NVIDIA ACABA DE SACAR UNA HERRAMIENTA PARA CONVERTIR CUALQUIER IMAGEN EN UN MUNDO 3D EXPLORABLE
y la han hecho 100% de código abierto.
→ modelo disponible en hugging face
→ código de la interfaz disponible en GitHub
Lyra 2.0 crea un mundo por el que puedes caminar, girar y hasta soltar un robot para simulación.
os dejo el repo abajo.
FastH3 secretly supports Reference-to-Video?! 🎬🤫
Official docs claimed FastH3 was T2V/I2V only with zero reference capabilities. But we plugged it into ComfyUI's Ref2V pipeline and it's INSANE:
✅ Fast 8-step sampling with ZERO extra LoRAs
✅ Dual-actor interactions (Passing tea without face-bleeding)
✅ 5-Image Asset Assembly (Actor A + Dress + Bag + Drink + Actor B)
✅ 2-Stage Latent Upscale for speed & crisp textures
Full benchmark & RunningHub workflows below: 👇
🔗 https://t.co/GCmDuTi9ND
#FastH3 #MiniMaxH3 #ComfyUI #Ref2Video #AIVideo #GenerativeAI
BREAKING NEWS🚨: NVIDIA has just achieved 423 tok/s/user 🚀 on the 2.8T Parameter Kimi K3 model via a new upstream vLLM optimization. 1.68x Faster on the Same Hardware!
This was done by NVIDIA CentML's Benjamin, who wrote NVIDIA's new decode-specialized kernels in Cute DSL via aggressive Megakernel fusion. This is a 12,000 LoC Megakernel and is the future of advanced AI-assisted kernel generation. NVIDIA is already working on DeepSeekv4.1 and other models for vLLM too.
Someone should give this NVIDIA 10x engineer a raise for this great new major feature!
Grok用视频提示词:
Master Prompt:
A cinematic wuxia fantasy battle: a woman warrior in black-and-white attire wielding a glowing teal sword duels a silver-haired swordsman with a coin-forged blade in a rain-soaked ancient Chinese courtyard at night. Flames clash with teal-green spectral energy, copper coins transform into a golden serpent, fire pillars erupt skyward. Deep black cinematic grade with warm orange firelight against cold teal energy glow, wet stone reflections, lantern light in the background, dramatic bloom on flames and energy. Epic dark fantasy atmosphere, photorealistic cinematic 3D render, fluid martial arts choreography. AUDIO: epic battle music with heavy percussion drums, sword clashes, fire whooshes, energy crackles, no dialogue. Avoid: warped hands, deformed faces, extra limbs, flickering flames.
Shot 1 (0-2s):
A woman warrior in black-and-white wuxia attire stands sword-ready amid a sea of flames, spectral teal fireballs orbiting her, then leaps into the air. Slow push-in. Flames as key light, embers flying, wet ground reflections. AUDIO: epic battle music begins, dense drums. Avoid: deformed face, weak flames.
Shot 2 (2-3s):
She lands and walks through the night courtyard, lanterns glowing in the background. Follow shot. Warm lantern light, wet reflections, dark tones. AUDIO: music continues. Avoid: floating walk.
Shot 3 (3-5.5s):
Close-up of the silver-haired swordsman caressing his long sword forged from strung copper coins, metal patterns gleaming. Slight push-in. Side light on coin metal, shallow depth of field, warm white highlights. AUDIO: music turns low. Avoid: warped fingers, blurry coins.
Shot 4 (5.5-7.5s):
The woman holds her sword glowing teal, facing off against the swordsman. Slow push-in. Teal glow lighting her face, cold midtones. AUDIO: tension building. Avoid: drifting faces.
Shot 5 (7.5-9.5s):
Fast duel with clashing blades, fire colliding with teal energy, motion blur on swings. Dynamic follow shots. Alternating warm and cold light sources, energy trails with bloom. AUDIO: drums at peak, sword clashes. Avoid: extra limbs, warped hands.
Shot 6 (9.5-12.5s):
Copper coins detach and transform into a golden serpent coiling through the air. Orbiting shot. Warm metallic reflections, motion blur. AUDIO: music continues. Avoid: deformed serpent.
Shot 7 (12.5-14.5s):
Wood chips burst around the swordsman as a pillar of flame erupts skyward. Tilt up. Flame pillar as strong light source, ember particles. AUDIO: music climax. Avoid: flickering flames.
Shot 8 (14.5-17.5s):
The woman leaps again amid fire and teal flames, spectral fireballs reappearing around her. Follow shot. Dual-color light, teal-orange glow. AUDIO: music continues. Avoid: morphing artifacts.
Shot 9 (17.5-20.5s):
A teal energy vortex erupts, camera rising to an overhead view of the courtyard. Descending crane to top-down. Self-lit green vortex, mist, cold highlights. AUDIO: music at its peak. Avoid: blurry vortex.
Shot 10 (20.5-23s):
The woman stands among scattered copper coins, holding the glowing teal sword, robes flowing. Slow push-in. Sword glow lighting her figure, coin reflections. AUDIO: music turns solemn. Avoid: drifting face.
Shot 11 (23-27.4s):
She turns with the sword, standing in a final heroic pose. Slow push-in. Rim light on her silhouette, vignette. AUDIO: music resolves, no dialogue. Avoid: deformed sword.
ok wow… video rebirth open-sourced hyperflow 🤯
data-free flow self-distillation, built on minimax h3. 49 sampling steps down to 8, ~3x faster end to end. video + synced audio.
tutorial + prompts below (sound on 🔊):
🎬 MiniMax-H3, now powered by Sol Engine in ComfyUI!
Generate 5s videos with native audio on a single GPU:
⚡ RTX 5090: 35.5s
⚡ DGX Spark: 64.3s
Meet Sol-H3 Studio: a visual workspace for local video generation, powered by our optimized inference engine.
Open source 👇
https://t.co/OYZxIPgxwU
Four free models. One API.
Try MiMo V2.6, DeepSeek V4.1 Flash, DeepSeek V4 Flash, and Qwen3.8 Flash on Token Harbor.
No credit card needed to get started.
Now we can use Fish Drama 3 in ComfyUI, and yes, it actually runs on just 8GB of VRAM!
massive 35B text-to-speech model does insane voice cloning, full dialogue, and even inline emotional cues like [whispers] /[sighs].
> FP8/int8 ConvRot/W6A8/W4A8
smart MoE routing,. it keeps the heavy lifting in system RAM
https://t.co/qCb3nGUb5V