Schedulers¶

Swap denoising schedulers at runtime. 10 aliases mapped to diffusers classes.

from strands_sana import sana_set_scheduler, sana_list_schedulers

print(sana_list_schedulers())

sana_set_scheduler(name="dpm-solver", model="sana-1.6b-1024")

Available schedulers¶

Alias	Diffusers class	Notes
`flow-match-euler`	`FlowMatchEulerDiscreteScheduler`	Default — best for Sana
`flow-match-heun`	`FlowMatchHeunDiscreteScheduler`	Higher fidelity, 2× cost
`dpm-solver`	`DPMSolverMultistepScheduler`	Fast, supports flow_sigmas
`dpm-solver-single`	`DPMSolverSinglestepScheduler`	Even faster
`dpm-solver-cosine`	`CosineDPMSolverMultistepScheduler`	Cosine schedule
`euler`	`EulerDiscreteScheduler`	Lossy on flow models
`euler-ancestral`	`EulerAncestralDiscreteScheduler`	Adds noise per step
`ddim`	`DDIMScheduler`	Lossy on flow models
`deis`	`DEISMultistepScheduler`	Fast
`heun`	`HeunDiscreteScheduler`	Lossy on flow models

Demo — same prompt, 4 schedulers¶

Flow-matching compatibility¶

Sana is a flow-matching model. Its scheduler config has prediction_type=flow_prediction.

✅ FlowMatch* schedulers — native fit
✅ DPMSolver* schedulers — accept use_flow_sigmas=True
⚠️ Other schedulers — work but quality degrades (we strip flow_prediction automatically and warn)

This was BUG #9 — caught on Thor.

Speed comparison (Thor, sana-0.6b-1024, 10 steps)¶

Scheduler	Time
flow-match-euler	48s (first run, ~12s warm)
dpm-solver	13.3s
euler	6.2s
euler-ancestral	6.3s
ddim	7.0s
deis	7.1s
heun	7.7s