MiniMax-H3 768p → 1440p: Open-H3-Regenerate-2K vs FlashVSR · blind

5 prompts × 4 arms. Every clip upscales the same production 768p draft (shown first, not voted) on one RTX 5090. Arms differ in HOW they reach 1440p: a video super-resolution model, or NVlabs' Open-H3-Regenerate-2K (latent ×2 upscale, then 3 denoising steps of H3 with a refiner LoRA) with different attention and VAE choices. The regen arms re-generate detail, so small content changes vs the draft are expected — judge each clip on whether you would ship it, at full size (fullscreen, 1:1 if you can), not on resemblance to the draft. Watch faces, hands, text, fur/foliage texture and motion; play the audio (it is the draft's audio in every arm). FlashVSR outputs 2688×1536, the others 2560×1440. Order is rotated per prompt and identities are hidden; pick first, then reveal.

p1 fox t2v

prompt
integrated_multimodal_description: [Shot 1] cinematic live-action, close-up, a red fox trots through fresh powder snow in a birch forest at golden hour, individual hairs of its thick winter coat catch the low sunlight, snow crystals sparkle and puff up around its paws, the camera tracks alongside it with small amplitude at slow speed.
overall_soundscape: soft crunch of paws in snow, light wind through bare branches, distant crow calls.
non_diegetic_music: none.
Reference production 768p draft (1344×768) — the input to every clip below, not voted
Clip 1 FLASHVSR x2 · FlashVSR-full (deAPI v1.1) x2 tiled 256 -> 2688x1536, image-server over HTTP · 191s on top of the draft
Clip 2 REGEN SAGE2 · Regen-2K: latent x2 + 3-step refine, dense sage2, our H3 VAE -> 2560x1440 · 206s on top of the draft
Clip 3 REGEN ISK-SOL · Regen-2K: latent x2 + 3-step refine, isk-sol sparse, our H3 VAE -> 2560x1440 · 115s on top of the draft
Clip 4 REGEN ISK-SOL + SOL-VAE · as ISK-SOL, decoded by Sol-H3-VAE (5-latent chunks, 5 ctx) -> 2560x1440 · 127s on top of the draft

p2 market t2v

prompt
integrated_multimodal_description: [Shot 1] cinematic live-action, medium close-up, at a crowded night market in Taipei a middle-aged woman vendor in a red apron stands behind steaming bamboo baskets, glowing neon signs with Chinese characters behind her, she looks at the camera and says warmly: "Fresh dumplings, just out of the steamer, come try one!", the camera holds steady at eye level.
overall_soundscape: crowd chatter, sizzling woks, clinking bowls, a scooter passing in the distance.
non_diegetic_music: none.
Reference production 768p draft (1344×768) — the input to every clip below, not voted
Clip 1 REGEN SAGE2 · Regen-2K: latent x2 + 3-step refine, dense sage2, our H3 VAE -> 2560x1440 · 207s on top of the draft
Clip 2 REGEN ISK-SOL · Regen-2K: latent x2 + 3-step refine, isk-sol sparse, our H3 VAE -> 2560x1440 · 118s on top of the draft
Clip 3 REGEN ISK-SOL + SOL-VAE · as ISK-SOL, decoded by Sol-H3-VAE (5-latent chunks, 5 ctx) -> 2560x1440 · 130s on top of the draft
Clip 4 FLASHVSR x2 · FlashVSR-full (deAPI v1.1) x2 tiled 256 -> 2688x1536, image-server over HTTP · 192s on top of the draft

p3 waterfall t2v

prompt
integrated_multimodal_description: [Shot 1] cinematic live-action, wide shot, a narrow waterfall pours over dark basalt rocks covered in thick green moss and tiny ferns in a misty rainforest gorge, fine water spray drifts through shafts of morning light, the camera slowly cranes down with small amplitude.
overall_soundscape: rushing water, dripping leaves, distant bird calls.
non_diegetic_music: a quiet ambient pad.
Reference production 768p draft (1344×768) — the input to every clip below, not voted
Clip 1 REGEN ISK-SOL · Regen-2K: latent x2 + 3-step refine, isk-sol sparse, our H3 VAE -> 2560x1440 · 116s on top of the draft
Clip 2 REGEN ISK-SOL + SOL-VAE · as ISK-SOL, decoded by Sol-H3-VAE (5-latent chunks, 5 ctx) -> 2560x1440 · 128s on top of the draft
Clip 3 FLASHVSR x2 · FlashVSR-full (deAPI v1.1) x2 tiled 256 -> 2688x1536, image-server over HTTP · 193s on top of the draft
Clip 4 REGEN SAGE2 · Regen-2K: latent x2 + 3-step refine, dense sage2, our H3 VAE -> 2560x1440 · 207s on top of the draft

p4 dance t2v

prompt
integrated_multimodal_description: [Shot 1] cinematic live-action, full shot, a young breakdancer in a white hoodie and baggy jeans performs fast windmills and freezes on the concrete floor of an abandoned warehouse, dust kicks up and hangs in hard beams of sunlight from broken windows, the camera circles around him with medium amplitude at medium speed.
overall_soundscape: sneakers squeaking on concrete, cloth whipping, echoing claps from onlookers.
non_diegetic_music: an energetic hip-hop beat with heavy drums.
Reference production 768p draft (1344×768) — the input to every clip below, not voted
Clip 1 REGEN ISK-SOL + SOL-VAE · as ISK-SOL, decoded by Sol-H3-VAE (5-latent chunks, 5 ctx) -> 2560x1440 · 128s on top of the draft
Clip 2 FLASHVSR x2 · FlashVSR-full (deAPI v1.1) x2 tiled 256 -> 2688x1536, image-server over HTTP · 191s on top of the draft
Clip 3 REGEN SAGE2 · Regen-2K: latent x2 + 3-step refine, dense sage2, our H3 VAE -> 2560x1440 · 206s on top of the draft
Clip 4 REGEN ISK-SOL · Regen-2K: latent x2 + 3-step refine, isk-sol sparse, our H3 VAE -> 2560x1440 · 116s on top of the draft

p5 chef t2v

prompt
integrated_multimodal_description: [Shot 1] cinematic live-action, close-up, an older chef with a grey beard finely chops fresh parsley on a worn wooden board in a busy restaurant kitchen, steam rises from a pot behind him, he glances up and says: "The secret is a sharp knife and patience.", the camera pushes in slowly with small amplitude.
overall_soundscape: rhythmic knife taps on wood, bubbling pot, muffled kitchen clatter.
non_diegetic_music: none.
Reference production 768p draft (1344×768) — the input to every clip below, not voted
Clip 1 FLASHVSR x2 · FlashVSR-full (deAPI v1.1) x2 tiled 256 -> 2688x1536, image-server over HTTP · 191s on top of the draft
Clip 2 REGEN SAGE2 · Regen-2K: latent x2 + 3-step refine, dense sage2, our H3 VAE -> 2560x1440 · 207s on top of the draft
Clip 3 REGEN ISK-SOL · Regen-2K: latent x2 + 3-step refine, isk-sol sparse, our H3 VAE -> 2560x1440 · 116s on top of the draft
Clip 4 REGEN ISK-SOL + SOL-VAE · as ISK-SOL, decoded by Sol-H3-VAE (5-latent chunks, 5 ctx) -> 2560x1440 · 127s on top of the draft

Result