wanx-troopers
Hidden Knowledge
2025.12.01 - VRAM Saver
Wan VAE can use way less VRAM in wrapper if you offload the cache to CPU, makes it also a lot slower, but still beats tiled VAE at 720p it’s like ~5GB less VRAM used
To similarly save RAM in native use WanVideo LatentsReScale node and plug wrapper VAE node into a native WF.
2025.11.30 Chunked RoPE - VRAM Saver When Torch Compile Off
Two of the highest peak VRAM hits on the model code are the RoPE appllication (rotary positional embeds) and ffn activation (feed forward), these happen separate from the attention, which is the 3rd highest.
The chunked option splits both the rope and the ffn calculations, doing it in chunks and thus reducing the peak VRAM usage to below the attention (which cant really be split without affecting results or slowing it down);
torch.compilealso optimises it similarly, that’s why it uses less VRAM when it works.
pytorch now has the fused RMS norm built in, that was one of the biggest benefits of compile; you could just not use compile, use the pytorch RMS norm and the chunked rope
2025.11.30 rms_norm_function - Time Saver When Torch Compile Off
WanVideo Model Loader: set rms_norm_function to pytorch not default for faster gens if torch is new enough and torch compilation is off.
No time saving if torch compilation is on.
2025.11.30 sageattn
sageattn and sageattn_compiled refer to Sage Attention 2; if you had Sage Attention 1 installed for some odd reason they’d refer to it instead.
Sage Attention 3 can be installed at the same time as Sage Attention 2, so there is a separate option for it sage_attn3.
The only difference between
sage attentionandsage attention compiledwas that the latter ALLOWED for torch compile; normally compilation is disabled for sage as it wouldn’t work anyway, and just causes unnecessary warnings and such; even with the new sage it still doesn’t actually compile it, the code just handles it graciously so there’s no graph breaks.
sageattn_compiledis sort of redundant now with latest version of the wrapper as the workaround to allow even the old version in the graph seems to work now
Difference now is that normal
sageattnis wrapped in a custom op so nothing in it is compiled (but also doesn’t cause graph break) and thesageattn_compileddoesn’t wrap it at all, but expects you to have the very latest version of sageattn that has built-in similar workaround
fullgraph feature in torch compile is a debug feture; it will error if there’s any graph break
2025.11.30
this was even more relevant for Kandinsky5 because it has the 10 second model, and it’s 24fps, so that’s 241 frames to do at once but same applies to Wan
xformers: No use for it with sageattn being so fast; been ages … back then it was maybe 5% ahead of sdpa
Unload Modulesin ComfyUI manager can potentially fix weird VRAM issues
Note cfg usage(>1.0) increases resources consumption
disable smart memory in ComfyUI - then it always offloads [does not try to keep models in VRAM to avoid unnecessary RAM - VRAM transfers]
Drozbay
drozbay is continuing to push boundaries on what can be achieved aroud Wan models. His nodes discussed on this website reside mainly in the following code repositories:
| Drozbay’s Repo | Purpose |
|---|---|
| GH:drozbay/ComfyUI-WanVaceAdvanced | Nodes for advanced Wan + VACE generation |
GH:drozbay/WanExperiments aka WanEx |
Nodes for preparing and examining embeds in new ways |
List of Drozbay’s contributions discussed on this website:
- Drozbay’s Study: nodes in
WanExto compose/examine Bindweave and advanced I2V embeds - HuMo - Drozbay: HuMo continuation workflow
- SVI-shot:
WanExcontainsWanEx I2VCustomEmbedswhich works well with SVI-shot - Drozbay’s Impact Pack SEGS Detailer For Wan Video And VACE
- Painter I2V Modular Replacment
WanExcontainsWanEx_PainterMotionAmplitudenode which replicates action ofPainter I2Vnode in a modular manner - Wan I2V T2V / VACE Tricks hoists avdvanced workflows created by Drozbay
- Wan 2.1 to Wan 2.2 LOW LoRA
2025.11.29
Self-forcing can mean models that generate video portions progressively, small frame batch by small frame batch, thus enabling longer generations. Mainstream today’s models are not SF however. Existing SF models offer lower quality of results generally.
2025.11.27
res_2s runs about 2x slower compared to Euler, the advantage is that with the same number of steps a better result is delivered; 2x steps on euler is a fair comparison for res_2s; the hope behind using res_2s is to get more mileage out of the same .safetensors file
The Mystery of Denoise Control Knob
Denoise on sampler is a widely used but often misunderstood parameter.
In WanVideo Sampler e.g. in wrapper workflows
its function is simply to override start_step - it doesn’t do anything else.
It is less confusing to always set start_step and never touch denoise.
In native Sampler the code is more nuanced. One bit of advice we have is
denoise_strengthand steps are mutually exclusive and changing both at the same time leads to weirdness, especially with the usual schedulers and small step values.
2025.11.24
Full non-accelerated generation of AI videos delivers the greatest quality but takes a long time. Faster generation is achieved by adjusting the model weights. The process is known as “distillation” producing a “distilled” model. In most cases the difference between distilled and non-distilled model can be extracted as a LoRa. Distillation can proceed along two separate axes: cfg and steps. A model can be distilled in on both at the same time.
cfg distilled makes it work at cfg 1.0, meaning cfg is disabled and only the positive prompt model pass is done, meaning 2x faster inference
step distillation just makes it work with less steps
2025.11.21
Wan latent space is compressed by factor of 8 spatially, 4 temporarily
2025.11.20
res_multistep in Wan Video Sampler (wrapper) is probably similar to res_2s in other samplers (Clownshark?)
most people using res_2s and similar are not comparing it to other samplers at 2x the steps; not denying it can be better, just skeptical how much, at least with distill loras
2025.11.15
umt-xxl-enc-bf16.safetensors only works with the WanVideoWrapper;
native workflows need the model without the enc
like umt5-xxl-fp16.safetensors.
2025.11.14
noise mesh/grids/dithering problem with Wan models: originally believed to be
wan video is only usable with qwen image inputs because those are so soft that they do not cross into the detail level where the vae will fail into noise grids
but it looks like the actual explanation is
It’s triggered by the mismatch between real (encoded) and generated latents. Generated latents are usually blurry, although adversarial distillation like lightx (dmd2) helps sharpen the generated latents. The decoder doesn’t know how to handle blurry latents, they’re out of distribution, so it generates speckle/grid artifacts as its failure mode.
The generated latents are blurry. More steps, use lightx loras, disable sage attn if relevant. 2.2 low noise + lightning or lightx2v is pretty good at minimizing the artifacts
another idea
Doing a second pass with the T2V LOW model with a low denoise (around 0.2) can fix that
2025.11.11
When switching on tiling in VAE to save VRAM working with WAN models it is advisable to set temporal_size setting on VAE to more than the actual number of frames generated.
For example when generating 81 frames a setting of 84 is advised. Otherwise tiled VAE may introduce flickering.
Massively useful list of WAN-realted models, LoRa-s, controlnets: link
2025.11.10
Wan 2.2 vae is for the Wan 5b not the rest though
For wrapper implementations the “null” image is black. For native implementations the “null” image is gray. Also the masks are inverted.
2025.10.25
WanVideo Set Block Swap and similar nodes do not require re-loading .safetensors to make changes - that is why they’re separate nodes.
Earlier
By default safetensors files are loaded via mmap.
Kijai’s nodes can convert between data types such as bf16 at safetensors loading time.
load_device = main_device in WanVideo Model Loader means “keep model in VRAM”.
| Kijai’s Repo | Purpose |
|---|---|
| kijai/ComfyUI-WanVideoWrapper | Alternative to Native KSampler, “wrapper” nodes |
| kijai/ComfyUI-KJNodes | Supplementary, can be used with Native KSampler |
| kijai/ComfyUI-WanAnimatePreprocess | Nodes for Kijai’s WanAnimate workflow |
| kijai/ComfyUI-MMAudio | Foley? |
| kijai/ComfyUI-GIMM-VFI | frame interpolator |
Resolutions to try with WAN: 1536x864, 1280x768, 1200x960, 1/2 of that, 832x480, 1024x576, 1440x816
Resolutions to try with Ovi v1.1: 1280x704, 704x1280
Colored inputs on ComfyUI node designate non-optional inptus, rather than inputs that have something connected to them.
Q: why using block swap is hardly any slower than not using it?
Kijai: because it’s async; it swaps the next block while current one is processing; so if your step takes longer than moving the block, you don’t really notice; it’s the prefetch option in the node and non-blocking; even without that on fast systems it’s not a huge speed hit tbh
Too high CFG on low steps can cause “burns” and flashes.
Resize Image v2 once reported that a 150 frame 480x832 video was taking slightly bellow 0.7Gb.
There is only one CLIP that works with a particular class of model like Wan. It can be in different precision however.
ViT means “vision transformer”. h = huge, l = large.
RoPE
When using going beyond trained model resolution with a wrapper workflow can try using WanVideo RoPE Function.
Particularly with Cinescale LoRA can try setting this node to comfy and f/h/w=1/20/20 or 1/25/25; Kijai:
when you use cinescale.. you mean just the lora? or the rope scaling too?
really large resolutions should benefit from the rope scaling
well technically anything that goes past the trained resolutions should
I don’t really remember the values but in the original they used something like that
basically it’s scaling the spatial rope in attempt to make larger resolutions work without the repeat effect
possibly the lora is trained with the scale of 20
Defaults are 1/1/1.
Skimmed CFG
Skimmed CFG from Extraltodeus/Skimmed_CFG. High CFG in the sampler helps to follow the prompt, skimming prevents the burn issue.