machin-diffusion-scheduler

name	machin-diffusion-scheduler
description	CRITICAL — SD-Turbo scheduler correctness. The

SD-Turbo Scheduler Correctness

The #1 bug we hit

Using timestep_spacing: "leading" (the diffusers default for some schedulers) produced abstract noise instead of images. The UNet was told the input was nearly clean (timestep=1, sigma~0.04) while supplying high-noise latents, so it left the noise mostly intact.

Fix: SD-Turbo requires timestep_spacing: "trailing".

Correct SD-Turbo scheduler config

timestep_spacing: "trailing"
steps_offset: 1
num_train_timesteps: 1000
beta_schedule: "scaled_linear"
beta_start: 0.00085
beta_end: 0.012
prediction_type: "epsilon"
use_karras_sigmas: false

One-step values

• Timestep: 999
• Sigma: ~14.6146
• init_noise_sigma: ~14.6146

Four-step values (for reference)

• Timesteps: [999, 749, 499, 249]
• Sigmas: [14.6146, 4.08173, 1.61289, 0.693205, 0.0]

One-step Euler operation

latent = noise * 14.6146
unet_input = latent / sqrt(14.6146² + 1)
noise_pred = UNet(unet_input, timestep=999, text_embeddings)
latent = latent - 14.6146 * noise_pred
latent = latent / 0.18215          # VAE scaling factor
VAE.decode(latent)

General multi-step epsilon prediction

latent = randn * init_noise_sigma

for each step:
    model_input = latent / sqrt(sigma² + 1)
    eps = UNet(model_input, timestep, embeddings)
    latent = latent + eps * (sigma_next - sigma)

latent = latent / 0.18215
VAE.decode(latent)

The carried state is the UNSCALED latent, not the model input and not x0_hat.

Guidance

• guidance_scale=0.0 is the official recommended setting for SD-Turbo.
• SD-Turbo's ADD (Adversarial Diffusion Distillation) incorporates guidance behavior.
• Do NOT add CFG for the one-step path.
• Test with guidance_scale=7.5 produced high-contrast, poor results.
• Negative prompts have no effect at 1 step.

How to verify the scheduler

If you're debugging noisy output, check:

1. timestep_spacing is "trailing" (not "leading")
2. One-step timestep is 999 (not 1)
3. Sigma is ~14.6146 (not ~0.04)
4. guidance_scale is 0.0
5. The latent is divided by 0.18215 before VAE decode

Reference

The reference pipeline is at /tmp/ref_pipeline.py (diffusers/PyTorch with the correct trailing scheduler). Validate against it stage-by-stage.