Software Engineering Courses (SECourses)•10mo ago

@Dr. Furkan Gözükara - you see skip-layer guidance has been implement for WAN in Comfy? Thoughts? h

@Dr. Furkan Gözükara - you see skip-layer guidance has been implement for WAN in Comfy? Thoughts?

https://github.com/kijai/ComfyUI-KJNodes/commit/7c488a16ef420acf0276a4f8e31fc024a969d24b

GitHub

Add SkipLayerGuidanceWanVideo · kijai/ComfyUI-KJNodes@7c488a1

TTomfoolery @Dr. Furkan Gözükara - you see skip-layer guidance has been implement for WAN in...

Furkan Gözükara SECourses•3/16/25, 10:41 PM

some placebo effect

Furkan Gözükara SECourses•3/16/25, 10:41 PM

0 explanation :d

TomfooleryOP•3/16/25, 10:42 PM

I know it was pretty well tested in Hunyuan video, that skipping layers 9 and 10 resulted in 'smoother' videos. Curious to see what the end results are for Wan

xlads•3/16/25, 10:47 PM

xlads•3/16/25, 10:48 PM

Going to try a 30 sec image to video 720p tomorrow.

Xxlads Click to see attachment

Furkan Gözükara SECourses•3/16/25, 10:59 PM

nice

Xxlads Going to try a 30 sec image to video 720p tomorrow.

JaviEM•3/16/25, 11:03 PM

you need a lot of VRAM. Today I tried to make a 7,5s video and ran out of VRAM

FFurkan Gözükara SECourses some placebo effect

OmegaBrain•3/16/25, 11:41 PM

Calling it placebo is dismissive when there’s a clear rationale: SLG adjusts how noise is removed across layers, affecting detail retention and structure consistency. Many users, including myself, report improved results—I’ve seen this across hundreds of generations, even with SD1.5—which aligns with how diffusion models handle denoising. Plus, tweaking SLG allows fine-tuning a scene that already follows a fixed seed without breaking its coherence. If it’s truly placebo, proper testing would confirm or disprove it… so why not let the results speak for themselves?

PPelolisu did you try this ? https://huggingface.co/spaces/primecai/diffusion-self-distill...

masood•3/17/25, 4:46 AM

tried it buddy, a bit too much. the ouput doesnt look like the input at all

FFurkan Gözükara SECourses you can use google gemini flash

masood•3/17/25, 4:47 AM

life saver, dr

masood•3/17/25, 4:47 AM

@Dr. Furkan Gözükara

masood•3/17/25, 4:48 AM

one more thing, suppose i have trained a person now i want that person to hold an item ( a bag for example ) do i have to train the model on that item as well? or is there any moe effective way

masood•3/17/25, 4:48 AM

and isnt training on alot of items gonna effect model weights?

�

𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮•3/17/25, 8:21 AM

@Dr. Furkan Gözükara Are you working on Wan 2.1 14b model lora training?

Mobs•3/17/25, 1:03 PM

Hello Everone, I am running into a newbie issue with the FLUX lora training. I am running a RTX6000 with 24GB of Ram on a Virtual Machine (ShadowPC) and when trying the Lora training my machine automatically switches to the Shared GPU memory even if my actual memory is using 6gb out of the 24. any indication as to what I may be doing wrong ?
I tried all the profiles in the training. Even got to the point where the 4GB version rank 9 was using 44GB total and gave me CUDA memory error.
Best try so far was the 24GB version which ran at 23GB with 2GB on shared but it would have taken me 96h to complete the training so I stopped it.

JaviEM•3/17/25, 1:13 PM

this is fast

Mobs•3/17/25, 1:38 PM

even 12GB training fails with CUDA error, I am at a loss. Anyone has any idea what I might be doing wrong ?

�𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮 @Dr. Furkan Gözükara Are you working on Wan 2.1 14b model lora training?

Furkan Gözükara SECourses•3/17/25, 2:16 PM

yes

MMobs Hello Everone, I am running into a newbie issue with the FLUX lora training. I a...

Furkan Gözükara SECourses•3/17/25, 2:16 PM

which config you have used?

JJaviEM this is fast 🔥

Furkan Gözükara SECourses•3/17/25, 2:16 PM

nice

FFurkan Gözükara SECourses yes

�

𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮•3/17/25, 2:17 PM

Are you kohya's musbi trainner or diff synth studio?

FFurkan Gözükara SECourses which config you have used?

Mobs•3/17/25, 2:17 PM

trying most of them they almost all fail with Cuda Error, the one that does not Fail is the 24GB tier 4 but it switches as well to Shared and after 35 min I am still at 0

Mobs•3/17/25, 2:20 PM

4GB config Screeshot before failure CUDA out of memory

JJaviEM this is fast 🔥

xlads•3/17/25, 2:46 PM

Is this for wan 720p image to video? @JaviEM

Xxlads Is this for wan 720p image to video? @JaviEM

JaviEM•3/17/25, 2:46 PM

480p img to video. 720p doesn't work with 16GB VRAM

xlads•3/17/25, 2:53 PM

@Dr. Furkan Gözükara i remember you had video where you show how to create presets inside of swarmui, which one the videos was it?

Xxlads @Dr. Furkan Gözükara i remember you had video where you show how to create prese...

Mobs•3/17/25, 3:01 PM

most likely in this one https://www.youtube.com/watch?v=HKX8_F1Er_w

YouTubeSECourses

Zero to Hero Stable Diffusion 3 Tutorial with Amazing SwarmUI SD We...

Do not skip any part of this tutorial to master how to use Stable Diffusion 3 (SD3) with the most advanced generative AI open source APP SwarmUI. Automatic1111 SD Web UI or Fooocus are not supporting the #SD3 yet. Therefore, I am starting to make tutorials for SwarmUI as well. #StableSwarmUI is officially developed by the StabilityAI and your m...

TOPUI•3/17/25, 3:18 PM

@Dr. Furkan Gözükara Which is the best image upscaling and enlargement currently available?

TTOPUI @Dr. Furkan Gözükara Which is the best image upscaling and enlargement currently...

xlads•3/17/25, 3:19 PM

supir

TTOPUI @Dr. Furkan Gözükara Which is the best image upscaling and enlargement currently...

cyberbol•3/17/25, 3:59 PM

or check
https://huggingface.co/spaces/prs-eth/thera

Thera Arbitrary-Scale Super-Resolution - a Hugging Face Space by pr...

Gasby•3/17/25, 4:30 PM

Question about WAN 2.1. I have a 5090 32GB VRAM and 64GB RAM.

For the I2V, should I be using 720p or 420? Does it depend on the resolution of the image? What about the clip model precision?

Those who use WAN right now, if you had my setup, how would you be generating your images? Right now, with the 720p I2V, 45 frames at 16FPS takes around 50 seconds with sage attention and teacache

xlads•3/17/25, 4:38 PM

Question are the new wan update automatically installed when i run "cd /workspace/Wan2.1
export HF_HOME="/workspace"
export PYTHONWARNINGS="ignore"
export HF_HUB_ENABLE_HF_TRANSFER=1
source venv/bin/activate
python App.py --share" inside of runpod or do i have to download new installers then transfer the models?

Ccyberbol or check https://huggingface.co/spaces/prs-eth/thera

Pelolisu•3/17/25, 5:23 PM

this is not so good to upscale but did good image restore im my tests, ive got better result in my upscaler with controlnet tile

PPelolisu this is not so good to upscale but did good image restore im my tests, ive got b...

smoke•3/17/25, 5:30 PM

I’ve heard good things about tile controlnet. Does upscaling with it preserve resemblance? If I use it with a lora for example.

Ssmoke I’ve heard good things about tile controlnet. Does upscaling with it preserve re...

Pelolisu•3/17/25, 5:33 PM

controlnet needs a input image to generate tile, i think first you need generate your image with lora after you reintroduce it with input image for controltile.

Pelolisu•3/17/25, 5:33 PM

to upscale in denoising process you need one upscaler in latent space

Pelolisu•3/17/25, 5:34 PM

like that from Flux in swarm ui

Pelolisu•3/17/25, 5:36 PM

This is a comparison between Thera and ControlnetTile, I didn't have time to put them in the same proportion but you can check the details between one and the other, both are 4k. https://imgsli.com/MzYwMDcy