wanx-troopers

SVI

SVI is a set of LoRas for Wan 2.1 480p I2V model

The intent is to generate videos longer than 81 frames with smooth transitions using I2V. Degradation artifacts such as exposure and contrast increasing with each 81 frames step as well as loss of character consistency are rampant.

Update: there has been a claim that SVI-film lora is a placebo and simply using base I2V model with embeds prepared in the same way produces exactly the same effect. Countered: “The degradation with svi-film is much much better than Vace or anything else when used how it was designed (with Wan 2.1 I2V without distill loras)”.

i2v model could even work with multiple frame inputs until recently, but it clearly isn’t designed for that because of how fast it degrades; Vace would keep consistency longer

2025-Oct-24 authors uploaded slightly tweaked models to Huggingface alongside old ones. New .safetensors are supposed to be better. One of the updated model weight files is called film-opt.

Advice for developers is to generate extensions with a seed different from the one used for the original video.

SVI stands for “Stable Video Infinity”.

The Whole Point

Normal I2V doesn’t do well with more than one input image; VACE (T2V) works but not the base I2V; the “film” loras make that [5 last frames motion continuation] work, and the shot/talk/dance loras add something new - the ability to use reference image.

Indeed shot/talk/dance loras hold the (initial? ref?) image for the whole span of generation. I2V only take in image as 1st frame.

the Old Pusa Lora works with the Holocine - Holocine on High noise with 2.1 Pusa 512 with Magref as low noise model

Discussint Workflows

Distill LoRa is applied by the authors in the article. Workflow/nodes to use are very similar to Fun InP model.

its meant to be chained through samplers

shot lora with 1 frame yes, film with 5

no 2.1 lora is ever gonna work proper on 2.2 high noise, it’s just too different s this was with distill lora actually 8 steps, euler

don’t know if the unmerged lora can affect it negatively, sometimes it does on special loras like this; just in case run merged and fp16

Discussion Around Feeding Inputs To SVI

This section is to be cleaned up. Right now it doesn’t make much sense but it will in the future.

Normally the I2V conditioning is added in extra channel.

The image cond channel normally is the start image + black pixels for rest.
The mask marks the image as 1 and the black pixels as 0.

normally with a single image, it’s the image + black pixels for the [video?] length, and the mask marks the image

the different thing with SVI is that those black images are the original input image, with some of the pipelines the mask remains same - it does it like normal I2V does

the I2V mask is just binary 1 for the images you keep, 0 for the frames you gen

normal I2V padding is black; SVI-shot etc. is the original input image; and mask for those frames in both cases is black, zeros, 0

Q: need to tick fun_or_fl2v_model in WanVideo ImageToVideo Encode ?
A: only if you use end img

the ref is the original, the first frame is the previous last

Artist:

and even masking the remaining frames the pixel value for those frames matters a lot; the best values are definitely 0.5 or close to it basically grey pixels seem to work the best; In the native WanImageToVideo node the remaining padded pixel values are set to 0.5 after the reference images

in the original codes [diffusers?] it’s -1 to 1, so middle is zero; in comfy it’s 0-1 so middle is 0.5

Artist:

so when they say they are padding with zeros in their code, that’s the same as padding with 0.5 in comfy

comfyui there is an inversion of the binary mask before sending it to the model - super intuitive

SVI-Film

WanImageToVideo node happens to construct the masks and latents properly for SVI-Film, no custom nodes are necessary.

in native comfyui it’s black mask is unchanged; in native you send black masks for the overlap frames and white for the rest; in wrapper you send white masks for the overlap frames and black for the rest; but if you are just doing 5 overlap frames using svi-film you don’t need to worry about masks, just send in the 5 overlap frames as the start image and that’s it

svi-5-frames

SVI-Shot

SVI-Shot allows us to supply an additional reference image to an AI video sampling using I2V models. One way to format conditioning accordingly is to use WanEx I2VCustomEmbeds node from GH:drozbay/WanExperiments code repository.

HuMO section contains a workflow which drozbay, the node’s author, has shared for HuMO. Please refer to this workflow as a general instruction on how to use WanEx I2VCustomEmbeds node with SVI-Shot LoRa.

with svi shot you need exactly 1 black mask frame and 80 white mask frames, at least in the wrapper

the reference padding only works with the shot, talk and dance loras

SVI-Dance

Q: How does svi dance is suppose to be used btw ? A: with UniAnimate

Q: svi-dance is an addition to unianimate and not a replacement? A: it just makes the extension [unianimate] better

Practice

Rought sketch of a workflow: link.

One way to replicate padding from the article: svi-fragment.webp

Additional Comments

If we build a workflow replicating what code in Github from the authors of SVI does it will however, end up different from the image above. It will not be using temporal_mask input at all.

The following quote gives some info on what the code in their Github repository does and also provides some background on what temporal_mask was originally created for. It basically says that originally temporal_mask was created specifically for Fun InP model. It also says that temporal_mask is generated automatically if the only inputs given are start/end frame.

in example scripts, the film -pipelines use normal I2V padding with zeros; the shot -pipeline (and probably infinitetalk) use the padding with reference image (the input image); the temporal_mask in the wrapper is just custom way to do that I2V channel mask, as that’s what the Fun InP model uses, and also start/end frame even if that’s done automatically as well if only those inputs are given

even if passing multiple start frames it’s not necessary to mask those - it’s done automatically

Hint: SVI lora on 1st clip not needed, just freezes motion.

when svi film is used with fun inp, it’s vanilla fun inp workflow with svi lora? yes

shot lora allows infinite gens with no degradation, film lora allows for motion continuation of generation, so combined would allow infinite continuation

This site is open source. Improve this page.