Software Engineering Courses (SECourses)•10mo ago

MIDI: Multi-Instance Diffusion for Single Image to 3D Scene Generation https://huanngzh.github.io

MIDI: Multi-Instance Diffusion for Single Image to 3D Scene Generation https://huanngzh.github.io/MIDI-Page/

MIDI: Multi-Instance Diffusion for Single Image to 3D Scene ...

MIDI is a novel paradigm for compositional 3D scene generation from a single image, extending pre-trained 3D object generation models to multi-instance diffusion models for simultaneous generation of multiple 3D instances.

theJindesan•3/17/25, 6:46 PM

@Furkan Gözükara SECourses Are you seeing the full 32GB in windows from your 5090? Mine only reports 31.5GB and its the difference between fitting the full Wan Model or not.. My 3090 reports 24.0 - in the same task manager so I'm thinking its not the 1024 calculation.

TtheJindesan <@205854764540362752> Are you seeing the full 32GB in windows from your 5090? M...

Weatherby•3/17/25, 8:19 PM

Which is your primary gpu connected to your display?

WWeatherby Which is your primary gpu connected to your display?

theJindesan•3/17/25, 9:18 PM

Onboard Intel. Both headless. 3090 shows the full 24.0

theJindesan•3/17/25, 9:19 PM

theJindesan•3/17/25, 9:20 PM

So thats why i was curious why the total was different.

Aart-ai_usa-girl MIDI: Multi-Instance Diffusion for Single Image to 3D Scene Generation https:...

Furkan Gözükara SECourses•3/17/25, 9:43 PM

wow nice

TtheJindesan <@205854764540362752> Are you seeing the full 32GB in windows from your 5090? M...

Furkan Gözükara SECourses•3/17/25, 9:43 PM

same here 31.5

Furkan Gözükara SECourses•3/17/25, 9:44 PM

they stolen vram

FFurkan Gözükara SECourses same here 31.5

theJindesan•3/17/25, 10:48 PM

Yeah 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.

TtheJindesan Yeah something different between 30xx/50xx. My 3090 = 24.0 in task manager, 5090...

Furkan Gözükara SECourses•3/17/25, 10:52 PM

true

Furkan Gözükara SECourses•3/17/25, 10:52 PM

shameless nvidia

Furkan Gözükara SECourses•3/17/25, 10:52 PM

but flux training fully fits

theJindesan•3/17/25, 10:54 PM

@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.

TtheJindesan <@205854764540362752> have you found perfect teacache settings for Wan2.1? with ...

Furkan Gözükara SECourses•3/17/25, 10:57 PM

0.15 is nice

Furkan Gözükara SECourses•3/17/25, 10:57 PM

but 0.2 made some degrade

Furkan Gözükara SECourses•3/17/25, 10:57 PM

but if working for you great

theJindesan•3/17/25, 10:58 PM

Will 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

Furkan Gözükara SECourses•3/17/25, 10:58 PM

i will first do wan 2.1 lora

FFurkan Gözükara SECourses 0.15 is nice

theJindesan•3/17/25, 10:58 PM

I think this number changes if you select 14B co-efficients. this is also a bit confusing. I also notice torchcompile kills loras

TtheJindesan I think this number changes if you select 14B co-efficients. this is also a bit...

Furkan Gözükara SECourses•3/17/25, 10:59 PM

0.15 only changes with 1.3b model

FFurkan Gözükara SECourses 0.15 only changes with 1.3b model

theJindesan•3/17/25, 11:00 PM

so you like 0.15 for I2V and T2V 14B right? I didn't buy 5090 to run 1.3B models

FFurkan Gözükara SECourses i will first do wan 2.1 lora

theJindesan•3/17/25, 11:03 PM

will 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.

TtheJindesan so you like 0.15 for I2V and T2V 14B right? I didn't buy 5090 to run 1.3B models...

Furkan Gözükara SECourses•3/17/25, 11:04 PM

ytes 0.15 for all 14b model

TtheJindesan will you use musubi-tuner? I just can't get good lora's for wan with the same d...

Furkan Gözükara SECourses•3/17/25, 11:04 PM

yes i plan to use musubi

Furkan Gözükara SECourses•3/17/25, 11:04 PM

lots of people using lots of bug fixes

FFurkan Gözükara SECourses lots of people using lots of bug fixes

theJindesan•3/17/25, 11:07 PM

Maybe 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

TtheJindesan Maybe I need to try again with later code. Also this guy just created a gui for...

Furkan Gözükara SECourses•3/17/25, 11:08 PM

i plan to make my own gradio

FFurkan Gözükara SECourses i plan to make my own gradio

theJindesan•3/17/25, 11:12 PM

I 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

Furkan Gözükara SECourses•3/17/25, 11:13 PM

sure. i am testing 0.2 teacache right now

Furkan Gözükara SECourses•3/17/25, 11:13 PM

it takes less than 10 min on h100 highest quality - 20 steps

theJindesan•3/17/25, 11:15 PM

Also 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.

TtheJindesan Also TorchCompileModelWan (for wan21) breaks Lora's on 50xx but not 30xx. Must...

Furkan Gözükara SECourses•3/17/25, 11:15 PM

very likely

Furkan Gözükara SECourses•3/17/25, 11:16 PM

wow teacache 0.2 not bad

Furkan Gözükara SECourses•3/17/25, 11:16 PM

and super fast

theJindesan•3/17/25, 11:16 PM

Its hard to find people with direct 5090 experience still. Shortages are tough.

theJindesan•3/17/25, 11:17 PM

I was using 0.16 then 0.2 - but i was reading another thread where they reported 0.25 being the perfect spot

theJindesan•3/17/25, 11:17 PM

Not 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

TtheJindesan Click to see attachment

Weatherby•3/18/25, 12:36 AM

That is odd...

Wolfparty7•3/18/25, 4:35 PM

Hello everyone

Wolfparty7•3/18/25, 4:35 PM

Can somone hel me to solve this?

Wolfparty7•3/18/25, 4:35 PM

Im trying to merge two models in SD

Wolfparty7•3/18/25, 4:35 PM

but im getting this error

Wolfparty7•3/18/25, 4:36 PM

Wolfparty7•3/18/25, 4:36 PM

Its a VRam problem?

art-ai_usa-girlOP•3/18/25, 8:48 PM

stable-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: Multi-View Video Generation with...

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.