Which is your primary gpu connected to your display?
Which is your primary gpu connected to your display?
TOnboard Intel. Both headless. 3090 shows the full 24.0
T
TSo thats why i was curious why the total was different.
Fsame here 31.5
Fthey stolen vram
TYeah something different between 30xx/50xx. My 3090 = 24.0 in task manager, 5090 = 31.5 in task manager - and I can't load the full FP16 WAN 2.1 model in 32GB because its 500MB short!! it goes to shared GPU which slows it down a lot.
Fshameless nvidia
Fbut flux training fully fits
T@Furkan Gözükara SECourses have you found perfect teacache settings for Wan2.1? with no visible quality loss to me: teacache at 0.250 T2V 14B model, with Sage and torch-acceleration i got a 49 frame Wan down to 101 seconds. I can get it faster, but start to notice degradation.
F0.15 is nice
Fbut 0.2 made some degrade
Fbut if working for you great
TWill you be doing wan/hunyuan lora training tests? I got hunyuan down well, but wan lora's are not coming out well for me with musubi-tuner, but can't tell if its 5090 problem. So much instability with cuda 12.8, sage, triton 3.3
Fi will first do wan 2.1 lora
TI think this number changes if you select 14B co-efficients. this is also a bit confusing. I also notice torchcompile kills loras
F0.15 only changes with 1.3b model
Tso you like 0.15 for I2V and T2V 14B right? I didn't buy 5090 to run 1.3B models 
Twill you use musubi-tuner? I just can't get good lora's for wan with the same dataset as used for good loras on hunyuan. I actually think captions might be required this time for more flexibilty during prompting. But not sure yet. I didn't caption hunyuan and they came out good, sometimes better than flux with kohya.
Fytes 0.15 for all 14b model
Fyes i plan to use musubi
Flots of people using lots of bug fixes
TMaybe I need to try again with later code. Also this guy just created a gui for it. https://github.com/Kvento/musubi-tuner-wan-gui there is also a gui from ttplanet https://github.com/TTPlanetPig/Gui_for_musubi-tuner - this is gradio
Fi plan to make my own gradio
TI just modified TTplanet's one. its a good starting point. Let me know how your Wan Lora success is. I'm going to keep experimenting, and will let you know what i discover too. I think there is a reason character lora's aren't being released as fast as "motion/animation" loras
Fsure. i am testing 0.2 teacache right now
Fit takes less than 10 min on h100 highest quality - 20 steps
TAlso TorchCompileModelWan (for wan21) breaks Lora's on 50xx but not 30xx. Must be the 3.3 triton, or torch nightlies. I just tested. Same workflow works fine on 30xx. but not 50xx with Torchcompile running on both.
Fvery likely
Fwow teacache 0.2 not bad
Fand super fast
TIts hard to find people with direct 5090 experience still. Shortages are tough.
TI was using 0.16 then 0.2 - but i was reading another thread where they reported 0.25 being the perfect spot
TNot my research - direct from Kijai: Benchmarks. Everything here is Wan2.1 720 @ 49 frames, sageattention, cuda 12.6, 3080 ti, listing thresh, starts, times, and quality observations
0.300 (rel_l1_thresh) a lot more subtle grainy noise than 0.250
0.00 (start) -> 13:08 (encoding time)
0.2875
0.00 -> 13:39 more grainy noise
0.280
0.00 -> 13:18 92% starts noisy
0.275
0.00 -> 14:55, near indistinguishable from 0.25
0.270
0.00 -> 14:41, 96%, nearly perfect
0.265
0.00 -> 16:09, 96%, near perfect
0.250 -> looks lossless
0.10 -> 16:30
2x run:
Best so far, but slower,
0.250
0.10 (start) -> 16:20
Near if not lossless at 0.250, at the cost of about 3 minutes when compared to 0.300.
in summary
0.300 ~ 13:00 I found a bit too grainy for my liking, though speed is nice
0.275 ~ 14:55 great balance between speed and quality
0.250 ~ 16:30 I found was basically lossless
0.300 (rel_l1_thresh) a lot more subtle grainy noise than 0.250
0.00 (start) -> 13:08 (encoding time)
0.2875
0.00 -> 13:39 more grainy noise
0.280
0.00 -> 13:18 92% starts noisy
0.275
0.00 -> 14:55, near indistinguishable from 0.25
0.270
0.00 -> 14:41, 96%, nearly perfect
0.265
0.00 -> 16:09, 96%, near perfect
0.250 -> looks lossless
0.10 -> 16:30
2x run:
Best so far, but slower,
0.250
0.10 (start) -> 16:20
Near if not lossless at 0.250, at the cost of about 3 minutes when compared to 0.300.
in summary
0.300 ~ 13:00 I found a bit too grainy for my liking, though speed is nice
0.275 ~ 14:55 great balance between speed and quality
0.250 ~ 16:30 I found was basically lossless
WThat is odd...
WHello everyone
WCan somone hel me to solve this?
WIm trying to merge two models in SD
Wbut im getting this error
W
WIts a VRam problem?
Astable-virtual-camera https://stability.ai/news/introducing-stable-virtual-camera-multi-view-video-generation-with-3d-camera-control?utm_source=x&utm_medium=social&utm_campaign=virtual+camera
Stability AI
Introducing Stable Virtual Camera, currently in research preview. This multi-view diffusion model transforms 2D images into immersive 3D videos with realistic depth and perspective—without complex reconstruction or scene-specific optimization.
AGitHub
Stable Virtual Camera: Generative View Synthesis with Diffusion Models - Stability-AI/stable-virtual-camera
Fyep. select cpu to load
Iwith the wan 2.1 zip file you provide is triton also included or do i need to install it manually? also is teacache using more vram?
Fteacache shouldnt impact vram
T@Furkan Gözükara SECourses have you seen a xformers fix for blackwell yet? I searched but haven't seen progress. I saw someone compiled wheels for python 3.12 but not 3.10, was thinking to upgrade venv for kohya to 312. Also for wan21. lora training, you must use captions, or the lora model can't do anything, no flexibility beyond the source images similarity. I tried at least 4 tests using musubi. Also still struggling to get good likeness as compared to hunyuan
