Home
Creators
Search Recent Random
Posts
Search DMs Popular Hash Lookup Tags Random
Importer
Import FAQ
Register Favorites Keys Review DMs Logout
Community
Telegram ThePornDude
Home Creators Posts Import Register Favorites Logout

FramePack Full Tutorial: 1-Click to Install on Windows - Up to 120 Second Image-to-Videos with 6GB (Patreon)

Published:
2025-04-17 20:29:54
Imported:
2026-06

Content

FramePack from legendary lllyasviel full Windows local tutorial with a very advanced Gradio app to generate consistent videos from images with as long as 120 seconds and as low as 6 GB GPUs. This tutorial will show you step by step how to install and use FramePack locall with a very advanced Graido app. Moreover, I have published installers for cloud services such as RunPod and Massed Compute for those GPU poor and who wants to scale.

🔗 Full Instructions, Installers and Links Shared Post (the one used in the tutorial) ⤵️

▶️ https://www.patreon.com/posts/click-to-open-post-used-in-tutorial-126855226

🔗 SECourses Official Discord 10500+ Members ⤵️

▶️ https://discord.com/servers/software-engineering-courses-secourses-772774097734074388

🔗 Stable Diffusion, FLUX, Generative AI Tutorials and Resources GitHub ⤵️

▶️ https://github.com/FurkanGozukara/Stable-Diffusion

🔗 SECourses Official Reddit - Stay Subscribed To Learn All The News and More ⤵️

▶️ https://www.reddit.com/r/SECourses/

🔗 MSI RTX 5090 TRIO FurMark Benchmarking + Overclocking + Noise Testing and Comparing with RTX 3090 TI ⤵️

▶️ https://youtu.be/uV3oqdILOmA

🔗 RTX 5090 Tested Against FLUX DEV, SD 3.5 Large, SD 3.5 Medium, SDXL, SD 1.5, AMD 9950X + RTX 3090 TI ⤵️

▶️ https://youtu.be/jHlGzaDLkto

Packing Input Frame Context in Next-Frame Prediction Models for Video Generation

FramePack, to train next-frame (or nextframe-section) prediction models for video generation. The FramePack compresses. Input frames to make the transformer context length a fixed number regardless of the video length.

Paper : https://lllyasviel.github.io/frame_pack_gitpage/pack.pdf

Project Page : https://github.com/lllyasviel/FramePack

Video Chapters

0:00 Intro: New Frame Pack Image-to-Video Framework

0:11 Mind-blowing Features & Why It's Different

0:24 Local Windows Install & Cloud Options (RunPod/Mass Compute)

0:39 Meet the Creator: Legendary LLL YassVL & His Track Record

1:02 Frame Pack Power: Up to 120s, Constant Speed/VRAM, Low VRAM/T-Cache Support

1:30 Live Demo: Generating 3 Videos at Once

1:52 CRITICAL Pre-Install Step: Windows Requirements Check

2:03 Preparing for Install: Moving File & Disk Space (50GB+)

2:14 How to Install: Extracting & Running the .bat File

2:28 Automatic Installation Process Overview

2:45 Technical Detail: Hunyuan Backend (Future 1.2.1 Support?)

2:57 Frame Pack Explained: Prediction Network & High Frame Counts

3:20 Live Generation Preview Feature (Second-by-Second Output)

3:37 Cloud vs. Local Performance: Delays & Comparisons

4:00 Installation Complete! How to Verify

4:16 Automatic Model Downloading on First Launch

4:27 Where Models Are Saved (models folder)

4:44 Troubleshooting Download Failures: Editing startup.bat Fix (Set to 0)

5:08 Download Fix Options: Restarting vs. Editing .bat File

5:27 Launching Pre-Installed Version & Checking Models Folder

5:39 Starting Application: Model Loading Process (CPU - VRAM)

5:53 Using Google Studio AI (Gemini Pro) to Generate Prompts

6:14 Crafting a "Hunyuan Image-to-Video Prompt"

6:34 Exploring Generation Options: T-Cache, Seed, Length, Steps

6:56 Options: Distill CFG, GPU Memory Preservation

7:11 Tool Tip: How to Install & Use nvitop for VRAM Monitoring

7:30 NEW FEATURE: Improved Video Quality Control (Encoding Fix)

7:41 Quality Settings: High, Medium, Low, Web Compatible

8:09 Starting Live Demo (10s) - Note: Version 1, More Features Coming!

8:26 Watching Live Generation: GPU Wattage Monitoring & Efficiency

8:51 Wattage Explained: Shared vs. Dedicated VRAM Power Draw

9:10 Live Preview: First Second Rendered & Displayed

9:23 Generating Subsequent Seconds & Benefit of Live Preview

9:36 Understanding T-Cache Speed (Slow Start, Fast Finish)

9:59 Live Preview: Second Second Rendered (Video Updated)

10:14 Watching Full Generation Progress & VRAM Check

10:35 Reviewing Completed Generations from Cloud Instances

10:55 Analysis Result 2: Another High Accuracy Example

11:03 Result 3 Failed - Likely Needs a Better Prompt

11:26 Understanding Output Video Resolution (Variable)

11:37 Future Plans: Resolution Control & Recap (Rush Release)

11:54 Cloud Setup: Overview & Mass Compute Details (Coupon)

12:21 Mass Compute GPUs: RTX A6000 Ada Rec, H100/A100 Pricing

12:48 RunPod Setup Details: Instructions & Required Template

13:06 Final Thoughts & Future Plans (Batch Processing, Presets)

13:25 Aiming for Feature Parity with 1.2.1 Application

13:39 Teaser: New Improved Super Upscaling App Coming Soon

14:01 Links & Thank You

Song: Cartoon, Jéja - Why We Lose (feat. Coleman Trapp) [NCS Release]

Music provided by NoCopyrightSounds

Free Download/Stream: http://ncs.io/whywelose

Watch: http://youtu.be/zyXmsVwZqX4

And few others : https://gist.github.com/FurkanGozukara/8fa15b67370a53af6394da239d2fce8e

Files

FramePack Full Tutorial: 1-Click to Install on Windows - Up to 120 Second Image-to-Videos with 6GB

FramePack from legendary lllyasviel full Windows local tutorial with a very advanced Gradio app to generate consistent videos from images with as long as 120 seconds and as low as 6 GB GPUs. This tutorial will show you step by step how to install and use FramePack locall with a very advanced Graido app. Moreover, I have published installers for cloud services such as RunPod and Massed Compute for those GPU poor and who wants to scale. 🔗 Full Instructions, Installers and Links Shared Post (the one used in the tutorial) ⤵️ ▶️ https://www.patreon.com/posts/click-to-open-post-used-in-tutorial-126855226 🔗 SECourses Official Discord 10500+ Members ⤵️ ▶️ https://discord.com/servers/software-engineering-courses-secourses-772774097734074388 🔗 Stable Diffusion, FLUX, Generative AI Tutorials and Resources GitHub ⤵️ ▶️ https://github.com/FurkanGozukara/Stable-Diffusion 🔗 SECourses Official Reddit - Stay Subscribed To Learn All The News and More ⤵️ ▶️ https://www.reddit.com/r/SECourses/ 🔗 MSI RTX 5090 TRIO FurMark Benchmarking + Overclocking + Noise Testing and Comparing with RTX 3090 TI ⤵️ ▶️ https://youtu.be/uV3oqdILOmA 🔗 RTX 5090 Tested Against FLUX DEV, SD 3.5 Large, SD 3.5 Medium, SDXL, SD 1.5, AMD 9950X + RTX 3090 TI ⤵️ ▶️ https://youtu.be/jHlGzaDLkto Packing Input Frame Context in Next-Frame Prediction Models for Video Generation FramePack, to train next-frame (or nextframe-section) prediction models for video generation. The FramePack compresses. Input frames to make the transformer context length a fixed number regardless of the video length. Paper : https://lllyasviel.github.io/frame_pack_gitpage/pack.pdf Project Page : https://github.com/lllyasviel/FramePack Video Chapters 0:00 Intro: New Frame Pack Image-to-Video Framework 0:11 Mind-blowing Features & Why It's Different 0:24 Local Windows Install & Cloud Options (RunPod/Mass Compute) 0:39 Meet the Creator: Legendary LLL YassVL & His Track Record 1:02 Frame Pack Power: Up to 120s, Constant Speed/VRAM, Low VRAM/T-Cache Support 1:30 Live Demo: Generating 3 Videos at Once 1:52 CRITICAL Pre-Install Step: Windows Requirements Check 2:03 Preparing for Install: Moving File & Disk Space (50GB+) 2:14 How to Install: Extracting & Running the .bat File 2:28 Automatic Installation Process Overview 2:45 Technical Detail: Hunyuan Backend (Future 1.2.1 Support?) 2:57 Frame Pack Explained: Prediction Network & High Frame Counts 3:20 Live Generation Preview Feature (Second-by-Second Output) 3:37 Cloud vs. Local Performance: Delays & Comparisons 4:00 Installation Complete! How to Verify 4:16 Automatic Model Downloading on First Launch 4:27 Where Models Are Saved (models folder) 4:44 Troubleshooting Download Failures: Editing startup.bat Fix (Set to 0) 5:08 Download Fix Options: Restarting vs. Editing .bat File 5:27 Launching Pre-Installed Version & Checking Models Folder 5:39 Starting Application: Model Loading Process (CPU - VRAM) 5:53 Using Google Studio AI (Gemini Pro) to Generate Prompts 6:14 Crafting a "Hunyuan Image-to-Video Prompt" 6:34 Exploring Generation Options: T-Cache, Seed, Length, Steps 6:56 Options: Distill CFG, GPU Memory Preservation 7:11 Tool Tip: How to Install & Use nvitop for VRAM Monitoring 7:30 NEW FEATURE: Improved Video Quality Control (Encoding Fix) 7:41 Quality Settings: High, Medium, Low, Web Compatible 8:09 Starting Live Demo (10s) - Note: Version 1, More Features Coming! 8:26 Watching Live Generation: GPU Wattage Monitoring & Efficiency 8:51 Wattage Explained: Shared vs. Dedicated VRAM Power Draw 9:10 Live Preview: First Second Rendered & Displayed 9:23 Generating Subsequent Seconds & Benefit of Live Preview 9:36 Understanding T-Cache Speed (Slow Start, Fast Finish) 9:59 Live Preview: Second Second Rendered (Video Updated) 10:14 Watching Full Generation Progress & VRAM Check 10:35 Reviewing Completed Generations from Cloud Instances 10:55 Analysis Result 2: Another High Accuracy Example 11:03 Result 3 Failed - Likely Needs a Better Prompt 11:26 Understanding Output Video Resolution (Variable) 11:37 Future Plans: Resolution Control & Recap (Rush Release) 11:54 Cloud Setup: Overview & Mass Compute Details (Coupon) 12:21 Mass Compute GPUs: RTX A6000 Ada Rec, H100/A100 Pricing 12:48 RunPod Setup Details: Instructions & Required Template 13:06 Final Thoughts & Future Plans (Batch Processing, Presets) 13:25 Aiming for Feature Parity with 1.2.1 Application 13:39 Teaser: New Improved Super Upscaling App Coming Soon 14:01 Links & Thank You Song: Cartoon, Jéja - Why We Lose (feat. Coleman Trapp) [NCS Release] Music provided by NoCopyrightSounds Free Download/Stream: http://ncs.io/whywelose Watch: http://youtu.be/zyXmsVwZqX4 And few others : https://gist.github.com/FurkanGozukara/8fa15b67370a53af6394da239d2fce8e

Comments

Justin

Any idea on how to make it available on 0.0.0.0 rather than localhost?

cool1

For new features, if it's generating it out of sequence (last bit first) and generating it in steps, a second at a time, and it's using the whole prompt for each section it generates, couldn't there be an option where you specified a different prompt for specific timecodes in the video, so you could make different things happen starting from different timecodes? eg. 0: the man picks up the sword, 10: the man walks away, etc.

Justin

Thank you! Just tried it, found it needs to be server_name='0.0.0.0',

Delvin

Sounds doable with comfyui node or programming side in gradio. Could set it up to self reference frames after certain count. Take that frame then generate with following prompt then next etc. populate all then stitch.

cool1

Maybe it could cache some text encoder stuff at the start (like one of the Flux trainer settings/configs does for the prompts). ie. cache each separate prompt & text encoding from it where there's different ones starting at different timecodes. If that couldn't work and make it faster then maybe the encode again and reload/deload could still be done. It would probably make things slower but it could still be an option if you wanted more control of what happened in the video.

Softtrack Live

I installed it, nicely done btw. when I do a picture to video it uses 24Gb Gpu, is not bad, but is not 6Gb, I tested 2 on the 3090 Suprime and RTX 5090. Is it possible to set somewhere how much gpu you use? it is very fast I must say. Greetings William

cool1

When it generates a new 1 second section of a video, is it taking the whole previous second into account or doing an image to video on the last frame from the previously generated 1 second video? Ideally it would take the previous 1 second video into account (or at least multiple frames of it rather than one frame), but when I generate an image 2 video of a scene that has a crowd of people in it, one person walks one way for one second then instantly switches to walking in the opposite direction for the next second which makes it seem like it's only taking 1 frame into account to continue from and so not being consistent for that part. Is that setting controllable? Also (like was maybe suggested in the video or somewhere else), it would be good if we can specify the resolution of the video we want to create (width, height and maybe aspect ratio selection) and maybe frame rate. Also if you set the video length to exactly 5 seconds in the UI, for me, the video it creates is 4 secs 834 millisecs (according to mediainfo), 30.0 fps, so not quite 5 seconds. So it seems that's a slight error. If there's a way to create Loras with it that would be good too. Also it could help if the output folder also could output a .txt file with the prompt that was used to create the video and the settings used.

cool1

Thanks. I don't get why the person walking switched directions instantly on (what seemed to be) the 1 second mark though if it was taking the whole previous second into context when generating the next frame. Maybe it was just a coincidence though. Another thing, if when generating a long video each rendered segment keeps getting longer, maybe it could have an option where it deletes the segments that are no longer needed and just the fullest are stored. eg. if I generate a 20 second video, the 17th video is 48.5 MB, the 18th is 52 MB, the 19th is 55.6 MB, the 20th is 59.5 MB. The file sizes would get even higher if we specified a higher resolution (if that option gets added) with no option for compression amount. And if we created a longer video (such as the max of 120 secs) we'd probably end up with many of those intermediate videos. But it seems only the most recent one is need for adding a new one onto when generating the video. If so, couldn't there be an option where, when it's creating the 20th video, that only the 19th video (that it will append to - maybe at the start) will exist at that time? I also wondered if, with longer videos, will the earlier parts of the video (that are already rendered as a file) keep getting re-compressed - or would it not re-compress them (maybe it's concatenating them on GOP boundaries or something ). I thought if it kept re-compressing them then the longer videos might get more compression artefacts. Someone on reddit said there's a way to do timecode based prompting (though I haven't seen the post that says how). I wonder if it's possible to do video to video where it creates a new video based on the original (changing something in it or changing the style). It might be easier to do that now it can do long videos. I also wondered, if there's no VRAM limit now for generating long videos, why the app limits it to 120 secs (ie. why not higher limit). If it wasn't for it keeping every intermediate 'up to' second video (with each getting bigger) then disc file size might not be a problem. Though maybe it would take very long on some PCs/GPUs.

Leonid Wolf

torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 108.00 MiB. GPU 0 has a total capacity of 12.00 GiB of which 6.12 GiB is free. Of the allocated memory 4.80 GiB is allocated by PyTorch, and 34.66 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables)

Leonid Wolf

I was rebooting. It still gives the same error.

Leonid Wolf

Really! 6gb gives an error. I set the parameter to 10gb and the generation worked. I probably didn't figure it out at first and I got an error. And the last question is, if the image resolution is 4k, does this affect the generation speed and memory usage?

Leonid Wolf

Thank you for your help because of your development

Leonid Wolf

I did several generations with the detailed prompta that florence gave me and realized that the model gives a memory error Offloading DynamicSwap_HunyuanVideoTransformer3DModelPacked from cuda:0 to preserve memory: 8 GB Loaded AutoencoderKLHunyuanVideo to cuda:0 as complete. VAE decoding completed in 61.45 seconds (2.276 sec/frame) Exception Code: 0xC0000005 LLVM ERROR: IO failure on output stream: bad file descriptor during generation. I didn't know about this and started writing prompta very briefly. Maybe my case will help you in the future.

Leonid Wolf

Yes, not only that. For example, if you set the value to 5 seconds, the generation fails after the second and third seconds of generation.

Loyd Keudem

The model refuse to download for me I have tried all your recommendations still don't work. It is like it freezes

Loyd Keudem

Alright thanks hope it will solve the issue

Loyd Keudem

I just want to let you know that most of the installer use this pip install --pre torch==2.7.0.dev20250311 torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128 Which do no more exist since it is stable so need to update those installer