KI-Video / Comparison

Testing distance facial distortion in BF16 vs INT8

Urheber@EndFolding79421 ↗

Kreative Referenz

Urheber: @EndFolding79421 ↗

01 / Deine eigene Version

Deine eigene Version

01
Betrachte das Ergebnis und sein Tempo
02
Lies die Originalnotizen des Urhebers
03
Entwickle anhand der Referenz deine eigene Idee

02 / Quelle und Urheber

Quelle und Urheber

Urheber@EndFolding79421

Originalbeitrag ↗

Notizen des UrhebersKostenlos

Modell: Hailuo 03 ↗
Kategorie: Comparison ↗

Notizen des Urhebers

MiniMax H3 — Testing the “Wall” | BF16 vs INT8 I'm testing two things here: the difference between a full BF16 model and a smaller INT8 model, and what I call “the wall” — the point where a character's face becomes smudged, distorted, or loses detail. For this test, the woman starts at the back of the scene "the wall" and walks toward the camera, gradually getting closer to the viewer. This was intentional. I wanted to see how each model handles facial detail at different distances, especially as the character approaches the camera. It always starts out distorted when she's farther away. What's interesting is seeing how quickly each model corrects that distortion as she gets closer. Models tested: (no audio) FeiHou Remix BF16 — 8 and 20 steps Link: https://t.co/2GMsNLyBDu Singularity INT8 (Non-Pruned) — 8 and 20 steps Link: https://t.co/h4LXOrNF8U Settings: Resolution: 1344 × 768p Single-pass generation No upscaling Fixed seed across all four tests Same scene, prompt, and references No LoRAs except Turbo 8-Step for the 8-step tests Hardware: RTX 5090 Laptop GPU — 24GB VRAM 64GB RAM Of course, these are two different model variants, so this isn't a pure precision comparison, but it's a useful real-world test of their output quality. Prompt used: DURATION: 6 SECONDS ASPECT RATIO: 16:9 <Picture 1> is the character reference. <Picture 2> is the gas station location reference. One woman walks toward the camera across the wet pavement of a quiet rural gas station at night, carrying a red gas can in her right hand. The camera tracks backward smoothly, maintaining a medium shot that gradually tightens toward her face. Preserve the woman's exact facial identity, hair, body proportions, and clothing. Keep the gas station layout consistent with the location reference. Natural walking movement, realistic hands, subtle facial expressions, and believable clothing movement. Cinematic live-action 35mm film appearance. Practical gas station lighting, natural shadows, muted colors, subtle film grain. Exactly ONE woman. No additional people or vehicles. No dialogue. No music. No whip-pan.

Originalbeitrag ↗