SoL-Refiner-H3 — NVlabs reference (bf16) vs image-server (int8)

The same FlashGen 4-step 768p clip refined to 1920×1080 twice. Left: NVlabs' own pipeline as released with the full bf16 checkpoint and exact attention. Right: our image-server (sol-refiner-h3: int8 transformer, int8 text encoder, SageAttention). Same input, same refine prompt, same seed number — but the two code paths draw their noise differently, so these are two takes of the same refinement, not one take at two precisions.

90s vs 41s refine per clip (NVlabs offloaded / ours)
1920×1080 · 121f both
6 clips 3 i2va + 3 t2va from the FlashGen page
σ 0.909, 1 step both
What to look for. Detail and sharpness, faces and hands, text and fine texture, temporal stability. Differences in content (a different face, a changed prop) are expected between two takes at σ 0.909 and are not by themselves a quality gap; a consistent difference in quality across clips would be.
Deviations forced on the NVlabs run (32 GB card): weights streamed from host RAM (sequential CPU offload), decoder tiles 512/384 instead of their 768/512 (theirs ran out of memory), torch 2.8 / NATTEN 0.21.1 instead of 2.9.1 / 0.21.5. Their pipeline outputs video only, so the left side has no audio.

Per-clip wall clock

Refine step only, same RTX 5090. NVlabs' pipeline streams its 38 GB bf16 transformer from host RAM (sequential CPU offload) because it does not fit 32 GB, so its time reflects offload, not the model.

clip NVlabs bf16 (s) image-server int8 (s) image-server faster by
Chef91.941.32.23×
Helicopter89.741.62.16×
Skateboarder89.741.52.16×
T2VA 0189.741.52.16×
T2VA 0889.841.42.17×
T2VA 1790.941.62.19×
mean (excl. warm-up) 90.341.5 2.18×

Image → video+audio (12)

lightx2v's i2va examples, images unmodified, prompts verbatim. The thumbnail is the first frame both arms were given.

Chef — animated from a still

1920×1080 · 1920×1080 · NVlabs 91.9s vs image-server 41.3s
◀ NVlabs reference · bf16 · no audio image-server · int8 + Sage ▶
mean/peak: NVlabs: no audio track · image-server -29.3/-0.0 dB · clipped 0/27
audio: 0.0s
prompt
Live-action, cinematic, a warm daylight medium close-up preserves the adult male chef's short brown hair, trimmed beard, focused expression, white double-breasted chef's coat with black piping and buttons, black trousers, and his position at the wooden cutting board established in the first frame. The crowded kitchen remains unchanged, with copper pans hanging behind him, jars and ingredients on the shelves, vivid vegetables and herbs filling the right foreground, and a metal pot steaming on the stove. Beginning with the knife resting against the cucumber, he curls the fingertips of his guiding hand inward and drives the broad blade through it in a quick sequence of clean, even downward cuts. The camera pushes in with small amplitude at slow speed toward his hands as thin cucumber rounds accumulate beside the chopped carrots and other vegetables. He finishes the final slice, turns the blade sideways, and uses its flat face to gather the fresh pieces into a neat pile while steam continues to billow behind him and sunlight glints along the moving knife.

Helicopter — animated from a still

1920×1080 · 1920×1080 · NVlabs 89.7s vs image-server 41.6s
◀ NVlabs reference · bf16 · no audio image-server · int8 + Sage ▶
mean/peak: NVlabs: no audio track · image-server -18.2/-2.8 dB
audio: 0.0s
prompt
Live-action, cinematic, a low-angle wide shot explicitly begins from the first frame: the same dark navy-black helicopter, with its rounded glazed cockpit, roof-mounted engine housing, long right-pointing tail boom, enclosed tail rotor, landing skids, and seated pilot silhouette, holds a very low hover above the paved landing area. Its scale, right-facing profile, central composition, warm backlit edges, and deep shadow remain consistent. The large sunrise glows behind the helicopter on the left, silhouetting the distant tree line beneath a pale blue-and-gold sky; small birds remain high in the distance, dark grass fills the foreground, and suspended golden dust drifts across the runway. The main and tail rotors accelerate into broader motion blur as their synchronized whir intensifies. Downwash flattens the nearby grass and drives dust and loose debris outward in expanding circular waves, glowing as they cross the sunbeams. The helicopter then rises smoothly from its low hover, its skids gaining clear separation from the ground while its moving shadow softens beneath it. It pitches forward only slightly and begins a controlled rightward acceleration while continuing to climb. The camera holds the restrained low viewpoint and tilts up with small amplitude at slow speed to keep the aircraft centered, allowing the sunlit dust cloud to spread below as the helicopter finishes several meters above the landing area.

Skateboarder — animated from a still

1920×1080 · 1920×1080 · NVlabs 89.7s vs image-server 41.5s
◀ NVlabs reference · bf16 · no audio image-server · int8 + Sage ▶
mean/peak: NVlabs: no audio track · image-server -13.1/-0.2 dB · clipped 0/32
audio: 0.0s
prompt
Live-action, cinematic, a warm daytime medium-wide shot begins from the exact side-on composition in the first frame: the long-haired skateboarder in a charcoal-gray T-shirt, black pants, black sneakers, and a red wristband rides from the left foreground along the broad concrete sidewalk beside the busy city intersection. Golden side light silhouettes the skateboarder's profile and streams through the roadside trees while cars and distant pedestrians continue moving through the background. A low tracking shot keeps pace beside the skateboarder as they compress into the existing crouch, snap the tail against the pavement, and lift into a clean ollie over a transverse sidewalk seam. Their hair and shirt move in the slipstream as the board levels beneath both feet; they land smoothly with knees absorbing the impact, carve gently toward the open center of the sidewalk, then build speed and execute a controlled frontside 180, turning shoulders, hips, and board together. The wheels reconnect with the pavement in alignment, and the skateboarder rides away toward the intersection while the camera continues tracking at the same warm, low angle.

Text → video+audio (24)

lightx2v's t2va examples, prompts verbatim.

T2VA 01 — Live-action, cinematic, a low-angle medium-wide tracking shot follows…

1920×1080 · 1920×1080 · NVlabs 89.7s vs image-server 41.5s
◀ NVlabs reference · bf16 · no audio image-server · int8 + Sage ▶
mean/peak: NVlabs: no audio track · image-server -13.3/0.0 dB · clipped 0/67
audio: 0.0s
prompt
Live-action, cinematic, a low-angle medium-wide tracking shot follows two athletic martial artists sparring on a rain-slick rooftop at blue hour, city lights glowing behind them. The camera arcs right at fast speed as the fighter in a red jacket ducks beneath a spinning kick, plants one hand on the wet concrete, sweeps the other fighter's legs, and both recover into guarded stances without injury. Water sprays from each footfall and a loose rooftop cable whips in the wind.

T2VA 08 — Live-action, cinematic landscape, an ultra-wide aerial view glides abo…

1920×1080 · 1920×1080 · NVlabs 89.8s vs image-server 41.4s
◀ NVlabs reference · bf16 · no audio image-server · int8 + Sage ▶
mean/peak: NVlabs: no audio track · image-server -15.3/-1.2 dB
audio: 0.0s
prompt
Live-action, cinematic landscape, an ultra-wide aerial view glides above a mirror-still alpine lake before sunrise, with dark fir trees in the foreground and snow-covered peaks filling the horizon. The camera pedestals up with large amplitude at slow speed as the first orange sunlight reaches the summit, travels down the ridges, and turns thin valley mist gold; a small flock of birds crosses the reflection and subtle ripples spread from the shore.

T2VA 17 — Hand-drawn 2D animation with watercolor backgrounds, a wide shot frame…

1920×1080 · 1920×1080 · NVlabs 90.9s vs image-server 41.6s
◀ NVlabs reference · bf16 · no audio image-server · int8 + Sage ▶
mean/peak: NVlabs: no audio track · image-server -29.0/-14.4 dB
audio: 0.0s
prompt
Hand-drawn 2D animation with watercolor backgrounds, a wide shot frames a young mage standing on a cliff above a floating island city at sunset as a long blue dragon dives from the clouds. The camera tilts up with large amplitude at fast speed while the mage spins a staff, draws a glowing circular shield, and redirects the dragon's harmless stream of golden fire into a ribbon that spirals around both characters. The dragon folds its wings and lands beside the smiling mage.