Yes, except i2v clean high_noise. It contains no loras. See workflow from the image.

Yeah, that’s a very common trap with chaining multiple I2V/T2V models on a 12 GB card. What’s happening here is that the models aren’t truly releasing their allocations even when you drop VRAM clearing nodes, so each new stage “leaks” into memory until you’re out of headroom.

Even if you add clear VRAM or unload model nodes, ComfyUI still keeps memory tied up until the whole run finishes. That means when you chain 4 KSamplers with 4 models, all 4 eventually end up sitting in VRAM at once → way more than 12 GB can handle.

👉 On a 12 GB card, the only fix is to split the run (2 models → save → reload → next 2).

Here are some ways to make it work (or at least limp along) on 12 GB

How to Run 4 KSamplers on 12 GB VRAM

WAN2.2 models (~13.3 GB fp8) are too big to load 4 at once. You have to split the workflow.

Steps:

First pass (2 models):

Load your first 2 models → connect them to 2 KSamplers.

After the 2nd KSampler, add a Save Latent node (saves to disk).

Second pass (next 2 models):

Start a fresh workflow or disable the first 2 samplers.

Add a Load Latent node (load the saved file).

Connect it to your 3rd and 4th KSamplers.

Final step:

After the 4th KSampler, decode the latent → get your final image/video.

Why this works:
You only keep 2 models in VRAM at once. VRAM fully resets between passes, so 12 GB is enough.

🧠 Rule of Thumb

On 12 GB VRAM, three full I2V passes at high res just isn’t sustainable in a single run. The stable way is to break it up into multiple workflows and stitch results. Anything else is a gamble with hangs.

❌ Why 4 KSamplers + 4 Models Don’t Work on 12 GB

Even with Unload or Clear VRAM nodes, ComfyUI keeps memory until the run ends, so all 4 models stack up in VRAM → crash.

✅ Solutions

Split the run:

Run 2 models (2 KSamplers) → Save Latent → reload → continue with the other 2.

Simplify models (easier on 12 GB):

Use only High + Low models with 4 KSamplers.

Example:

KSampler 1 → Low model

KSampler 2 → High model

Save the output to latent.

Load the latent and continue.

KSampler 3 → Low model

KSampler 4 → High model

This way you only keep 2 models in VRAM, but still run 4 sampling passes.

❌ Why 4 KSamplers + 4 Models Don’t Work on 12 GB

Even with Unload or Clear VRAM nodes, ComfyUI keeps memory until the run ends, so all 4 models stack up in VRAM → crash.

Why Save Latents?

VRAM Reset – lets you split big workflows into smaller passes; memory is fully freed between runs.

Checkpointing – you can stop after any stage and continue later without starting over.

Consistency – the latent is a “frozen state,” so reloading it gives the exact same base to experiment from.

Flexibility – you can try different models, samplers, or strengths on the same saved latent without re-running the whole chain.

Speed – you don’t need to redo heavy steps (like denoising/first models) when testing later stages.

I forgot to add that 2 models High and Low are actually enough to make high quality videos.
Use the IV2 CLEAN HIGH model with IV2 LOW. First ksampler cfg at 4 or 5, second ksampler cfg at 1.5 or 2.

If you prefer IV2 High (which contains lightning) both ksampler's cfg should have 2 or lower.

I can update the workflow with 2 models if you want :)

Try this workflow. Load the video in comfyui :)
https://civitai.com/images/103351612

@Seeker360 Glad it worked for you, Enjoy making awesome animated videos :)

@Seeker360 blurry bad results has nothing to do with using too many high noise models as long the last ksampler is using the Low noise. I think you messed up the ksampler steps. Check if they have the proper steps.

1st ksampler start step = 0, end step = 4.
2nd ksampler start step = 4, end step = 8.
Also max total steps is important else it will get bad results.

example:

Stage 1: 2 steps.

Stage 2: 4 steps.

Stage 3: 8 steps.

Total Steps: 12.

Try giving 4 steps for IV2 CLEAN High noise, 2 steps for other high noise, and 4 steps for T2V Low noise

Stage 1: 4 steps.

Stage 2: 6 steps.

Stage 3: 10 steps.

Total Steps: 14.

@Seeker360 Can you post an short video with meta (workflow attached). I can check what you did wrong.
You can also try to load this workflow from this video: https://civitai.com/images/103345891

2 ksmaplers should work as long you use high for the 1st ksampler and low on 2nd.
If its a wan2.2 lora, then the high noise should have strength: 2.00 and low 0.85.

Both I2V high and low except the clean version have Lightning and seko at different strenght.
Lightning, Lightx2v, seko, Mps, etc do mess up loras so thats why you should always use a base model on the 1st ksampler without Lightning, Lightx2v, seko, mps, etc and CFG should be between 3 and 5.

First ksampler: start step 0, end step 4,
Second ksampler: start step 4, end step 8.
Max total steps = 8. ( 10 would be better but it online refines little bit more)

@bogysmsm129 When setting up your model loaders with LoRAs, the strength depends on whether you’re using Lightning and how many model loaders are in play:

LoRA Strength Guidelines

First model loader (with Lightning): LoRA strength = 2.00

First model loader (without Lightning / clean version): LoRA strength = 1.00

Second model loader (with Lightning): LoRA strength = 1.00

4-model loader method: LoRA strength = 0.80 (keep it lower to avoid oversaturation)

CFG (Classifier-Free Guidance) Tips

Use CFG between 4 and 6 with IV2 CLEAN HIGH→ gives more stable motion and stronger prompt adherence without overcooking the video.

Lower CFG (closer to 4) = smoother, more natural motion.

Higher CFG (closer to 6) = more detail and prompt accuracy.

@bogysmsm129 I am guessing the power lora was causing the issue. I could not find the creampie lora you were using so instead i tried this one: https://civitai.com/models/1941545/creampie-anime-inspired-vaginalanaloral-wan-22-i2v-14b

I couldn't fully test it out without the lora you were using but its working now.

I edited th steps for better motion. You dont really need to put too much details in wanimage2video. Little bit caption, and the rest tags.

Here is the workflow. Included a video with meta too.
https://limewire.com/d/lZI3M#UohPQUJi2X

@bogysmsm129 Still doesn't work? I tested and it was working tho.

@bogysmsm129 Try the last Unet loader as Ultima T2V Low.

Try with NSFW VAE too :)

@bogysmsm129 https://limewire.com/d/1C665#A0C75R3i9H

Because i created them with the Fp8 pruned base models called... Ultima I2V :)

@Tophness The full fp16 version will have a much better quality but also requires more vram.