v3.0 Is Finally Here!

make sure you re-download it if you got the file within the first few hours. I've made some last-minute tweaks and updated the file as soon as I could

I just like the simplicity of KSampler nodes. Feel free to add/remove nodes as you see fit!

@Asaghon I do it because I run multiple iterations of the image using different prompting techniques and/or different samplers, schedulers, and shift values for testing. if I start with the same image each time, I can use the scientific process to document and adjust for changes

It’s a wan2.1 Lora that still works in Wan2.2. Just keep the strength of Detailz Lora on the lower end.

link: https://civitai.com/models/1385506/detailz-wan-detail-enhancer-for-wan-videos

yeah, with 8GB RAM on an older card like that I suggest using the Q4 quantized (GGUF) models and instead of using the full fp8_scaled text encoders, there are quantized (GGUF) text encoder models you can plug into the formula as well. Your GTX1080 will likely have significantly slower generation times as well as some potential loss in quality and prompt adherence because of the structure and age of the technology (GDDR5X) behind the current generation (GDDR7)

I haven't been able to get face-swap or face-restore nodes to work 100% of the time. Therefore, I do not add them to my publicly released versions of my workflows. You might want to try disconnecting the Color Match node to utilize a face node more effectively, or set up the face-swap/restore nodes in a separate post-processing flow/extension.

I have experienced this before when my ComfUI got too bloated with other models and cached files/images. I would recommend a fresh installation of ComfyUI Windows Portable, or alternatively a fresh Manual Install of ComfyUI with its own virtual environment. Make a backup of ONLY your most important models, unet, text encoders, and vae. Everything else (especially Custom Nodes) should be completely deleted until a new fresh install is complete. Make sure you also delete temporary files and shader cache from your computer system prior to reinstalling ComfyUI.

@Renaphine I usually recommend a virtual environment, because it prevents "system creep." That's where your python dependencies can "creep" into the rest of your computer and accidentally override or overwrite other necessary dependencies in your computer. If you work with multiple AI tools and programs, it's especially important to be mindful of that. Therefore, a Virtual Environment (Manual Install) of ComfyUI is my recommendation. It creates a completely separate environment for that specific system and set of files. It blocks "system creep" by requiring that Virtual Environment to be activated before any files can be added or removed within it.

@chunksmobile756 try updating your Python Dependencies if needed, and/or you may need to delete everything (keep a backup of your most important LoRAs, diffusion models, unet models, vaes, and generated content) and re-install with a clean copy of ComfyUI Portable or Manual Installation methods. I recommend manual for reducing these types of issues. CogniBuild aka Get Going Fast on YouTube is very informative on how to do it.

It means a lot to know that you’ve found this workflow to be preferable!

The Sigmas Preview graph is a visualization of the de-noise process. If you're disabling the Lightning LightX2V Lora, you can play around with more schedulers such as Bong Tangent and Beta57. To adjust for better quality with less noise, you can change the Model Shift weights (make sure they are equal in both High and Low noise settings). Your scheduler with a proper shift weight should make the graph look something like the phot I posted in the Full Info Article about this version: https://civitai.com/articles/21684/torstens-wan22-i2v-gguf-workflow-version-30

***If you are ENABLING the LightX2V Speed/Lightning Lora, I do not recommend changing model weights or schedulers too much***

Thank you for the kind words! To prevent loop-back, make sure you've enabled the RifleXRoPE nodes for both high-noise and low-noise sections. Try to increase or decrease the value by increments of 1-2 until you get satisfactory results. Wan models tend to try to loop back no matter what, and RifleXRoPE can lower the amount of loop-back. It may not get rid of it completely, but you should end up with more usable frames if you're looking to do a sequential video generation (like telling a story through multiple videos strung together for a final project)

@iamstreamin that's odd. Luckily you don't NEED to enable those VRAM and Cache cleaning nodes as long as you keep randomizing the seed or changing some parameters in the flow on a fixed seed for testing.

I may also recommend updating your ComfyUI. If that doesn't work, you may want to try updating your Python Dependencies... but be careful with Python Dependency updates. Depending on your specific version of ComfyUI and how you installed it, they may try to overwrite each other which would then require a re-installation. If you have to re-install, try the Manual method with virtual environment if possible.

I’m considering making one for t2v based on what I now know! I’m just a hobbyist with a busy full time career so it may take a bit to get it out there publicly. There are multiple new projects in the works that I want to do.

@vicaut I’ve adapted my steps and lora strategies recently. Enable LightX2V on both high and low, 8 steps total, (4 high, 4 low) dpmpp_2m_sde sampler with beta scheduler, plus a very detailed prompt with descriptive instructions for realism, camera lens and angle…. I barely need any detail-enhancing loras anymore.

@vicaut The reason I use 8 steps is because for my specific project that I'm currently working on, I am looking for maximum quality out of the Quantized models (currently using Q8 on my 16-GB NVidia RTX 5070) - but your values seem like a good balance as well, so if it works for you, then go for it! I'm happy that people are getting some versatility with my workflows, and using it to test new ideas and methods!

@rajivenator639 yeah, I believe the other GGUF loader is now deprecated based on the new Comfyui-Front-End version and Node2.0 update. I had to use the same fix on mine.

Yeah, I was going to suggest to use native Loaders since RGThree have been deprecated now. I may make an updated V4 soon, if I have the time. I've personally switched away from GGUF in favor of native FP8 support on my Nvidia 5060ti Blackwell GPU.

@fstubbs1701 I appreciate the compliment, and I'm glad you get something meaningful out of the workflows I develop! Hopefully in a few weeks I'll put a small workflow update through. Until then, V3 is till very stable, just update the RGThree nodes to the latest version — the developer released a new update just a few days ago at the time of writing this comment — or swap the RGThree nodes out for native ones. I've also noticed that Video Combine can be a little finicky sometimes so, again, native Create Video and Save Video nodes can be put in place of VideoCombine, plus it gives you the option to use GetAnyFrame to grab a still frame image from your rendered video and connect it to Preview Image or Save Image to see/save the still frame.

Firstly, sequential rendering (last frame being used as a new first frame) will always produce some artifacts and image burn if you’re rendering more than 3-5 clips in a sequence. You may want to generate a new master frame using image editing on Flux.2 Klein9b for every 4th or 5th clip, and editing the video together afterwards in Davinci Resolve (it’s free). DO NOT USE THE LAST FRAME AS A NEW FIRST. Instead, use a near-last frame (the third-to-last frame usually gives me better results) for sequencing multiple clips together.

In other instances, image burn and severe color shifting is usually caused by too many steps on Low Noise pass, incorrect hand-off between the High and Low phases, or CFG being set too high. Here are a few things to try:

1. You can swap out the k-sampler setup with some custom sigmas, or just adjust the Model SD3 sampling value. It’s best for i2v model to trade-off from High Noise to Low Noise at exactly 0.900 sigma value, and t2v model to trade-off at 0.875. Make your High Noise end at the appropriate value and then start Low Noise.

2. Check if you’re using a Wan2.2 model that already has Lightning/Lightx2v architecture in it. CFG must ALWAYS be 1.0 when Lightning architecture is present in the model or when using those Lightning/distill LoRAs.

— If your model already has it built-in, DO NOT add additional Lightning/Distill/Lightx2v LoRAs. WAN2.2 Remix (both i2v and t2v models) has that built-in, and Wan2.2-Smoothmix T2v has it too (but not Smoothmix i2v).

— If the model you use does NOT have Lightning/Distill architecture built-in, try lowering the Lightning LoRA strengths to 0.75 or even down to 0.5. You may need to add slightly more steps when doing this. (This will also add some length to the total render times, but it will produce less “bake” in the output)

3. If those two approaches don’t fix it, perhaps you have too many steps on the Low Noise Pass. Try reducing the number of steps during Low Noise by a very small amount.