ComfyUI Node

🎲 LDM Sampler

Runs the diffusion sampling loop to denoise the latent. Samplers (alpha-space, native to the LDM paper): • ddim — deterministic DDIM (eta controls stochasticity: 0=det, 1≈ddpm); • ddpm — ancestral DDPM (always stochastic); • plms — pseudo-linear multistep. Compatible schedulers: ddim_uniform, linear, beta. Samplers (sigma-space, k-diffusion family): • euler / euler_ancestral / heun / dpmpp_2m. Compatible schedulers: all (ddim_uniform, linear, beta, karras, exponential, cosine). denoise=1.0 generates from pure noise (txt2img). Values < 1.0 partially denoise an existing latent (img2img). NOTE: This model was trained without classifier-free guidance. There is no negative prompt and no CFG scale.

By Winlensky·Created a day ago·Updated about 12 hours ago· 1
🎲 LDM Sampler
  • model
  • positive
  • latent_image
  • LATENT
seed42
steps50
sampler_name
scheduler
eta0.00
denoise1.00
CategoryLDM

Inputs (9)

NameTypeDefaultDescription
modelLDM_MODELMODEL output from Load LDM Checkpoint.
positiveLDM_CONDITIONINGCONDITIONING from BERT Text Encode node.
latent_imageLDM_LATENTStarting latent: empty (txt2img) or encoded image (img2img).
seedINT420–18446744073709550000Random seed for initial noise.
stepsINT501–1000Number of denoising steps. k-diffusion samplers converge well at 20–30 steps.
sampler_nameCOMBOddim/ddpm/plms are alpha-space (LDM paper native). euler/heun/dpmpp_2m are sigma-space (k-diffusion).
schedulerCOMBOSigma schedule. Alpha-space samplers (ddim/ddpm/plms) accept: ddim_uniform, linear, beta. Sigma-space samplers accept all schedulers.
etaFLOAT0.000–1Stochasticity for DDIM only. 0 = deterministic, 1 ≈ DDPM. Ignored by other samplers.
denoiseFLOAT1.000–11.0 = full generation from noise; < 1.0 = partial denoising (img2img).

Outputs (1)

NameTypeDescription
LATENTLDM_LATENT