Simon Langrieger (@simlang_) profile picture

Machine Learning Research Engineer @ Pruna AI

Munich
Joined November 2024
Photo by Simon Langrieger (@simlang_)
Simon Langrieger retweeted
P-Video-Upscale just got faster and cheaper! โšก Introducing Turbo mode: faster video upscaling at 65% lower cost compared to Standard (up to 30 fps), with a slight trade-off in quality. Here's what you need to know: - ๐—ฃ๐—ฟ๐—ถ๐—ฐ๐—ถ๐—ป๐—ด: Starting at just $0.0035/s - ๐—ฆ๐—ฝ๐—ฒ๐—ฒ๐—ฑ: Faster processing across resolutions - ๐—ฅ๐—ฒ๐˜€๐—ผ๐—น๐˜‚๐˜๐—ถ๐—ผ๐—ป: 720p, 1080p, 2K, or 4K - ๐—™๐—ฟ๐—ฎ๐—บ๐—ฒ ๐—ฟ๐—ฎ๐˜๐—ฒ: Up to 60 fps, at no extra cost in Turbo mode Available via our inference partners @scenario_gg, @replicate, @WiroAI, @TellersAI, @eachlabs, @wavespeed_ai, Sociaro, and @runware. ๐Ÿ“š Model docs: docs.api.pruna.ai/guides/modโ€ฆ ๐Ÿงฉ Try it through the API or Playground: dashboard.pruna.ai/login
6
21
5
49
3,844
Simon Langrieger retweeted
Today, we open-source P-Q-Image-2.1-Remove-Object, 3 LoRA adapters for Qwen-Image-2.1 by @Alibaba_Qwen. Remove the object you point at, along with its shadow or reflection, and change nothing else! - ๐—›๐—ผ๐˜„ ๐—ฒ๐—ณ๐—ณ๐—ถ๐—ฐ๐—ถ๐—ฒ๐—ป๐˜ ๐—ถ๐˜€ ๐—ถ๐˜? Remove objects in 6 steps instead of 40, with no CFG. Use the 6-step adapter as the default. An 8-step adapter is also available, and a 40-step adapter runs only on the base model, without the few-step model's drift. - ๐—›๐—ผ๐˜„ ๐—ด๐—ผ๐—ผ๐—ฑ ๐—ถ๐˜€ ๐—ถ๐˜? Shadows and reflections just go away along with the object, even when they sit outside the outline. The rest of the photo stays put and lines up with the input, so you can use the result as is. - ๐—›๐—ผ๐˜„ ๐—ฑ๐—ผ๐—ฒ๐˜€ ๐—ถ๐˜ ๐˜„๐—ผ๐—ฟ๐—ธ? The adapters load directly on top of Qwen-Image-2.1, so the pipeline stays unchanged. The few-step variants stack on our P-Q-Image-2.1 distillation LoRA. Mark the object with a rough red outline, or just name it in the instruction. ๐Ÿ‘‰ Blog: pruna.ai/blog/p-q-image-2.1-โ€ฆ ๐Ÿค— Open Weights: huggingface.co/PrunaAI/P-Q-Iโ€ฆ โšก๏ธ ComfyUI: huggingface.co/PrunaAI/P-Q-Iโ€ฆ
10
27
10
182
8,307
Simon Langrieger retweeted
Introducing P-Video-2-Pro-Replace, based on @MiniMax_AI H3! ๐ŸŽฌ Replace the people. Keep the scene. Swap one or two people in an existing video using reference images, while preserving the original motion, camera, framing, background, and audio And for launch, it's 50% off for the first two weeks, until November 19 at 6:00 PM CET! Here's what you need to know: - ๐—ฃ๐—ฟ๐—ถ๐—ฐ๐—ถ๐—ป๐—ด: Starting from $0.05/s at 480p - ๐—ฆ๐—ฝ๐—ฒ๐—ฒ๐—ฑ: ~1.5โ€“2 minutes for a 10-second video - ๐—ฅ๐—ฒ๐—ณ๐—ฒ๐—ฟ๐—ฒ๐—ป๐—ฐ๐—ฒ๐˜€: Replace up to two people using reference images - ๐——๐˜‚๐—ฟ๐—ฎ๐˜๐—ถ๐—ผ๐—ป: Supports videos up to 120 seconds Available via our inference partners @runware, @wavespeed_ai, Sociaro, @TellersAI, @WiroAI, @Segmind_ai, @replicate, and @eachlabs ๐Ÿ“š Model docs: docs.api.pruna.ai/guides/modโ€ฆ ๐Ÿงฉ Try the model through the API or Playground: dashboard.pruna.ai/login
22
22
26
162
21,027
Simon Langrieger retweeted
Spent the last weeks doing model brain surgery on the MiniMax-H3 video model, turning it into a world-class image model. Everything is open, and we wrote a technical blog post about how we did it. Go read it. pruna.ai/blog/oss-p-h3-imageโ€ฆ
2
6
1
14
675
Simon Langrieger retweeted
Today, we open-source P-H3-Image, an image generation and editing model built from MiniMax-H3 by @MiniMax_AI. A SOTA video model, turned into an efficient open image model. - ๐—›๐—ผ๐˜„ ๐—ฒ๐—ณ๐—ณ๐—ถ๐—ฐ๐—ถ๐—ฒ๐—ป๐˜ ๐—ถ๐˜€ ๐—ถ๐˜? Generate or edit images 5ร— faster, in 8 steps instead of 50. A 4 MP image takes 2.6 s instead of 12.4 s. The model is 7ร— smaller than H3, so high-resolution images are a light download and run on a much smaller GPU. - ๐—›๐—ผ๐˜„ ๐—ด๐—ผ๐—ผ๐—ฑ ๐—ถ๐˜€ ๐—ถ๐˜? Strongest at photorealism, documentary and editorial photography, film stills, and graphic styles with short quoted text. It renders 2 to 16 MP and edits with up to 8 input images. -๐—›๐—ผ๐˜„ ๐—ฑ๐—ผ๐—ฒ๐˜€ ๐—ถ๐˜ ๐˜„๐—ผ๐—ฟ๐—ธ? We reworked every component of H3: the prompt upsampler, the text encoder, the diffusion schedule, the transformer, and the decoder. Use it with Diffusers or ComfyUI. ๐Ÿ“„ Tech Report: pruna.ai/blog/oss-p-h3-imageโ€ฆ ๐Ÿค— @huggingface Open Weights: huggingface.co/PrunaAI/P-H3-โ€ฆ ๐Ÿ‘‰ @ComfyUI Native Support: huggingface.co/PrunaAI/P-H3-โ€ฆ
107
175
58
1,620
163,749
Simon Langrieger retweeted
Introducing P-Video-Upscale! ๐ŸŽฌ Fast video upscaling to 720p, 1080p, 2K, or 4K, while preserving the original aspect ratio and audio. Need smoother motion? Increase the frame rate to up to 60 fps. And for launch, it's 50% off for the first two weeks, until November 12 at 18:00 CET! Here's what you need to know: โ€ข ๐—ฃ๐—ฟ๐—ถ๐—ฐ๐—ถ๐—ป๐—ด: Starting from $0.01/s at 720p โ€ข ๐—ฆ๐—ฝ๐—ฒ๐—ฒ๐—ฑ: ~1.7s per second of output video, depending on resolution โ€ข ๐—™๐—ฟ๐—ฎ๐—บ๐—ฒ ๐—ฟ๐—ฎ๐˜๐—ฒ: Keep the original or interpolate up to 60 fps โ€ข ๐—ง๐˜‚๐—ฟ๐—ฏ๐—ผ ๐—บ๐—ผ๐—ฑ๐—ฒ: Faster processing with a slight quality trade-off No need to regenerate a video just to get it in higher resolution. Available via our inference partners @scenario_gg, @replicate, @WiroAI, @TellersAI, @eachlabs, @wavespeed_ai, Sociaro, and @runware. Benchmarks by our benchmark partners @datapointai and @RapidataAI are coming soon. ๐Ÿ“š Model docs: docs.api.pruna.ai/guides/modโ€ฆ ๐Ÿงฉ Try the model through the API or Playground: dashboard.pruna.ai/login
46
202
44
738
530,581
Simon Langrieger retweeted
Today, we open-source v0.2 of P-Q-Image-2.1, two new 6-step LoRA adapters for Qwen-Image-2.1 by @Alibaba_Qwen. The first distillation that removes the prompt-upsampling step! - ๐—›๐—ผ๐˜„ ๐—ฒ๐—ณ๐—ณ๐—ถ๐—ฐ๐—ถ๐—ฒ๐—ป๐˜ ๐—ถ๐˜€ ๐—ถ๐˜? Generate or edit images in just 6 steps instead of 40. Use 6-steps for preserving inputs when editing and 6-steps-upsampled for improved visual quality and shorter end-to-end generation. - ๐—›๐—ผ๐˜„ ๐—ด๐—ผ๐—ผ๐—ฑ ๐—ถ๐˜€ ๐—ถ๐˜? 6-step distillation provides improved quality with few generation steps. 6-steps-upsampled is the first distillation that removes the need for a separate prompt-upsampling step. This means 41 to 80 seconds saved, depending on your prompt style. - ๐—›๐—ผ๐˜„ ๐—ฑ๐—ผ๐—ฒ๐˜€ ๐—ถ๐˜ ๐˜„๐—ผ๐—ฟ๐—ธ? The LoRA adapters load directly on top of Qwen-Image-2.1, so the pipeline stays unchanged. Few-step distillation cuts the diffusion cost, and the upsampled adapter also removes the prompt-upsampling step before image generation. ๐Ÿค— Open Weights: huggingface.co/PrunaAI/P-Q-Iโ€ฆ ๐Ÿ‘‰ HF Space: huggingface.co/spaces/PrunaAโ€ฆ
15
39
6
221
18,602
Simon Langrieger retweeted
Introducing P-Video-2-Pro-Avatar, based on @MiniMax_AI H3 ! Turn a single portrait into a realistic talking video with natural expressions and accurate lip sync. And for launch, it's 50% off for the first two weeks, until November 5 at 18:00 CET! โ€ข ๐—ฃ๐—ฟ๐—ถ๐—ฐ๐—ถ๐—ป๐—ด: $0.02/s at 480p ยท $0.04/s at 768p โ€ข ๐—ฆ๐—ฝ๐—ฒ๐—ฒ๐—ฑ: ~0.97 s at 480p ยท ~2.2 s at 768p per second of output video โ€ข ๐—œ๐—ป๐—ฝ๐˜‚๐˜: One portrait with a text script or uploaded audio โ€ข ๐—ฉ๐—ผ๐—ถ๐—ฐ๐—ฒ๐˜€: 30 voices across 10 languages โ€ข ๐——๐˜‚๐—ฟ๐—ฎ๐˜๐—ถ๐—ผ๐—ป: 1sโ€“3min at 24 fps, in 480p or 768p โ€ข ๐—–๐—ผ๐—ป๐˜๐—ฟ๐—ผ๐—น: Voice delivery, scene motion, seed, and prompt upsampling One image, two ways to bring it to life. Use your own audio or generate speech directly from text. Available via our inference partners @replicate, @eachlabs, @Cloudflare, @scenario_gg, @wavespeed_ai, @WiroAI, Sociaro, @TellersAI, and @togethercompute. ๐Ÿ“š Model docs: docs.api.pruna.ai/guides/modโ€ฆ ๐Ÿงฉ Try the model through the API or Playground: dashboard.pruna.ai/login
69
302
49
734
632,513
Simon Langrieger retweeted
p-video-2-pro is now the undisputed king of video models on @DesignArena. #1 ELO, Only model on Pareto plots
5
1
12
408
it's just fun playing around with it! Iterating with it goes so fast, insane! Everything with a *single* prompt and a few references ๐Ÿซช greetings from eisbachwelle in munich ๐Ÿค™
Introducing P-Video-2-Pro-Reference based on @MiniMax_AI H3! ๐ŸŽฌ Combine images, videos, and audio references to guide how your video looks, moves, and sounds. Keep characters consistent, transfer motion, and bring different visual elements together in a single generation. โ€ข ๐—œ๐—ป๐—ฝ๐˜‚๐˜: Text prompt with image and/or video references, plus optional audio โ€ข ๐—ฅ๐—ฒ๐—ณ๐—ฒ๐—ฟ๐—ฒ๐—ป๐—ฐ๐—ฒ๐˜€: Up to 9 images, 3 videos, and 3 audio files โ€ข ๐——๐˜‚๐—ฟ๐—ฎ๐˜๐—ถ๐—ผ๐—ป: 5โ€“15s at 480p or 768p โ€ข ๐—ฃ๐—ฟ๐—ถ๐—ฐ๐—ถ๐—ป๐—ด: Output starting at $0.015/s in Cost mode, with reference inputs billed separately โ€ข ๐—–๐—ผ๐—ป๐˜๐—ฟ๐—ผ๐—น: First and last frames, seed, aspect ratio, and prompt upsampling Two modes: Speed and Cost. Choose the trade-off that fits your workflow. Available via our inference partners @Cloudflare, @scenario_gg, @replicate, @WiroAI, @TellersAI, @eachlabs, @wavespeed_ai, @inference_sh, @_segmind, @togethercompute, @EdenAIco, and @runware Validated by our benchmark partners @datapointai, @RapidataAI, and @DesignArena ๐Ÿ“š Model docs: docs.api.pruna.ai/guides/modโ€ฆ ๐Ÿงฉ API: dashboard.pruna.ai/login
2
10
119
Simon Langrieger retweeted
P-video-2-pro slowly climbed up to #1 ELO model on quality! Not only that, it is also the fastest and cheapest model on the market. There is literally 0 reason to use any other video model anymore.
Image-to-Video: P-Video-2-Pro is the #1 ELO model on @DesignArena. โ€ข Speed: #1 ยท 1328 Elo โ€ข Quality: #2 ยท 1326 Elo โ€ข Cost: #3 ยท 1324 Elo One model, three modes. Depending on whether you want to optimize for cost, speed, or quality. Benchmark by @DesignArena. ๐Ÿ“š Documentation: buff.ly/bHQHKV0 ๐Ÿงฉ API: dashboard.pruna.ai ๐Ÿฅ‡ P-Bench: buff.ly/81e3AuS
3
8
247
Simon Langrieger retweeted
P-Video-2-Pro now has a Cost mode! Get the same output quality as Speed mode, now starting from just $0.01/s. Optimised for cost when generation time matters less. And for launch, itโ€™s 50% off until October 13! โ€ข ๐—œ๐—ป๐—ฝ๐˜‚๐˜: Text or first-frame image, with optional last-frame conditioning โ€ข ๐——๐˜‚๐—ฟ๐—ฎ๐˜๐—ถ๐—ผ๐—ป: 5โ€“15s with generated audio โ€ข ๐—ฃ๐—ฟ๐—ถ๐—ฐ๐—ถ๐—ป๐—ด: $0.01/s at 480p ยท $0.025/s at 768p โ€ข ๐—–๐—ผ๐—ป๐˜๐—ฟ๐—ผ๐—น: First + last frame, seed, aspect ratio, and prompt upsampling Three modes: Cost, Speed, Quality. Pick the tradeoff that fits your workflow. Available via our inference partners @Cloudflare, @Scenario_gg, @replicate, @wiroai, @TellersAI, @eachlabs, @edenaico, @runware, @wavespeed_ai, @inference_sh, @togethercompute, and @prodialabs. Validated by our benchmark partners @datapointai, @RapidataAI, and @DesignArena. ๐Ÿ“š Model docs: docs.api.pruna.ai/guides/modโ€ฆ ๐Ÿงฉ API: dashboard.pruna.ai/login
16
16
9
108
27,458
Simon Langrieger retweeted
P-Video-2-Pro dominates the quality-speed frontier on @DesignArena for both Text-to-Video and Image-to-Video! Image-to-Video: โ€ข Quality: #3 ยท 1325 Elo ยท 8.0s โ€ข Speed: #2 ยท 1325 Elo ยท 4.5s Text-to-Video: โ€ข Quality: #4 ยท 1308 Elo ยท 8.6s โ€ข Speed: #6 ยท 1298 Elo ยท 4.7s Thatโ€™s a 19โ€“22 place jump from P-Video-2.
12
23
6
125
10,037
Simon Langrieger retweeted
Shoutouts to our distillation wizard @simlang_. It was a lot of fun to build together, and I'm proud of how far we got in the limited time we had.
1
3
55
Simon Langrieger retweeted
Today, we open-source Pruna-Qwen-Image-2.1, a set of a few-step LoRA adapters that make Qwen-Image-2.1 by @Alibaba_Qwen up to 6.3ร— faster for image generation and editing. - ๐—›๐—ผ๐˜„ ๐—ฒ๐—ณ๐—ณ๐—ถ๐—ฐ๐—ถ๐—ฒ๐—ป๐˜ ๐—ถ๐˜€ ๐—ถ๐˜? Generate or edit images in just 5 or 8 steps instead of 40. Choose 5 steps for maximum speed or 8 steps for the best balance. - ๐—›๐—ผ๐˜„ ๐—ด๐—ผ๐—ผ๐—ฑ ๐—ถ๐˜€ ๐—ถ๐˜? The 8-step adapter is our recommended default. It supports text-to-image and single- or multi-image editing at 1K resolution, with up to three reference images. - ๐—›๐—ผ๐˜„ ๐—ฑ๐—ผ๐—ฒ๐˜€ ๐—ถ๐˜ ๐˜„๐—ผ๐—ฟ๐—ธ?The LoRA adapters load directly on top of Qwen-Image-2.1, while the pipeline remains unchanged. Each adapter uses its own optimized sigma schedule and runs without CFG. ๐˜›๐˜ฉ๐˜ช๐˜ด ๐˜ช๐˜ด ๐˜ซ๐˜ถ๐˜ด๐˜ต ๐˜ฐ๐˜ถ๐˜ณ ๐˜ง๐˜ช๐˜ณ๐˜ด๐˜ต ๐˜ณ๐˜ฆ๐˜ญ๐˜ฆ๐˜ข๐˜ด๐˜ฆ, ๐˜ธ๐˜ช๐˜ต๐˜ฉ ๐˜ฆ๐˜ท๐˜ฆ๐˜ฏ ๐˜ฃ๐˜ฆ๐˜ต๐˜ต๐˜ฆ๐˜ณ ๐˜ท๐˜ฆ๐˜ณ๐˜ด๐˜ช๐˜ฐ๐˜ฏ๐˜ด ๐˜ฐ๐˜ฏ ๐˜ต๐˜ฉ๐˜ฆ ๐˜ธ๐˜ข๐˜บ. ๐Ÿš€ Want to try the fast open-source integration? Check it here on Diffusers: buff.ly/1CpiRPp Want to try the fastest image generation and editing endpoint? Check P-Image-Ideogram and P-Image-Edit here on API: buff.ly/gw5tvUV
32
102
11
949
93,207
Simon Langrieger retweeted
In case you ever wondered how I got good at optimizing models... hereโ€™s my story. Made with P-Video-2-Pro for less than $2 and in under 2 minutes.
3
4
1
14
925
Simon Langrieger retweeted
P-Video-2-Pro by @PrunaAI built on MiniMax H3 ๐Ÿ’œ Now on available on all our inference partners @replicate, @Cloudflare, @eachlabs, @scenario_gg, @TellersAI, @wavespeed_ai, @wiroai, @runware, @togethercompute. ๐Ÿ“š Docs: docs.api.pruna.ai/guides/modโ€ฆ ๐Ÿงฉ API: dashboard.pruna.ai/login
1
2
1
14
999
Simon Langrieger retweeted
There are already some very fast @MiniMax_AI H3 implementations out there, both open- and closed source. But how much performance can you get per dollar? Today @PrunaAI is launching P-Video-2-Pro, fast, affordable and built for high-quality video generation at scale
3
6
1
8
452
Simon Langrieger retweeted
Today, we launch P-Video-2-Pro, based on MiniMax H3: The best quality. The best speed. Text or image in. Video and audio out. โ€ข ๐—œ๐—ป๐—ฝ๐˜‚๐˜: Text or first-frame image, with optional last-frame conditioning โ€ข ๐—ฆ๐—ฝ๐—ฒ๐—ฒ๐—ฑ: ~2.0s for a 5s clip at 480p ยท ~4.3s at 768p in Speed mode โ€ข ๐—ฃ๐—ฟ๐—ถ๐—ฐ๐—ถ๐—ป๐—ด: From $0.02/s Speed ยท $0.04/s Quality โ€ข ๐— ๐—ผ๐—ฑ๐—ฒ๐˜€: Speed or Quality โ€ข ๐—–๐—ผ๐—ป๐˜๐—ฟ๐—ผ๐—น: First and last frame, seed, aspect ratio, and prompt upsampling with Off, Turbo, or Max Available via our inference partners @replicate, @Cloudflare, @eachlabs, @scenario_gg, @TellersAI, @wavespeed_ai, @wiroai, @runware, and @togethercompute Validated by our benchmark partners @datapointai, @RapidataAI, and @DesignArena. Benchmarks to be released soon ๐Ÿ‘€ ๐Ÿ‘‰ Free Playground: playground.pruna.ai/ ๐Ÿ“š Model docs: docs.api.pruna.ai/guides/modโ€ฆ ๐Ÿงฉ API: dashboard.pruna.ai/login
39
55
23
354
105,028
Simon Langrieger retweeted
P-Image-Ideogram, built with @ideogram_ai, now has a Very High mode. Very High uses test-time scaling to push final image quality further. Built for hero assets, difficult prompts, and quality-critical generations when people and products have to pop. โ€ข ~5.55s at 1K ยท $0.033/image (1K) ยท $0.066/image (2K) โ€ข Five modes from Very Low ($0.003) through Very High Again, benchmarks agree: This new very high mode is on efficiency-quality based on human and automated evals from @datapointai, @DesignArena, and P-Judger. Full benchmarks results follow in a later post on P-Bench. Available on @cloudflare @ComfyUI @GammaApp @inference_sh @kittldesign @LeonardoAi @lovart_ai @magnific @picsart @replicate @runware @scenario_gg @togethercompute @wavespeed_ai @wiroai ๐Ÿ‘‰ Free Playground: playground.pruna.ai/p-image-โ€ฆ ๐Ÿงฉ API: docs.pruna.ai/ ๐Ÿ“š Model page: pruna.ai/p-image-ideogram
8
10
7
67
10,496