This guide covers some intermediate aspects of video generation based on knowledge (not copy-pasted IA text bllshift) with the intention to help you understand what are you doing when using WAN2.2, SVIPro, Lightx2v all together.

Introduction

If you need basic guidance about the concepts the basic guide is HERE. The basic guide is meant to be used side by side with your "template workflows" for video generation: includes an explanation of the nodes and why are linked that way. Concepts. I encourage you to know what your basic nodes do: will change your prespective.

This intermediate guide covers some aspects and general problems when you add more nodes and features to your basic template because you want to increase the video duration and reduce the processing time. These are which are SVIPro (SVI) and Lightx2v respectively.

Why these loras are important is HERE including links to SVIPro default templates, Lightx2v templates comes with comfyUI.

This guide is meant to let you add some nodes on default comfyUI templates and play with them.

Although I try to follow an order, there isn't. You can select one aspect of this guide and try to implemented it, then another, etc.

From this point on I asume you've read the basic guide of concepts.

Lightx2v: Practical Recomendations

Lightx2v as a LoRA

Lightx2v can be used as the whole model or as a LoRA. Notice that as a whole model, the model contains the motion, and thus you have some limitations in prompting, there are tricks though.

Using it as a LoRA has some advantages the main ones:

• Select you base checkpoint/model

• modulate when lightx2v works

These two points are covered in the Triple Sampler chapter of this guide.

Regarding the use of the loras, there is a parallel guide HERE. It covers which combinations of light loras works and how can you "preview" your long videos before fully compute them.

Did you know lightx2v 1030 High does not have the ability to listen to camera or viewer movements and relies on the Low lora?

That explain why your videos remained static or instantly exploded in motion.

Triple Sampler

Wan 2.2 requires two samplers one for each expert: High and Low noise. It is unlikely that you're doing long WAN videos, or wan videos in general without lightx2v, either have very low quality or very high computing time. Or you have server-grade resources... Steping down from 20-40 steps to 10-15 have its caveats. lightx2v will drastically reduce the number of steps, but also your possibilities to guide and prompt adhere since lightx2v "explodes" when CFG is >1.

There are several solutions, one of them is adding additional samplers. For Wan 2.2 add an aditional HIGH sampler and split the sigmas of HIGH between the two samplers. All samplers will have the same model and loras except for the first one which won't have the lightx2v lora.

You have 3 samplers: HIGH_NOLIGHT, HIGH, and LOW.

HIGH_NOLIGHT with 1-3 steps, where, withouth lightv2x, you're free to increase the CFG.

HIGH, which is your regular HIGH sampler with the remaining high samples. CFG = 1.

LOW, which is your regular LOW sampler. CFG = 1

Then you can use a 6+6 scheme where: 2 are for HIGH without light, 4 for HIGH with light and 6 for low.

So by complicating things you have part of both worlds: control CFG in the early steps, and the power to reduce steps by lightx2v.

Three samplers: Note how a new lora is added in the second sampler containing all loras plus Lightx2v, giving space on the first sampler to increase CFG at your will. Beware of high CFG. Sigma Split splits the high Sigmas between the two samplers. The non lightx2v must go first. This is the exact template for SVIPro with 3 additional elements: the sampler boxes, a load lora box and a split sigma, it takes 10 minutes to implement and test.

You're free to split low as well and increase CFG as well.

This configuration allows to play with 1217 light2xv loras which are even worse to control than 1030.

Use different weights and sets of loras for each video segment on SVIPro but do not deactivate them unless the item/action prompt by that lore is no more required, otherwise will vanish.

Remember you have to have sigmas pre-calculated and split them accordingly for each sampler, there is a great node I recommende for that: WanMoScheduler

You can use: VisualizeSigmasKJ node and Preview Image to see how sigmas are split.

You can use: SplitSigmas and step to split your sigmas in all the samplers you consider.

Sidenote on Negative Prompt for WAN

The whole point of CFG>1 is to evaluate force the positive prompt and evaluate the negative. Remember in WAN (and most video encoders) negative prompt to avoid stuff. So if you say "no trees" you don't want no trees. Its very English for that thus will fill it with trees. Also, for WAN2.2 the negative prompt is for concepts, clip-like, contrary to the positive prompt which is phrase-like.

Slightly Advanced Guidance

As you can see in the previous figure all samplers are have Scheduled CFG. Don't ask AI: don't know how they work, it seems most AI models can't read documentation.

Lets see some examples of how the system works based on two parameters on that node: start_percent and end_percent.

I will refer to High expert as H and Low expert as L. so 3H+3L: The usual scenario you know is 3+3, i.e. 3 steps in High and 3 steps in Low.

Remember CFG>1 implies double the time, since both Positive and Negative prompt is evaluated. With CFG=1 negative prompt is not evaluated. That implies CFG>1 takes roughtly double the time to compute per step.

The 3H+3L combination is what you'll find in most guides and workflows with the note: "use 4 steps for better quality", which means simply increase from 3 to 4. This is case A) for 3H/L, and you can imagine how it looks for 4H/L steps.

Lets add the additional sampler at the begining. With this additional sampler you can force guidance, and set CFG at 3 (or whatever number you want). But remember you have to split sigmas, you still have 3 steps for H. One step, the first one, will be for non light lora: where you can freely push the CFG while the remaining two will be for light lora. Thats case B).

If you watch my content I usually say if its 6+6 or 6+2+6 etc. This is the usual fast run for quality and prompt adhesion: 6H 6L. if there is a 2 in the middle its 6H = 2nonlight + 4light, +6L. This is case C).

Here it comes the use of the start_percent and end_percent parameters. D). Until now we left those values at 0 and 1. this is a per unit setting: its stating that from 0 to 1 (0 to 100%) the CFG applied will be the one in the box, outside the specificed range the CFG will be 1. So if you leave it 0-1 it will always apply the CFG on the box. In case D you actually have 3 steps to work with, this means those percentages must be split in thirds. 0-0.33 then 0.34-0.66 then 0.67-1. So if you set: start_percent = 0.34 end_percent 1, for 3 steps, you're saying: use CFG=1 for first step, then use the set CFG for the 2 remaining steps.

Imagine you set 0.01-0.99. you'll need 100 steps to appreciate the fist and the last step are CFG=1 while the other 98 are CFG as set.

Imagine you set 0.1-0.9 and you have 3 steps, all steps will be computed as CFG as set.

Look at the green arrow, for three steps those are the points where the Scheduled CFG Guider checks if the step is within the 0-1 range. Example D) is 0.34-1. So first step is CFG = 1, and the other two are CFG = 3.

Four examples of steps combination based on A) 2 samplers B), C) and D) 3 samplers.

The "creativity" (I would say vorticity or chaos) comes at lower CFG. so it is up to you to decide to enforce the prompt on the very first step, on the 3 steps or maybe on the middle step. Remember a loose CFG allows for better continuity, seamless movement or less degeneration at the cost of less prompt following.

If you understand this you also understan why it has no point to use start 0 end 0.5 on D. The motion has been heavely forced on the fist two steps and you still have a step before swaping to high light sampler but that step is at CFG==1, no push, no negative prompt check. I recommend not to do it, but may help to calm down crazy loras which if the CFG was quite high can "explode" when light kicks in.

[I'll post some examples soon :D]

Lets see the sigmas involved, using simple scheduler.

First the sigmas without splitting:

For 3H and 3L.

For 7H and 7L

Case B

Case C

Case D

Slow Motion

There are severa reasons you may experience slow motion effect when doing long videos with SVI and LightX2V here are some reasons I know produce these effect:

Too much/little lora strength

Qualitative speaking when you mix loras their strenght are ponderated before use. If you use a movement lora and its not working other loras my interfeer. Isolate the lora and prompt text. The prompt may be guiding two oposite loras. If its isolated at doesnt work may be the strenght, if strenght is high and doesnt work may be the prompt. This process methodically used allows you yo ponderate the weights of loras. Maybe you haven't noticed this until you added two 1.0 strengh loras to your workflow: light and svipro.

Too much lightx2v strength

As commented in other guides let me remind you that too much lightx2v may leed to slow motion. There is a delicate balance between the strength of that lora, the spatial size of a video and the motion. Once decided a resolution, try to lower it in 0.1 increments. Prepare a clip with your loras and estimated strenght, fix the seed, and test from 0.9 to 1.5 if you have no experience with that combination. Remember, if I2V there is no MOTION on the first frame, so the video needs to "kick" the motion: don't experiment with short (<50 frames) for this matter.

Ideally do it with 2 clips (50+50 frames f.e.), that way the second clip will have the motion of the first one and you'll be able to evaluate the motion lightx2v injects. The first second (16 frames) of the first clip is lost motion unless the image contains all the elements in motion of a lora of a specific motion. A wheel spining on the image will start already spining, A still wheel will require frames to accelerate even if you say "the wheel is spining fast" on the first phrase of the prompt.

If using 3x Sampler: Low CFG

If you apply a triple sampler non-light steps will produce very few push if CFG = 1, since the model is meant to work with around 20 total steps. You're also losing the ability to evaluate the negative prompt. Check the processing time to see if your CFG>1 is being applied (should doble the time of light steps). If you play the league of 4+3 or 3+3 raise the CFG and raise it more as bigger the canvas is.

Short prompt

A short prompt on WAN may lead to several effects, specially when passed the 480 boundary. If your characters "dance", "jitter", "walk in small steps", or "doubt" probaby your prompt is very short and WAN doesnt know what to do with the time he has. You may:

• reduce the lightx2v lora strength: this will force the model to do slower motions, which wont be slow-motion. F.e. turn around will turn your character slower.

• reduce the frames of the segment: if you dont need the time don't compute it.

• increase the detail in the motion: add some details on he facial expression, arm movement, background change.

Look for the following points you may be experiencing one of them.

[BUG] in LoadLora nodes its CORRECTION VERY IMPORTANT!

In the image above you can see how LoadLightx2v HIGH LORA is applied in the 2nd of 3 samplers.

HIGH no light, HIGH light, and LOW light. The load lora box of comfyUI at least until release 0.28 (2026-07) contains a bug that makes loras not apply alphas. This means light lora is not properly applied, which leads to: slow motion, no prompt following, etc. There are two solutions:

a) Apply at least one lora at 0.1 strengh 1.0 clip to all segments.

b) Use power lora loader node (from rg-tree) beware because this node wont allow you to "text" select neither the lora or its strength.

Many batch lora loaders may not load WAN loras correcty generating the same effect.

FPS Awarenes on lightx2v 16fps and 24 fps [important!]

Lightx2v for wan 2.2 in the 1030-1022 combo or 1022-1022 combo or the 26-26 combo (H-L) have a particular trick. When working on 480p mode they seem to work at 16fps but when resolution is increased to 720 they seem to work at 24fps some loras are trained at 24 fps and may mandate on the clock of events.

This is a curious effect, try to prompt by seconds and youll notice some actions dont fit in time when you increase the resolution. Its not about steps, strenght or anything but time seems to tick slower on higher resolutions. Just save the video at 24fps and you'll see how the motion match!

SVI/SVIPro: Practical Recommendations

I asume you have red the SVI point of this guide. SVI has two features:

1) it uses the initial frame of WAN as an anchor point, so for the model if "you dont know" then "look at that reference frame (latent information)" lets say. That's whats creates the character consistency. WAN 2.5 and further versions, is character consistent with more than one character among many improvements. Lets limit the discussion to WAN 2.2.

2) it can translate motion vectors to the next clip. That is why SVI pro videos are seamless. This is done by the motion latent count (counted in latent frames (i.e. lframe = 4 frames). "it can" it means "it can't" as well if you require an abrupt cut: 0 motion frames allow for seam complete abrupt motion cut, a scene cut f.e. 1 is for high motion continuity, the regular value, 2 translates more information but also its more rigid, its also good for valleys of motion (the "stop point" from going forward to going backward).

To do so you overlap lframes which will give the motion continuity (and also port noise and errors and slop to the next clip!). Look at the figure. You can select how many motion bring to the next clip by the number of overlaping lframes: 0, 1, 2 etc. In red its an example of 1 lframe overlap. in green an example of 2 frame overlap with the final result. When overlaping you loose the number of lframes you overlaped, which are the motion latent frames. Both motion latent count and overlap frames must be coherent.

Here is an example of how to control it dynamically:

the Iff() function checks the index of the segment (a), you can nest iff. If the index (a) its the desired value, apply a given motion latent count (1) else apply 1 as well. (a=1, apply 1, else 1). This is linked to the overlap frames which must be 1 (the anchor) plus n latent frames: 4*motion latent count.

Another example iif(a==3, 2, iif(a==5,2,1) --> if segment is 3, motion = 2, else if segment is 5 motion = 2 else motion = 1.

*If you dig deeper it is true that the latent can be processed (and in some workflows it is) in chunks. But that doesnt change the fact that the state is formed equally in the spatial and the time dimensions of the video.

Notes

Personal notes

• I started with ComfyUI and WAN on last week of december 2025, this guide has been published 3 months later (end of march 2026) after about 2000 generated videos in local hardware with the implied limitations.

• I've lots to learn but I haven't found, even with the help of IA, a comprehensive way to understand what was this video generation stuff is about, that is why I wrote this guide. Any corrections will be appreciated.

• The reason why the denoised stuff I publish does not contain the "recipe" is because I've found reverse engineering other's workflows lets to:

  • hundred of hours lost on dead ends

  • installation of lots of missing components which end up breaking your comfyUI installation

  • lots of specific user-corrections without any practical sense but to correct mistakes on the workflow

  • workflow monsters which achieve 96% of what templates achieve with small modifications.

• I decided to split the guide because the cens.. moderation system delays its publication several days every time I update it. Also the other article was becoming very long.

• WAN may be obsolete, since the models are released by political means, i.e. if occident releases a new powerful paid model someone else counterweights. Good for us. The conceps beneath these guide probably are not. Guidance, samplers, denoise, concatenation, sigmas are used across all models and work similarly for all image generation.

In time these guides will become a meme, an antiquity, reminiscent of a period where hardware was kidnapped by the big tech and you had to squezee every pirric resource you had available. This can be next year or in a couple of years.

They don't seem to understand two facts: technology will advance and they don't own it, they own some chips and some current models. Another person will fabricate more chips, another person will develop new models and all their business will be gone, because theres nothing so special a gigantic model can do that regular person needs.