From my personal experience the quality difference with Comfy-Org FP8 isn't huge at first glance, but the subtle improvements in details grew on me over time, now I clearly notice them when using Kijai quantizations. It's hard to explain exactly, but those small enhancements make a real difference and makes you go back to it. For NSFW videos, I used SmoothMix High noise with one LightX2V set at 0.23 strength to eliminate bounce issues. Use ComfyUI default workflow or WanVideoWrapper in the Zip file: I2V | T2V
In practice: can use a bit more VRAM at load time in some workflows, but once running it’s often competitive or slightly faster than the default Comfy‑Org workflow. Quality is extremely close to FP16, especially in subtle detail, texture and motion.
“better fp8 scaled models (when measured against fp16)” with a quantization method inspired by Tencent’s HunyuanVideo FP8 code.
large linear weights in FP8, many small tensors (biases, norms, embeddings, head, modulation) kept in FP32 for extra stability.
Designed and tested together with the ComfyUI‑WanVideoWrapper, but also works with native WanVideo nodes.
Comfy-Org or KJ, Which one should you use?
Use Kijai FP8 if:
You care about squeezing every bit of quality out of the FP8 and value subtle detail, texture and motion coherence across many renders.
You’re using ComfyUI‑WanVideoWrapper or don’t mind experimenting a bit with workflows.
You can tolerate slightly higher VRAM use during load in exchange for potential quality/speed wins in actual generation.
Use Comfy‑Org FP8 if:
You want the “official” default that matches ComfyUI Wan2.2 tutorials and examples.
You prefer a stable, well‑documented setup that “just works” with native WanVideo nodes.
You mainly care about solid quality + VRAM savings, without chasing the last 1-2% of detail.
Think of Comfy‑Org FP8 as a reliable baseline, and Kijai FP8 as an enthusiast‑tuned variant with extra precision in key places and a slight edge in subtle detail,
FAQ
Comments (21)
Hi, could you please tell me the values for CFG and Steps for High and Low?
Using LightX2V or without? With LightX2V: CFG 1, Steps 4, sampler Euler, shift: 5.
Without LightX2V: CFG 3.5, steps 20 or 25. You are not bound to any specific sampler, unlike with LightX2V, use whatever works best for you.
@sweetmax797 ok,ty
Hi, I experimented with your checkpoint today and noticed that videos with the same prompts but different seeds look extremely similar. For simplicity, I used your workflow with the FP16 text encoder. (T2V)
Hi, yeah that's normal! seeds like 0 or 222222 aren't special at all, just random starting noise.
If you keep the prompt exactly the same, changing the seed only gives subtle differences, nothing crazy. T5 (FP16 just understand more complex sentences) turns the text into fixed math stuff that guides everything, so different noise mostly ends up looking pretty similar.
Play with the shift!
Wow, thank you so much, I've learned something new again! I tried it right away and it actually worked :-)
Are these the same checkpoints obtainable from Kijai's HF or are these a new/different release?
exact copy
@sweetmax797 gotcha. Thanks for the smoothmix with lightx2v at 0.23 tip. I was having issues using smoothmix having it ruin the appearance of my characters or just doing wacky shit I didn't want. Lowering the high lora like you suggested has been a game changer.
@s373ns you're welcome, SmoothMix one of the best. UniPC/Normal 6 steps and most of the loras at 0.65 works best with it.
@sweetmax797 Sorry, I'm just wanting to make sure I get it...are you saying this is what you usually do?
HIGH LORAS:
LIGHTX2V 0.23
SMOOTHMIX 1.0
LOW LORAS:
SMOOTHMIX 1.0
@SinglePage
- base_model High [SmoothMix i2v H] -> Lora_whatever 0.65 -> Lightx2v 0.20
- base_model Low [KJ i2v] or any other low model you like -> Lora_whatever 0.65 -> Lightx2v 0.2 (if u r using untouched base model 1.0)
SmoothMix and most other base models shared here already have LightX2V merged.
You can use FP16 (high) or FP8 (low) -> higher precision is always better, lower-precision models will work with whatever signal they receive.
Regardless of the LoRA or base model, if the text encoder is FP8, prompt is limited. If the resolution is unsupported, the video will break or drift from the prompt no matter what you do.
WAN2.2 prefers resolutions under 1 megapixel, which is where it gives the best balance of motion, detail, stability.
720×1280 is supported
640×960 || 576×896 is where WAN2.2 feels alive
Portrait video sweet spot :
512×768 -> light, stable, but calmer motion
576×896 -> better motion without breaking consistency
640×960 -> magic zone
Upper safe zone (still good if VRAM allows):
704×1024
768×1024
Beyond 1 MP:
motion weakens
model will break
Tip: It's hard to memorize all those sizes. whatever size you choose, see if it is divisible by 8 like 512x848 : 512/8 = 64. 64/8 = 8. 8/8 = 0
@sweetmax797 OOOOH I've been using the Smoothmix loras, and thought you meant that. I never considered mixing low and high models! Thanks so much for all the details, this helps a lot!
@SinglePage Doesn't matter I think they're merged with the base model ( probably an older version, the merged one performs better, personal experience.) so even u use those loras, 0.8 is to high for them, at 0.8 water looks creamy, bounces are everywhere. if u r using a pure base model then 1.23 ( i did 40 videos between 1.2 to 1.33, majority of videos i liked, had 0.23 extra strength) see this workflow https://pastebin.com/TyGhSgcH
@sweetmax797 Thanks for tips and workflow. Much appreciated
@sweetmax797's account deleted? What happened?
@null Actually, 8/8=1. Just saying.
Is there a special workflow needed? I loaded these in to one that can switch between gruff and safetensors but I just get a blurry video with very little movement.
workflows are in the zip files and even added the links in the description text else you can use the ComfyUI default workflow for WAN2.2. regarding blurry videos: if CFG is 1 and steps are 4 or 6, then you need to use LightX2V LoRA, Sampler: Euler/Simple Shift:5, if you're not using LightX2V LoRA, then CFG 3.5 steps 20 or 25 as a start.
@sweetmax797 Thanks, I got it working with the Loras.
Any chance of getting the GGUF versions of these?