EasyKSampler (SDTurbo)

Turbo models are trained differently from a normal checkpoint, and they punish you for treating them normally. SD Turbo and SDXL Turbo are distilled models - trained with an adversarial loss on top of score distillation so they can produce a usable image in as few as 1-4 steps instead of the usual 20-50. That speed comes from baking the classifier-free guidance into the model during training, which means applying CFG again on top of that at inference time doesn't help - it actively hurts, pushing images toward oversaturated, artifacted results. easy kSamplerSDTurbo exists so you don't have to remember all of that every time: it's the Easy-Use sampler shaped for Turbo's settings rather than a standard model's.

Why a dedicated sampler for this

If you tried to run a Turbo checkpoint through a normal 20-step, CFG-7 sampler, you'd get a mess - not because Turbo is broken, but because you'd be applying settings tuned for a completely different training regime. A dedicated node nudges you toward the right defaults instead of making you rediscover, the hard way, that Turbo wants roughly 1-4 steps and a CFG close to the model's own baked-in guidance rather than the 5-9 range that's normal for SD 1.5/SDXL.

How it works

Same pipe-driven pattern as every sampler in this pack: it reads model, prompts, latent, and the sampling configuration off the pipe, which your loader and presampling nodes set up upstream. The distinguishing part isn't visible on the node itself - it's that this sampler is meant to be paired with a pipe configured for Turbo's low-step, low(er)-CFG regime, rather than one set up for a standard checkpoint.

The inputs and outputs that matter

• pipe - the required wire carrying model, prompts, latent, and sampling settings. Make sure whatever set those settings upstream actually configured them for Turbo (few steps, appropriate sampler/scheduler) - this node runs what the pipe tells it to, it doesn't override your step count for you.
• image_output - Hide, Preview, Save, Hide&Save, Sender, or Sender&Save, the standard output-routing choice shared across Easy-Use samplers.
• link_id / save_prefix - the usual Sender-pairing and filename-prefix fields.

Optional model overrides the pipe's model at the last moment. Outputs: pipe (pass-through) and image.

Installing it

Ships with the base pack. ComfyUI Manager: search ComfyUI Easy Use, install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/yolain/ComfyUI-Easy-Use

then install.bat on Windows or pip install -r requirements.txt, restart. The sampler itself needs no extra download - but you do need an actual SD Turbo or SDXL Turbo checkpoint loaded upstream; this node doesn't turn a normal checkpoint into a Turbo one.

Common issues & troubleshooting

Images look blown out, oversaturated, or artifacted. This is the classic Turbo mistake: CFG set too high. Turbo's guidance is baked in during training, so a low CFG close to the model's intended baseline (roughly 1-2 for these adversarially-distilled models) is what you want, not the 5-9 range that's normal for SD 1.5/SDXL. If your pipe's presampling node still has a "normal" CFG configured, that's almost certainly the problem.

Running way more steps than the model needs. A distilled model trained for 1-4 steps doesn't get better with 20 - it was trained to denoise in a small number of large jumps, and forcing more/smaller steps than it expects tends to produce artifacts rather than extra polish. Keep step count in Turbo's expected range.

Wrong scheduler for the model. Distilled models are typically trained against a specific noise schedule shape. If your results look off even with step count and CFG corrected, check that the scheduler on your pipe matches what the specific Turbo checkpoint's documentation recommends rather than defaulting to whatever you'd normally use on a standard SDXL model.

Licensing. SD Turbo and SDXL Turbo both ship under a non-commercial research license - more restrictive than most checkpoints you'd casually build a paid workflow around. Check the specific checkpoint's license before using its output commercially.