GOONING WORKFLOW FOR THE VRAM POOR!

Some recommended upscalers:

RealESRGAN_x2plus.pth (3x faster, better details/less 'grainy', comparing 2x-AnimeSharpV4_RCAN)

2x-AnimeSharpV2_MoSR_Sharp.pth (2.5x faster)

2x-AnimeSharpV2_MoSR_Soft.pth (2.5x faster)

BSRGANx2.pth (a bit slower, but gives better details than RealESRGAN_x2plus. in case you need very detailled video and not in a hurry)

There are some others interesting upscalers around (like 2x_Text2HD_v.1-RealPLKSR.pth and 2xVHS2HD-RealPLKSR.pth), but these are very dedicated upscalers. (Text and VHS restoration only)

Thank you for the nice workflow :)

Cannot execute because a node is missing the class_type property.: Node ID '#190'

Any suggestions on what to do if I want to run this on 8GB Rtx4060 and 16 GB 5600mhz ram?

I loved it, thank you so much!

Do you think there's a way to add a feature to the workflow to increase the video generation time? At least by 10 seconds?

That's a good one! With 8GB rtx3070 and 16GB ram it took only 11 minutes from start to upscaled and interpolated 544*960 res video. The only downside is 59GB swap file it required.

ModuleNotFoundError: No module named 'sageattention'
please help to solve this

PainterI2V is missing and idk what the hell it is

Im facing this problem, anyone know how to solve this? plz help me......

KSamplerAdvanced

'utf-8' codec can't decode byte 0xc7 in position 3: invalid continuation byte

Hi everyone, I've already installed all the dependencies mentioned by Taaty Sin. When I disable frame interpolation and upscaling, I can run the workflow with three LoRAs loaded without any issues. However, once I enable interpolation and upscaling – using the recommended settings and two LoRAs (High+Low) – I run out of memory at the second KSampler (OOM, about 800MB short according to the log). My suspicion is that ComfyUI loads the interpolation and upscaling models at the same time as the rest of the pipeline and never releases them during video generation; if those models were loaded only after the initial generation finished, it might work correctly. I also noticed a custom node – I think it's called "PainterI2V" (not listed in the Custom Nodes Manager) – that crashes every second time I use it, forcing me to restart ComfyUI and regenerate. The author and some commenters mention that they run the workflow smoothly on 8GB VRAM + 16GB RAM, and I honestly don't understand how they tuned it.

First of all, thank you for this amazing workflow – it makes high‑quality, slightly higher‑resolution results with Wan2.2 possible on 8GB VRAM. I really appreciate the extreme parameter tuning and structural optimizations. I'm not nearly as skilled at debugging as you; my only VRAM management trick is setting "Lowvram" in ComfyUI's server config. According to your README and user feedback, my hardware (8GB VRAM + 16GB RAM) should be able to complete a full generation using your recommended 4:3 resolution (not the test one – sorry, I'm away from my PC right now so I don't recall the exact numbers), with two LoRAs (High+Low), and with interpolation + upscaling enabled. But at the second KSampler I hit OOM. Could anyone share your ComfyUI settings (like --lowvram, --reserve-vram, etc.) and any VRAM‑saving details (e.g., node ordering, model unloading tricks, custom node settings)? I'd be extremely grateful for any help.

I've asked AI before, but the solutions I got were all very generic and didn't really help. So I'm really hoping the author or someone who can run this smoothly could give me some specific advice or hints – I would truly appreciate it. By the way, my English isn't very good, and I'm afraid of making mistakes or misunderstandings in technical terms, so I used a translator. Thanks in advance.

(If someone running into similar issues, maybe you'll find a solution here)

Thank you very much for this workflow. After several days of tweaking, I finally got it running stably on my machine. I'd like to share some settings, adjustments, and suggestions for getting this workflow to run reliably on a PC (I'm using a laptop) with 8GB VRAM and 16GB RAM.

1. ComfyUI version & environment setup

I recommend using the portable version of ComfyUI, but don't start with the very latest release – the PyTorch and Python versions are too new, and I couldn't find compatible versions of sageattention and triton. I downloaded an older build with Torch 2.9.1 + cu130 and Python 3.13.9. First install triton (<3.6). Find a precompiled wheel for sageattention (version 2.2.0, compatible with cu130 and torch 2.9.0). Also get the triton‑windows package, download the matching version, extract it, and copy the include and libs folders into your ComfyUI environment directory, overwriting when prompted. This step is very important – otherwise the KSampler will crash. (I'm not sure if newer versions already include these libraries; my version was missing them, so I did this manually.) Note that you may need MSVC installed on your PC – I saw some triton compilation errors in the terminal. Once the environment is set up, you can update ComfyUI (or you'll have to tolerate various low‑version warnings and hints).

2. Launch arguments

I recommend adding the following arguments:

set PYTORCH_ALLOC_CONF=expandable_segments:True,max_split_size_mb:256 (this helps prevent OOM at the second KSampler)

--reserve-vram 0.5

--lowvram (required if you use the Q8 Unet)

3. Model choices

Unet: Q8 quantization works great.

CLIP: Q5KM.

For other models, follow the author's recommendations.

Loading two LoRAs (High+Low) is the upper limit – at least on my system.

4. Workflow modifications – removing "junk nodes"

In the workflow, go to the two Sage+Torch+NAG+Shift nodes and delete the SD3 sampler – this node is problematic, and the shift value is unnecessary. It affects the text condition vector and produces bad results (e.g., text written on the image + wavy noise). To compensate for the author's point about shift values reducing slow motion, you can adjust the PainterI2V parameters instead.

Also delete the get nodes hidden behind the WIDTH and HEIGHT nodes connected under PainterI2V – those two are completely redundant.

Since this node runs very fast (both TastySin and Dasiwa's Unets already have baked lightning + sageattention acceleration), I recommend using the unipc_bh2 simple sampler for KSampler; no other parameter changes are needed.

Replace the VAE decoder with a tiled VAE decoder with parameters: 512, 128, 32, 8.

5. Two‑step generation and VRAM cleaning

As the author suggested, first generate a test‑resolution video. Then on the second run, switch to 896×672 (or another resolution) – this way ComfyUI doesn't need to reload CLIP, LoRAs, etc., making the process more stable. (Generating directly at high resolution also works fine.)

For frame interpolation and upscaling, follow the author's two‑step method, otherwise OOM occurs. From my testing, you also need to add a VRAM Clean node before both interpolation and upscaling to prevent OOM.

Final thoughts

Overall, this workflow is excellent. The sageattention mechanism in particular enables fast and stable generation of high‑resolution, low‑quantization (i.e., high‑quality) videos on 8GB VRAM – something I never thought possible given my limited knowledge.

That said, some of the steps above might be redundant – my debugging process was a bit too convoluted (laughs). So take this only as a reference (but if you're running into similar issues, maybe you'll find a solution here). The general approach is:

Install dependencies according to your versions → If KSampler crashes, manually copy the missing environment libraries → Adjust launch arguments → Tweak the workflow moderately.

Thanks again!