Papers
arxiv:2607.26004

Parallel Decoding Distillation for Fast Image and Video Generation

Published on Jul 28
· Submitted by
Julius Berner
on Jul 29
Authors:
,
,
,

Abstract

PDD accelerates diffusion and flow matching video generation by predicting multiple denoising steps per evaluation, improving speed, quality, and diversity without complex adversarial or distillation losses.

Generation in video diffusion or flow models is computationally expensive due to the slow and iterative sampling process. Current state-of-the-art (SOTA) acceleration methods heavily rely on variational score distillation (VSD) and adversarial losses to distill diffusion models into few-step generators. Albeit achieving high-quality video generation, these training losses are notoriously hard to optimize and suffer from mode collapse, leading to loss of video diversity and lack of motion. In this paper, we introduce Parallel Decoding Distillation (PDD), a simplified and scalable trajectory-based distillation method for fast inference of diffusion and flow matching models. Our architecture and training procedure are compatible with any pre-trained model and support sampling with a varying number of function evaluations (NFE). PDD accelerates generation by predicting multiple denoising steps per network evaluation. Conceptually, it learns a representation of the mean velocity without regressing its derivative using JVPs or finite-difference approximations. Our method achieves SOTA performance with 4-8 NFE on LTX-2.3 Text-to-Video/Audio, Wan 14B Text-to-Video, and Qwen-Image Text-to-Image. Moreover, PDD presents a significant improvement in generated video diversity.

Community

Paper submitter

This is an automated message from the Librarian Bot. I found the following papers similar to this paper.

The following papers were recommended by the Semantic Scholar API

Please give a thumbs up to this comment if you found it helpful!

If you want recommendations for any Paper on Hugging Face checkout this Space

You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: @librarian-bot recommend

A 60-second high-quality realistic cinematic commercial for "حضانة دعوة الحق الإسلامية بدمياط" (Daawet Al-Haq Islamic Nursery, Damietta), celebrating October Victory Day (نصر أكتوبر - Egypt's 6th of October victory).

STYLE: Photorealistic live-action style, ultra-realistic human characters, natural cinematic color grading, warm patriotic lighting with soft red-white-black Egyptian flag color accents, shallow depth of field, documentary-commercial hybrid look, NOT cartoon, NOT 3D animation — fully realistic human skin texture, realistic hair strands, realistic fabric wrinkles, natural eye blinking and micro-expressions.

CHARACTERS (Group of Children) — USE THIS EXACT DESCRIPTION THROUGHOUT: Five realistic Egyptian nursery children, ages 4-6, diverse natural appearances (varying hair lengths, light to olive skin tones), each wearing a plain white t-shirt with a small printed Egyptian flag on the chest, dark trousers or skirts, each holding a small handheld Egyptian flag (red-white-black horizontal stripes with the golden eagle emblem in the center).

CENTER BOY — USE THIS EXACT DESCRIPTION THROUGHOUT: A realistic Egyptian little boy, age 6, short black straight hair neatly combed, warm brown eyes, light olive skin tone, round friendly face, wearing the same plain white t-shirt with printed Egyptian flag and dark trousers as the group, holding his small Egyptian flag.

CENTER GIRL — USE THIS EXACT DESCRIPTION THROUGHOUT: A realistic Egyptian little girl, age 5, dark brown hair tied in two small neat pigtails with small red ribbons, warm brown eyes, light olive skin tone, round cheerful face, wearing the same plain white t-shirt with printed Egyptian flag and a dark skirt as the group, holding her small Egyptian flag.

TEACHER CHARACTER — USE THIS EXACT DESCRIPTION THROUGHOUT: A realistic Egyptian female teacher, early 30s, wearing a simple light beige hijab pinned neatly, a modest long-sleeve olive-green dress with a small nursery logo pin, warm proud smile.

SETTING: A bright decorated nursery classroom throughout the entire video. Walls decorated with Egyptian flag bunting, red-white-black balloon clusters in the corners, a large printed banner reading "نصر أكتوبر" in bold gold Arabic calligraphy on a red backdrop, warm golden sunlight streaming through classroom windows. A corner of the room has small wooden tables with white paper sheets showing printed outlines of the Egyptian flag, small trays of red, white, and black watercolor paints, paintbrushes, colored pencils, and a corkboard with finished flag drawings pinned up.

FULL ACTION SEQUENCE (timed precisely across 60 seconds):

[0:00-0:10 — Opening] Wide realistic establishing shot of the classroom door. The five children enter in a joyful little line, each holding their small Egyptian flag, the teacher welcoming them warmly at the door. They gather into a semi-circle, waving their flags gently in unison with big genuine smiles. Camera slowly pushes in as they cheer softly together, teacher clapping warmly in the background. No dialogue, only warm uplifting patriotic instrumental music (light drums and strings, no copyrighted anthem) and ambient joyful children's chatter and flag fabric fluttering.

[0:10-0:20 — Boy's line] Camera holds a natural medium shot on the center boy, who steps forward from the group, raising his flag higher, chest puffed with pride, other children waving flags gently in the blurred background. He speaks clearly, looking at the camera with pride: "إِحْنَا فَخُورِينَ بِنَصْرِ أُكْتُوبَر, لَمَّا الْجَيْش الْمِصْرِي عَدَّى قَنَاة السُّوَيْس بِبُطُولَة!" — spoken warmly and proudly, paced naturally, NOT robotic, NOT formal news-reader tone despite the diacritization. Background music continues softly under the dialogue.

[0:20-0:30 — Girl's line] Camera holds a natural medium shot on the center girl, stepping forward from the group, holding her flag close to her chest with both hands. She speaks clearly with warmth and innocent pride, tilting her head slightly: "وَفِي حَضَانَة دَعْوَة الْحَقّ بِدُمْيَاط, بِنِتْعَلِّم إِنَّ حُبّ الْوَطَن جُزْء مِنِ الْإِيمَان!" — spoken warmly, sincerely, naturally paced. She finishes with a bright smile and a small gentle flag wave. Background music continues softly.

[0:30-0:40 — Art activity] Wide realistic shot of the children now sitting together at the small wooden tables, focused and happy, carefully coloring the Egyptian flag outlines with red and black paint, tongues slightly poking out in concentration. Camera slowly pushes in on one child's hands painting the golden eagle emblem. The teacher walks by warmly, kneeling to admire one child's finished flag drawing, giving an encouraging thumbs-up, the child beaming with pride. No dialogue, ambient soft chatter and paintbrush sounds, music continues.

[0:40-0:50 — Group cheer] Wide realistic shot of all five children standing together in a line again, flags raised high, teacher standing proudly behind them. They raise their flags higher and cheer joyfully in unison, jumping gently in place with excitement. Camera slowly pushes in on their joyful faces. All children together, with natural overlapping joyful energy like real children cheering (not a perfectly synchronized chant): "يَحْيَا الْجَيْش! يَحْيَا مِصْر! يَحْيَا نَصْرِ أُكْتُوبَر!" Background music swells joyfully.

[0:50-1:00 — Closing] Warm group shot of all five children and the teacher together, calming from cheering into warm gentle smiles, flags held at rest, teacher's hand gently on the center boy's shoulder. The center boy speaks the closing line warmly and clearly, looking directly at the camera: "حَضَانَة دَعْوَة الْحَقّ بِدُمْيَاط.. بِتِرَبِّي جِيل يِحِبّ رَبُّه وَيِحِبّ بَلَدُه!" — spoken heartfelt and sincere, not rushed. Camera gently pulls back to a full warm wide shot of the whole group, soft warm golden light glow gently brightening the frame as the scene softly fades to end, music reaching a soft heartfelt peak then fading out with the visual fade.

CAMERA (overall): Smooth realistic handheld-style but stable cinematic movement throughout — push-ins during dialogue moments, wide establishing shots at the opening and art-activity segment, one final gentle pull-back at the very end. No shaky footage, no whip pans, no jarring cuts — natural continuous cinematic flow between segments.

LIGHTING (overall): Warm golden natural sunlight throughout, soft even exposure, natural skin tones, gentle warm highlight on flags, banner, and paint trays, intensifying into a soft warm glow in the final ten seconds, no harsh shadows on faces at any point.

AUDIO (overall): Warm uplifting patriotic instrumental background music continuous throughout (light drums, strings, and gentle rising melody, proud and joyful, subtly evoking Egyptian patriotic themes, absolutely no copyrighted anthem), layered with ambient natural sounds appropriate to each segment (children's chatter, flag fluttering, paintbrush sounds, joyful cheering).

DIALOGUE LANGUAGE NOTE: All spoken Arabic lines are fully diacritized with tashkeel for precise pronunciation accuracy, but must be performed with natural warm Egyptian child voices and authentic Cairo/Damietta colloquial rhythm — NOT a formal robotic news-reader delivery despite the written diacritization.

No text overlays. No on-screen captions. No subtitles beyond what naturally appears on the classroom banner prop itself. No cartoon elements. No 3D rendering. No copyrighted anthem music. No logos other than the nursery name spoken naturally in dialogue. No watermarks. No additional characters beyond the five children and the teacher throughout the entire video.

Sign up or log in to comment

Get this paper in your agent:

hf papers read 2607.26004
Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash

Models citing this paper 3

Datasets citing this paper 0

No dataset linking this paper

Cite arxiv.org/abs/2607.26004 in a dataset README.md to link it from this page.

Spaces citing this paper 9

Browse 9 spaces citing this paper

Collections including this paper 0

No Collection including this paper

Add this paper to a collection to link it from this page.