MBZUAI’s IFM releases open 29B PAN 2 world model
Apache 2.0 weights and code ship; IFM reports 34% full-plan completion, ahead of MiniMax-H3 on multi-step tasks.
1
Stories per day · 30 days
14 RSS
Relatedimage-generation7video-generation7autoregressive5imagenet4text-to-image4training4benchmark3embodied-ai3medical-imaging3open-weights3reinforcement-learning3world-models3computer-vision2ct2distillation2flow-matching2flux2huggingface2
121 stories, newest first · one card per story
Stories tagged #diffusion
Patch-wise PCA links v-prediction failure to a bandwidth bottleneck
A Raw-injected Transformer restores effective v-prediction and is competitive on ImageNet 256×256.
WaveGen generates speech waveforms end to end without external codecs
Scaled from 0.1B to 0.6B parameters, it reports a 1.47 word error rate and 0.75 similarity on Seed-en.
Compressed 31B video diffusion model to 14B parameters
Hyperflow-H3-Flash reduces a 31B parameter video+audio diffusion model to 14B using 6 steps and INT8 quantization.
D3D enables asynchronous speculative decoding
D3D drafts during verification without target hidden states for 5.71x speedup.
LiveFoley streams video-to-audio with look-ahead below perception
A trailing window and wider local diffusion keep alignment with offline models while cutting frontend latency.
VA-TDM boosts few-step diffusion models
VA-TDM uses velocity adversarial distillation to reach state-of-the-art four-step text-to-image and text-to-video generation.
MiniMax releases open-weight MiniMax-H3-Image for generation and editing
The checkpoint outputs about 16-megapixel images in 6 steps and accepts up to 16 reference images.
Paper ties diffusion training speed to residual-gate layout
A teacher-free radial objective on ImageNet 256 with SiT-B/2 improves FID over dispersive loss while widening gate participation.
Training-free multi-bit image watermark resists removal and forgery
Authors report a scalable 32–128-bit payload, a 94.6% decode rate against CtrlRegen+, and 98.8% forgery detection.
sDMD adds adaptive Sinkhorn correction to diffusion distillation
With 2–4 steps on FLUX.1-dev, Qwen-Image-20B, and Wan2.1-14B, authors report gains over DMD2 and other baselines.
HairDeck generates animatable stylized hair cards from one image
On CHARM's test set it lowers Chamfer distance by 44% and card-orientation error by more than half, while running 33 times faster.
PhysInsert inserts objects with geometry and scene interaction
The authors built Insert World and PhysInsert-Bench and report better fidelity plus a user study on placement and lighting.
Tutorial covers generative models for music audio
It compares codecs, token models, diffusion, and flow matching, and discusses metrics, listening tests, and legal disputes.
Paper splits visual generation into autoregressive knowledge and diffusion rendering
A 3B text-to-image model trained in 6,600 H100 hours scored 86 on GenEval and 87 on DPG-Bench.
Hy-Nomas generates multi-turn interactive video worlds from text and camera
The 8B diffusion model uses action-aware causal attention and a prompt enhancer; authors release weights, code, and benchmarks.
Wiener-informed decoding speeds pixel-space flow model training
On ImageNet 256, PixelDiT-XL with WiD reports FID 1.53 after 500 epochs, matching prior quality with 37.5% fewer epochs.
AMBIT generates and edits first-order ambisonics with executable plans
A shared Transfusion model plans scenes, renders FOA audio, and self-distills on student prefixes.
SceneBloom generates collision-free indoor layouts for embodied agent training
Navigation policies trained on its scenes beat policies trained on the same number of real scans in held-out scans.
HAD-VLM pairs diffusion with language for ultrasound measurements
It cuts four EchoNet-LVH measurement errors 21.4–37.6% versus a Vicuna baseline and lowers X-ray vertebral error from 11.4 to 6.7 pixels.
Learned Unmasking Policies Improve Diffusion Language Model Efficiency
A 856K-parameter policy trained on trajectory-contraction targets achieves best accuracy-NFE frontier on LLaDA-8B-Instruct.
CausalAlign aims to keep camera ego-motion stable in video models
Fine-tuned Wan2.2 gains more than 20% relative geometric consistency and lifts RoboCasa365 world-action success by 10 points.
Future Chunk Prediction trains autoregressive video models on later chunks
The auxiliary objective predicts separated future chunks without intervening frames and does not change autoregressive inference.
UniWALM unifies video, action and language for robot control
Authors report 34.5% on RMBench memory tasks without a pretrained language model, and Find-the-Item success rising from 11.1% to 61.1%.