StepX-OPD Rethinking Multi-Teacher On-Policy Distillation in Diffusion Models

Capability integration and trajectory compression, trained together on the student’s own rollout.

Chenru Wang^1,*, Xuanyang Zhang¹, Yucheng Han¹, Wenbo Nie^1,*, Wei Cheng¹, Xianfang Zeng¹, Gang YU¹, Chi Zhang²

¹StepFun·²Westlake University

^*Work done during an internship at StepFun.

Qualitative comparisons

The comparisons shown in the paper

These are the twenty prompts the paper shows. All four models generate in four sampling steps. Each cell uses one fixed noise seed, chosen per cell rather than shared across a row. The prompts are held out from training.

Generation for the prompt: A white cat sleeping on an orange armchair beside a stack of green books, soft afternoon l — Attribute binding“A white cat sleeping on an orange armchair beside a stack of green books, soft afternoon light through lace curtains.”

Generation for the prompt: Four candles of unequal height burning on an old church altar, wax pooling, dark stone beh — Counting“Four candles of unequal height burning on an old church altar, wax pooling, dark stone behind.”

Generation for the prompt: A wooden crate of lemons stamped "CITRUS CO", sunlit market stall, striped awning shadow f — Text rendering“A wooden crate of lemons stamped "CITRUS CO", sunlit market stall, striped awning shadow falling across the fruit.”

Generation for the prompt: Iridescent origami crane on a dark surface, with crisp paper folds highlighted by soft mac — Held-out narrative prompt“Iridescent origami crane on a dark surface, with crisp paper folds highlighted by soft macro lighting.”

Generation for the prompt: A black apple and a green backpack on a sunlit wooden bench, shallow depth of field, warm — Attribute binding“A black apple and a green backpack on a sunlit wooden bench, shallow depth of field, warm afternoon light.”

Generation for the prompt: A brass diving helmet and a coil of blue rope on a weathered dock, sea spray, low sun behi — Attribute binding“A brass diving helmet and a coil of blue rope on a weathered dock, sea spray, low sun behind.”

Generation for the prompt: Three wooden rowing boats moored on a still alpine lake at dawn, mountains mirrored in the — Counting“Three wooden rowing boats moored on a still alpine lake at dawn, mountains mirrored in the water.”

Generation for the prompt: Three ripe pomegranates split open on a dark table, seeds glistening, dramatic side light, — Counting“Three ripe pomegranates split open on a dark table, seeds glistening, dramatic side light, baroque still life.”

Generation for the prompt: Four seagulls in flight against a stormy grey sky above cliffs, wind-torn spray below. — Counting“Four seagulls in flight against a stormy grey sky above cliffs, wind-torn spray below.”

Generation for the prompt: Five jellyfish drifting in dark water, translucent bells glowing blue, fine tentacles trai — Counting“Five jellyfish drifting in dark water, translucent bells glowing blue, fine tentacles trailing.”

Generation for the prompt: Three drones hovering above a terraced rice field at sunrise, mist between the terraces. — Counting“Three drones hovering above a terraced rice field at sunrise, mist between the terraces.”

Generation for the prompt: Four crows lined up on a snowy fence rail, black against white, flat overcast light. — Counting“Four crows lined up on a snowy fence rail, black against white, flat overcast light.”

Generation for the prompt: Stacked wooden animal totem with bear at the bottom, eagle in the middle and fish on top, — Held-out narrative prompt“Stacked wooden animal totem with bear at the bottom, eagle in the middle and fish on top, with weathered paint, forest bokeh and warm light.”

Generation for the prompt: A hummingbird hovering in front of a red trumpet flower with blurred green foliage behind, — Spatial relation“A hummingbird hovering in front of a red trumpet flower with blurred green foliage behind, high-speed macro.”

Generation for the prompt: An old typewriter with a sheet of paper rolled in, the typed line reading "CHAPTER ONE", d — Text rendering“An old typewriter with a sheet of paper rolled in, the typed line reading "CHAPTER ONE", dust motes in a slanting sunbeam.”

Generation for the prompt: A hand-stitched sampler framed on a wall reading "HOME SWEET HOME", late afternoon light, — Text rendering“A hand-stitched sampler framed on a wall reading "HOME SWEET HOME", late afternoon light, faded floral wallpaper behind.”

Generation for the prompt: A construction site hoarding painted with "COMING SOON", scaffolding above, crane silhouet — Text rendering“A construction site hoarding painted with "COMING SOON", scaffolding above, crane silhouette against an orange evening sky.”

Generation for the prompt: An enamel mug on a campfire stone printed with "WILDERNESS", flames glowing, pine forest d — Text rendering“An enamel mug on a campfire stone printed with "WILDERNESS", flames glowing, pine forest dark behind, blue hour.”

Generation for the prompt: A vast, sun-baked desert landscape with a lone highway stretching into the distance. A wea — Text rendering“A vast, sun-baked desert landscape with a lone highway stretching into the distance. A weathered road sign stands by the side, warning "Next Gas 100 Miles", under a blazing sky.”

Generation for the prompt: A green sign that says "Very Deep Learning" and is at the edge of the Grand Canyon. — Text rendering“A green sign that says "Very Deep Learning" and is at the edge of the Grand Canyon.”