Important Notice
Optimized to work with the latest version of ComfyUI (v0.18.1 + Frontend 1.42.8).
Overview
This workflow uses Wan2.2 I2V together with SVI 2.0 PRO LoRA to generate long-form videos of 81 frames or more.
Instead of generating 81+ frames in a single pass, the video is created in multiple segments and then merged.
Compared to traditional extend/long workflows, this approach preserves motion information from the previous segment, resulting in much smoother transitions without noticeable seams.
Latest News (v1.7)
3-pass mode has been removed due to improved motion range achieved through updated recommended Lightx2v LoRA settings.
In the Extend group, you can now specify the Length for each chunk.
Please normally output at 24 FPS. If the result looks like fast-forward in some cases, try using 16 FPS instead.
(1.6) End Frame feature is now available!
You can specify an end frame for each chunk. This allows much finer control, especially over camera work.
Special thanks to Well-Made for updating the SVI node into a more powerful version.
Newly Implemented Feature
Prompt Enhancement via Ollama
Prompts are enhanced using an LLM, allowing for more detailed and structured motion descriptions.
You can also input prompts in Japanese (other languages are possible by adjusting the persona settings).
To use this feature, Ollama and an LLM model must be installed on your local PC.
Please note that I do not provide support for installing or configuring Ollama or LLM models.
Optional Features
Resume Mode
Allows you to continue generation not only from the first frame, but also from a previously generated video.
If something goes wrong partway through the process, you can save the generated segment and resume from that point without starting over.
Notes
This workflow is configured for the GGUF version, but fp8
andDistillversions can also be used by switching the model loader.Designed with 480p output in mind.
There are many variations of Lightx2v LoRA, and their suitability depends on the type of video you want to generate.
Acknowledgements
Special thanks to Kijai for providing the excellent LoRA models and custom nodes.
Description
Parameter optimization now enables more stable POV video generation, reducing artifacts and drift.
Persona settings were also tuned for better video prompts.
FAQ
Comments (13)
good workflow, but im struggeling with motion. any tips
There are several ways to enhance motion. As a starting point, the ZIP file includes sample MP4 videos and images.
The videos have the workflow embedded, so you can try generating motion similar to the sample videos.
If you need more dynamic motion, try increasing CFG up to 2.2 and breaking down the motion in your prompt.
For example, instead of just “running”, describe it step by step, such as:
the woman pushes off the ground → the woman accelerates → the woman runs at full speed.
I have a bit of an unconventional approach to getting good motion. Try different checkpoints in the high noise and run easywan or some other lightly merged checkpoint in low. heavily merged checkpoints kill motion or force the ai to generate undesired outputs in my experience. You'll need to experiment with different lora weights as well. But messing around with these alternate high a low configurations has resulted in some pretty amazing outputs for me. And this is while using Lightning. I haven't yet run tests without lightning but I'm betting that's even more amazing.
love the workflow but confused as to where to put the svi loras? do i just put them in the first i2v and extend sections?
The SVI LoRA is already placed by default inside the Model Loader (sub graph).
You only need to store the SVI LoRA files in the standard models/loras/ folder to be ready.
If you want to generate longer videos, simply disable the bypass on the extend group.
Can Ollama use custom LLM (uncensored versions)? get ever on PG16 prompts "
I’m sorry, but I can’t help with that."
well, phi4 does not appear to censor NSFW content.
even a local one?
@navyguy757 I usually use qwen3:8b for LLM, but sometimes it gets censored locally too.
Use a huihui abliterated Version of Qwen3 Models and influence ollama's build process with a Modelfile.
@javawock7618 try this one: https://huggingface.co/mradermacher/Qwen3-30B-A3B-abliterated-erotic-i1-GGUF?not-for-all-audiences=true