SpecEyes: Accelerating Agentic Multimodal LLMs via Speculative Perception and Planning
Paper • 2603.23483 • Published • 61
ZeroGPU Explorers
community
AI & ML interests
None defined yet.
Recent Activity
Jinfa
authored a
paper 18 days ago
Jinfa
submitted a
paper to Daily Papers 19 days ago
Jinfa
authored a
paper 26 days ago
Jinfa
submitted a
paper to Daily Papers 26 days ago
BestWishYsh
authored a
paper about 1 month ago
BestWishYsh
posted an update about 1 month ago
Post
3520
🚀 Introducing Helios: a 14B real-time long-video generation model!
It’s completely wild—faster than 1.3B models and achieves this without using self-forcing. Welcome to the new era of video generation! 😎👇
💻 Code: https://github.com/PKU-YuanGroup/Helios
🏠 Page: https://pku-yuangroup.github.io/Helios-Page
📄 Paper: Helios: Real Real-Time Long Video Generation Model (2603.04379)
🔹 True Single-GPU Extreme Speed ⚡️
No need to rely on traditional workarounds like KV-cache, quantization, sparse/linear attention, or TinyVAE. Helios hits an end-to-end 19.5 FPS on a single H100!
Training is also highly accessible: an 80GB VRAM can fit four 14B models.
🔹 Solving Long-Video "Drift" from the Core 🎥
Tired of visual drift and repetitive loops? We ditched traditional hacks (like error banks, self-forcing, or keyframe sampling).
Instead, our innovative training strategy simulates & eliminates drift directly, keeping minute-long videos incredibly coherent with stunning quality. ✨
🔹 3 Model Variants for Full Coverage 🛠️
With a unified architecture natively supporting T2V, I2V, and V2V, we are open-sourcing 3 flavors:
1️⃣ Base: Single-stage denoising for extreme high-fidelity.
2️⃣ Mid: Pyramid denoising + CFG-Zero for the perfect balance of quality & throughput.
3️⃣ Distilled: Adversarial Distillation (DMD) for ultra-fast, few-step generation.
🔹 Day-0 Ecosystem Ready 🌍
We wanted deployment to be a breeze from the second we launched. Helios drops with comprehensive Day-0 hardware and framework support:
✅ Huawei Ascend-NPU
✅ HuggingFace Diffusers
✅ vLLM-Omni
✅ SGLang-Diffusion
Try it out and let us know what you think!
It’s completely wild—faster than 1.3B models and achieves this without using self-forcing. Welcome to the new era of video generation! 😎👇
💻 Code: https://github.com/PKU-YuanGroup/Helios
🏠 Page: https://pku-yuangroup.github.io/Helios-Page
📄 Paper: Helios: Real Real-Time Long Video Generation Model (2603.04379)
🔹 True Single-GPU Extreme Speed ⚡️
No need to rely on traditional workarounds like KV-cache, quantization, sparse/linear attention, or TinyVAE. Helios hits an end-to-end 19.5 FPS on a single H100!
Training is also highly accessible: an 80GB VRAM can fit four 14B models.
🔹 Solving Long-Video "Drift" from the Core 🎥
Tired of visual drift and repetitive loops? We ditched traditional hacks (like error banks, self-forcing, or keyframe sampling).
Instead, our innovative training strategy simulates & eliminates drift directly, keeping minute-long videos incredibly coherent with stunning quality. ✨
🔹 3 Model Variants for Full Coverage 🛠️
With a unified architecture natively supporting T2V, I2V, and V2V, we are open-sourcing 3 flavors:
1️⃣ Base: Single-stage denoising for extreme high-fidelity.
2️⃣ Mid: Pyramid denoising + CFG-Zero for the perfect balance of quality & throughput.
3️⃣ Distilled: Adversarial Distillation (DMD) for ultra-fast, few-step generation.
🔹 Day-0 Ecosystem Ready 🌍
We wanted deployment to be a breeze from the second we launched. Helios drops with comprehensive Day-0 hardware and framework support:
✅ Huawei Ascend-NPU
✅ HuggingFace Diffusers
✅ vLLM-Omni
✅ SGLang-Diffusion
Try it out and let us know what you think!
- 6 replies
Zhengyi
authored a
paper about 2 months ago
Taishi-N324
authored a
paper about 2 months ago
wangfuyun
authored a
paper 2 months ago
Post
1895
PromptRL: Language Models as Co-Learners in Flow-Based Image Generation RL 🚀
We found two critical failure modes in flow-based RL:
1️⃣ Quality-Diversity Dilemma: High-quality models produce similar outputs, bottlenecking RL exploration
2️⃣ Prompt Linguistic Hacking: Models overfit to surface patterns—paraphrase the prompt and performance tanks
Solution: **Jointly train LM + FM** — the LM dynamically generates semantically-consistent but diverse prompt variants
📊 Results:
• GenEval: 0.97
• OCR accuracy: 0.98
• PickScore: 24.05
• 2×+ fewer rollouts than flow-only RL
Paper: arxiv.org/abs/2602.01382
Code: github.com/G-U-N/UniRL
#AI #TextToImage #ReinforcementLearning #Diffusion
We found two critical failure modes in flow-based RL:
1️⃣ Quality-Diversity Dilemma: High-quality models produce similar outputs, bottlenecking RL exploration
2️⃣ Prompt Linguistic Hacking: Models overfit to surface patterns—paraphrase the prompt and performance tanks
Solution: **Jointly train LM + FM** — the LM dynamically generates semantically-consistent but diverse prompt variants
📊 Results:
• GenEval: 0.97
• OCR accuracy: 0.98
• PickScore: 24.05
• 2×+ fewer rollouts than flow-only RL
Paper: arxiv.org/abs/2602.01382
Code: github.com/G-U-N/UniRL
#AI #TextToImage #ReinforcementLearning #Diffusion
wangfuyun
submitted a
paper to Daily Papers 2 months ago
BestWishYsh
authored a
paper 3 months ago
wangfuyun
authored 2
papers 4 months ago
wangfuyun
submitted a
paper to Daily Papers 4 months ago
Paper99
authored 2
papers 4 months ago
dawood
authored a
paper 5 months ago
yuanshengni
authored a
paper 6 months ago