Stability-AI/generative-models · error · ValueError

Sampler and loss function need to be set for training.

Error message

Sampler and loss function need to be set for training.

What it means

DiffusionEngine.on_train_start asserts, at the moment training actually begins, that both self.sampler and self.loss_fn were configured. If either is None it raises this ValueError — meaning you instantiated a DiffusionEngine without a sampler or loss and tried to call .fit() on it.

Source

Thrown at sgm/models/diffusion.py:191

            "global_step",
            self.global_step,
            prog_bar=True,
            logger=True,
            on_step=True,
            on_epoch=False,
        )

        if self.scheduler_config is not None:
            lr = self.optimizers().param_groups[0]["lr"]
            self.log(
                "lr_abs", lr, prog_bar=True, logger=True, on_step=True, on_epoch=False
            )

        return loss

    def on_train_start(self, *args, **kwargs):
        if self.sampler is None or self.loss_fn is None:
            raise ValueError("Sampler and loss function need to be set for training.")

    def on_train_batch_end(self, *args, **kwargs):
        if self.use_ema:
            self.model_ema(self.model)

    @contextmanager
    def ema_scope(self, context=None):
        if self.use_ema:
            self.model_ema.store(self.model.parameters())
            self.model_ema.copy_to(self.model)
            if context is not None:
                print(f"{context}: Switched to EMA weights")
        try:
            yield None
        finally:
            if self.use_ema:
                self.model_ema.restore(self.model.parameters())
                if context is not None:

View on GitHub (pinned to e8cd657656)

Solutions

1. Add a `sampler` section to the model config (e.g. a DDIMSampler target with its discretization/guider configs).
2. Add a `loss_fn` section to the model config (e.g. StandardDiffusionLoss target with its noise schedule).
3. If you only intended inference, don't call trainer.fit on this model; run text_to_image-style sampling instead.

Example fix

// before
model:
  target: sgm.models.diffusion.DiffusionEngine
  params:
    network_config: ...
    # no sampler / loss_fn
// after
model:
  target: sgm.models.diffusion.DiffusionEngine
  params:
    network_config: ...
    sampler:
      target: sgm.samplers.EulerEDMSampler  # or any configured sampler
    loss_fn:
      target: sgm.modules.diffusionmodules.loss.StandardDiffusionLoss

params = config.model.params
assert params.get('sampler') is not None, "config.model.params.sampler missing"
assert params.get('loss_fn') is not None, "config.model.params.loss_fn missing"
# before calling trainer.fit(config.model, ...)

try:
    trainer.fit(model, data=dm)
except ValueError as e:
    if "Sampler and loss function need to be set" in str(e):
        raise RuntimeError("Add sampler and loss_fn sections to the model config before training") from e

When it happens

Trigger: Instantiating DiffusionEngine from a config where the `sampler:` or `loss_fn:` key is missing/null, then calling trainer.fit(model). Inference-only usage without these keys works; training does not.

Common situations: Reusing an inference config for training; a YAML config edit deleting or commenting out loss_fn/sampler; target class instantiated with only network params (ckpt, conditioner) for finetuning without defining the loss.

Understand the failure class

Background: "X is required", "must be set", "cannot be empty": the missing-required-config error family, from Vertex AI project/location to WeChat keys — this error's family across 18 libraries.

Related errors

• Model {model_id} not supported
• unknown discretization {params.discretization}
• unknown sampler {params.sampler}!
• Initializing ActNorm in reverse direction is disabled by def
• unsupported dimensions: {dims}

AI-assisted analysis of Stability-AI/generative-models@e8cd657656 (2026-08-29). Data as JSON: /api/errors/c2261eb31b5ea627. Report an issue: GitHub.