScheduler, trailing + prediction_type="sample" | Experimental, x₀ prediction | HF |
| FameGrid | 20–30+ | 2–7 | DPM++ 3M SDE / Uni Pc / DPM SDE; Karras scheduler | Standard SDXL, not distilled | Civitai |
| Sleipnir [TLH] | 8+ | High range | Not specified | Checkpoint merge; Hyper component undocumented | Civitai |
Why Turbo and Lightning Both Need CFG=0
Turbo and Lightning use distillation-based training. Turbo employs Adversarial Diffusion Distillation, which combines adversarial loss (to match the real image manifold) with distillation loss (to match a frozen teacher model's outputs). This hybrid objective replaces the need for classifier-free guidance at inference time — the student model learns to generate directly from noise. The ADD paper states explicitly that "during inference, our approach does not use classifier-free guidance" (ADD paper).
Lightning uses Progressive Adversarial Diffusion Distillation, which similarly distills the SDXL Base teacher into a student that can run in 1–8 steps. The Lightning model card warns: "Ensure using the same inference steps as the loaded model and CFG set to 0" for every step count variant (SDXL Lightning). Attempting to use CFG>0 with either model introduces a conditioning signal the student was never trained to interpret, which degrades output quality.
What Happens If You Ignore CFG=0
The source documentation is unambiguous: both Turbo and Lightning specify CFG=0 as a hard requirement, not a suggestion. Turbo's model card states "SDXL-Turbo does not make use of guidance_scale or negative_prompt" (HF). Lightning's code examples repeat guidance_scale=0 in every inference snippet for all four step variants (HF). No quantitative measurement of how much quality degrades at CFG>0 exists in authoritative sources.
What You Trade Away at Fewer Steps
Prompt alignment and detail
1-step ADD-XL outperforms LCM-XL at 4 steps in human preference for both image quality and prompt alignment (SDXL Turbo). But the ADD paper cautions that single-step samples, while already high quality, can show reduced attention to detail in complex compositions (ADD paper).
Iterative refinement capability
ADD paper states Turbo retains "the ability to improve results through iterative refinement," an advantage over one-step GAN approaches (ADD paper).
Mode coverage and diversity
SDXL Lightning paper discusses achieving "balance between quality and mode coverage" (Lightning paper). Distillation can reduce diversity for underrepresented subject types (Lightning paper).
Negative prompt support
Turbo drops negative prompts entirely. Lightning does not explicitly remove the mechanism, but since guidance_scale=0 is required, negative prompts have no effect regardless — the model's classifier-free conditioning is bypassed. If your workflow relies on negative prompts to suppress artifacts (e.g., extra fingers, watermark text), neither Turbo nor Lightning will honor them.
The FameGrid Case: What "Fast" Really Means
FameGrid is a standard SDXL fine-tune checkpoint, not a distilled few-step model. It requires 20–30+ diffusion steps with CFG 2–7, uses DPM++ 3M SDE or Uni Pc samplers with a Karras scheduler, and has three variants (Photo Real, Bold, UltraReal) (FameGrid).
The "ultra-real" quality FameGrid produces comes from training on a curated top-100-image subset and photorealistic fine-tuning — not from distillation. Treating FameGrid as a "fast" checkpoint is a category error.
The FameGrid workflow recommends ControlNet and Skin Detail node — but these are ComfyUI workflow choices, not model constraints (FameGrid workflow).
Common Errors and Fixes
| Error | Symptom | Cause | Fix |
|---|---|---|---|
| Using CFG>0 with Turbo or Lightning | Blurry, misaligned, or degraded outputs | Both models were trained without classifier-free guidance | Set guidance_scale=0.0 for both Turbo and Lightning |
| Using Turbo on 1024×1024 without resizing | Poor composition or OOM errors | Turbo's native resolution is 512×512 | Resize to 512×512 before inference; higher sizes work but are not native |
| Treating Lightning 1-step as equivalent to Turbo 1-step | Inconsistent quality | Different distillation methods and scheduler configs | Turbo 1-step uses standard Euler; Lightning 1-step requires Euler + trailing + prediction_type="sample" |
| Assuming "hybrid" means Turbo+Lightning mixed | Confusion about model architecture | "Hybrid" is a community merge label, not a formal distillation category | Verify each component of a merged checkpoint separately |
| Running FameGrid at 1–4 steps | Noisy, underdeveloped images | FameGrid is not a distilled model — it needs 20–30+ steps | Use 20–30 steps with CFG 2–7 |
FAQ
Q: How few steps can SDXL Turbo really use? A: 1 step. Turbo was designed for single-step generation and produces usable output at 1 step. Increasing to 2–4 steps refines consistency and detail, but the quality jump is marginal — the main trade-off is deterministic reliability, not dramatic improvement. A single step already approaches SDXL Base quality in human preference studies.
Q: What's the CFG constraint for SDXL Lightning? A: CFG must be 0, regardless of step count. Lightning removes classifier-free guidance through its distillation process. Using CFG>0 introduces a conditioning signal the student was never trained to process.
Q: Can Turbo and Lightning checkpoints be mixed in one workflow? A: No evidence supports mixing them. Turbo is an ADD student; Lightning uses progressive adversarial distillation with separate scheduler configs. They are architecturally incompatible.
Q: How many steps does FameGrid need? A: FameGrid is a standard SDXL checkpoint, not a distilled one. Use 20–30+ steps with CFG 2–7. Its "instant" quality comes from curated training data and photorealistic fine-tuning, not from few-step distillation.
Q: What is Sleipnir TLH and how does it differ? A: Sleipnir TLH is a checkpoint merge on SDXL HEL base that combines Turbo, Lightning, and Hyper distillation flavors. It supports 8+ steps with a high CFG range.
Q: Is Lightning's 1-step model ready for production? A: No. ByteDance labels the 1-step model experimental and less stable, and recommends the 2-step variant for production. The 1-step model also requires prediction_type="sample".
Q: Do I need a negative prompt with Turbo or Lightning? A: No. Turbo's model card states it does not use negative_prompt. Lightning's CFG=0 requirement means negative prompts have no effect. Use prompt engineering or post-processing instead.
See also our CFG scale guide for tuning guidance beyond these hard constraints. See our model merging guide for how hybrid checkpoints like Sleipnir TLH are constructed.