diff --git a/swift/Sources/CoreAIDiffusionPipeline/Pipelines/Flux2Pipeline.swift b/swift/Sources/CoreAIDiffusionPipeline/Pipelines/Flux2Pipeline.swift
index d292bbd8..246bf6f6 100644
--- a/swift/Sources/CoreAIDiffusionPipeline/Pipelines/Flux2Pipeline.swift
+++ b/swift/Sources/CoreAIDiffusionPipeline/Pipelines/Flux2Pipeline.swift
@@ -420,14 +420,32 @@ public struct Flux2Pipeline: DiffusionPipeline {
                 }
             }
 
+            // Capture the denoising state BEFORE the scheduler advances so the
+            // preview below can form the x0 estimate.
+            let previewSigma = scheduler.currentSigma
+            let sampleBeforeStep = packedLatents
             packedLatents = scheduler.step(output: output, timeStep: t, sample: packedLatents)
             try checkLatentsAreFinite(packedLatents, step: step)
 
             if let progressHandler {
+                // Preview the DENOISED estimate, not the raw post-step sample. The
+                // sample after the Euler step is still mostly noise until the last
+                // step or two, so on a few-step model (e.g. FLUX.2 Klein at 4 steps)
+                // the early previews look like static. Flow-matching gives the
+                // estimate for one multiply-add: with x_t = (1-σ)·x0 + σ·ε and the
+                // model predicting v = ε - x0, x0 = x_t - σ·v. Blurry on step one,
+                // but it shows the composition and converges to the final image.
+                var previewPacked = sampleBeforeStep
+                if previewSigma > 0 {
+                    var negSigma = -previewSigma
+                    vDSP_vsma(
+                        output, 1, &negSigma, sampleBeforeStep, 1, &previewPacked, 1,
+                        vDSP_Length(output.count))
+                }
                 // Unpack → denorm → unpatchify: [1, 128, 64, 64] → [1, 32, 128, 128]
                 // These are array copies, no model call.
                 let spatial = unpackLatentsSpatialFlatten(
-                    packedLatents, channels: inChannels, height: spatialSide, width: spatialSide)
+                    previewPacked, channels: inChannels, height: spatialSide, width: spatialSide)
                 let denormed = applyBatchNormDenorm(
                     spatial, channels: inChannels, height: spatialSide, width: spatialSide)
                 let unpatchified = Self.unpatchifyLatents(
diff --git a/swift/Sources/CoreAIDiffusionPipeline/Schedulers/DiscreteFlowScheduler.swift b/swift/Sources/CoreAIDiffusionPipeline/Schedulers/DiscreteFlowScheduler.swift
index feb96cb4..57b7257c 100644
--- a/swift/Sources/CoreAIDiffusionPipeline/Schedulers/DiscreteFlowScheduler.swift
+++ b/swift/Sources/CoreAIDiffusionPipeline/Schedulers/DiscreteFlowScheduler.swift
@@ -15,6 +15,11 @@ public final class DiscreteFlowScheduler {
     /// The first scheduled sigma after all shifts are applied — use this for img2img noise addition.
     public var startSigma: Float { sigmas.first ?? 1.0 }
 
+    /// Sigma for the step the NEXT `step(...)` call will consume — i.e. the current
+    /// point on the noise schedule, before advancing. Used to form the denoised x0
+    /// estimate for live previews: `x0 = sample − σ·v`.
+    public var currentSigma: Float { counter < sigmas.count ? sigmas[counter] : 0 }
+
     let trainSteps: Float
     let shift: Float
     let mu: Float?