Sajjad Pakdamansavoji's picture
šŸ‘‹ Open to Work

Sajjad Pakdamansavoji

savoji
3

AI & ML interests

VLM, VLA, LLM, Robotics, Computer Vision

Recent Activity

posted an update about 22 hours ago
Jeveration draws a 512x512 image in 77.8 ms with a single forward pass of a frozen Qwen3.5-4B. That is up to 122x faster than Janus-Pro-7B (9,508 ms), at 0.74 GenEval vs its 0.80. Autoregressive and diffusion models need many sequential evaluations per image. Here all 128 image codes are decided in parallel, conditioned on one shared random state so they fit together. Only rank-256 LoRA adapters and output heads are trained. The same checkpoint also does image-to-image and dense prediction. šŸ” How it works • Phase 1: a masked model learns the code distribution. Alone, it reaches only 0.32 GenEval in one pass • Phase 2: Di[M]O-style distribution matching plus a correction target trains the one-pass model • Fixed state instead of random: GenEval 0.73 to 0.70, HPSv2.1 28.27 to 26.63 šŸ“Š The numbers • Show-o Turbo-512 (4 steps): 0.63 GenEval, 94.1 ms. Jeveration: 0.74, 77.8 ms • Image-to-image at 91.7 ms per request; NYUv2 segmentation 67.7 mIoU vs 53.5 for DeepLabV3+ • Joint training vs text-to-image only: GenEval 0.74 vs 0.73 āš ļø Caveats • DPG 78.24 trails FLUX.1-schnell (85.97) and Janus-Pro-7B (83.91) • Dense training is NYUv2 indoor only; normals lag (24.2 vs 21.2 degrees) • Latency is one H200 at batch 1; the repo has inference code only; not independently reproduced Weights are 6.5 GB, Apache-2.0: hf download linlinlinlinsdada/Jeveration --local-dir weights /papers/2611.24738 https://github.com/Carol-lyh/Jeveration If you run a few-step generator today, would 78 ms at 0.74 GenEval beat it for your latency budget? šŸ“š Read it elsewhere • The deep dive on Medium: https://medium.com/p/e129423fab53 • The long read on Substack: https://sajjadpakdamansavoji.substack.com/p/jeveration-a-frozen-qwen35-4b-draws • The thread on X: https://x.com/saavoji/status/2123847044502864155 • Every version on sajjadpsavoji.github.io: https://sajjadpsavoji.github.io/blog#jeveration-makes
reacted to theirpost with šŸ”„ 1 day ago
Apple's FastFrames cuts 1.3B Wan 2.1 video latency from 272 ms to 36 ms per frame on one B200 at 480p, with similar VBench scores. The denoiser was not the bottleneck. In Causal Forcing++, the fastest prior 1.3B model, diffusion sampling is only about 30% of a 272 ms frame, while the VAE decoder alone takes 147 ms. @apple keeps the DiT frozen and shrinks what surrounds it. šŸ” What changed • Pixel decoder: 3 causal ConvNeXt blocks plus a linear layer replace the Wan VAE decoder, 63.9M vs 73.3M params, 1 ms vs 147 ms • KV repair network (about 2 ms) replaces the extra DiT pass that refreshes the KV cache • Systems work: persistent KV buffers, fused QKV, CUDA graphs; one DiT call goes from 41.6 ms to 16.4 ms šŸ“Š The numbers • FastFrames-S: 36 ms (27.8 RPS), VBench Total 84.23 vs 84.11 for Causal Forcing++ • FastFrames-L (14B): 45 ms on 8x B200, 158 ms on one GPU; Total 84.44 vs 84.42 for Krea-Realtime at 1,675 ms • Judge interactivity by responses per second, not FPS: a 50 ms target is 20 RPS āš ļø Caveats • S scores 79.92 on Semantic vs 80.20 for Causal Forcing++ • The pixel decoder is fine tuned per model; tested on B200s at 480p only • License tag is apple-amlr; LingBot parts are CC BY-NC-SA 4.0 šŸ› ļø Reproduce the profile from the repo: ``` uv run python evaluation/latency/benchmark.py ff-s ``` /apple/FastFrames-L /papers/2611.23226 Before you distill more steps, have you ranked your KV update and pixel decoding by wall time? šŸ“š Read it elsewhere • The deep dive on Medium: https://medium.com/p/589b4109f380 • The long read on Substack: https://sajjadpakdamansavoji.substack.com/p/apples-fastframes-shows-the-denoiser • The thread on X: https://x.com/saavoji/status/2123477247952879830 • Every version on sajjadpsavoji.github.io: https://sajjadpsavoji.github.io/blog#apples-fastframes
posted an update 2 days ago
Apple's FastFrames cuts 1.3B Wan 2.1 video latency from 272 ms to 36 ms per frame on one B200 at 480p, with similar VBench scores. The denoiser was not the bottleneck. In Causal Forcing++, the fastest prior 1.3B model, diffusion sampling is only about 30% of a 272 ms frame, while the VAE decoder alone takes 147 ms. @apple keeps the DiT frozen and shrinks what surrounds it. šŸ” What changed • Pixel decoder: 3 causal ConvNeXt blocks plus a linear layer replace the Wan VAE decoder, 63.9M vs 73.3M params, 1 ms vs 147 ms • KV repair network (about 2 ms) replaces the extra DiT pass that refreshes the KV cache • Systems work: persistent KV buffers, fused QKV, CUDA graphs; one DiT call goes from 41.6 ms to 16.4 ms šŸ“Š The numbers • FastFrames-S: 36 ms (27.8 RPS), VBench Total 84.23 vs 84.11 for Causal Forcing++ • FastFrames-L (14B): 45 ms on 8x B200, 158 ms on one GPU; Total 84.44 vs 84.42 for Krea-Realtime at 1,675 ms • Judge interactivity by responses per second, not FPS: a 50 ms target is 20 RPS āš ļø Caveats • S scores 79.92 on Semantic vs 80.20 for Causal Forcing++ • The pixel decoder is fine tuned per model; tested on B200s at 480p only • License tag is apple-amlr; LingBot parts are CC BY-NC-SA 4.0 šŸ› ļø Reproduce the profile from the repo: ``` uv run python evaluation/latency/benchmark.py ff-s ``` /apple/FastFrames-L /papers/2611.23226 Before you distill more steps, have you ranked your KV update and pixel decoding by wall time? šŸ“š Read it elsewhere • The deep dive on Medium: https://medium.com/p/589b4109f380 • The long read on Substack: https://sajjadpakdamansavoji.substack.com/p/apples-fastframes-shows-the-denoiser • The thread on X: https://x.com/saavoji/status/2123477247952879830 • Every version on sajjadpsavoji.github.io: https://sajjadpsavoji.github.io/blog#apples-fastframes
View all activity

Organizations

None yet