Wan2.2 14B on GTX1050 with 4Gb : ok.
Latest ComfyUI versions are wonderful in memory management :
I own an old GTX1050Ti with 4Gb VRAM, in an even older computer with 24Gb RAM. I've been using LTXV13B-distilled since august, creating short image to video 3s 768×768 clips with various results on characters. Well rendered bodies on slow movements. But often awful faces. It was slower on lower resolutions, with worst quality.
I tend not to update a working solution, and at the time, Wan models were totally out of reach, hiting 00M error or crashing during the VAE decoding at the end.
But lately, I updated ComfyUI. I wanted to give another try to Wan.
• Wan2.1 Vace 1.3 — failed (ran but results unrelated to initial picture)
• Wan2.2 5B — awful ;
And...
• Wan2.2 14B — worked... !!!
How ?
1) Q4KM quantization on both low noise and high noise models) ;
2) 4 steps Lightning Lora ;
3) 480×480, length 25, 16 fps (ok, that's really small) ;
4) Wan2.1 VAE decoder.
That very same workflow didn't work on older ComfyUI version.
Only problem: it takes 31 minutes and uses a huge amount of RAM.
Tested on Fedora 42.
https://redd.it/1okn7z1
@rStableDiffusion
Latest ComfyUI versions are wonderful in memory management :
I own an old GTX1050Ti with 4Gb VRAM, in an even older computer with 24Gb RAM. I've been using LTXV13B-distilled since august, creating short image to video 3s 768×768 clips with various results on characters. Well rendered bodies on slow movements. But often awful faces. It was slower on lower resolutions, with worst quality.
I tend not to update a working solution, and at the time, Wan models were totally out of reach, hiting 00M error or crashing during the VAE decoding at the end.
But lately, I updated ComfyUI. I wanted to give another try to Wan.
• Wan2.1 Vace 1.3 — failed (ran but results unrelated to initial picture)
• Wan2.2 5B — awful ;
And...
• Wan2.2 14B — worked... !!!
How ?
1) Q4KM quantization on both low noise and high noise models) ;
2) 4 steps Lightning Lora ;
3) 480×480, length 25, 16 fps (ok, that's really small) ;
4) Wan2.1 VAE decoder.
That very same workflow didn't work on older ComfyUI version.
Only problem: it takes 31 minutes and uses a huge amount of RAM.
Tested on Fedora 42.
https://redd.it/1okn7z1
@rStableDiffusion
Reddit
From the StableDiffusion community on Reddit
Explore this post and more from the StableDiffusion community
Tencent SongBloom music generator updated model just dropped. Music + Lyrics, 4min songs.
[https://github.com/tencent-ailab/SongBloom](https://github.com/tencent-ailab/SongBloom)
* **Oct 2025**: Release songbloom\_full\_240s; fix bugs in half-precision inference ; Reduce GPU memory consumption during the VAE stage.
https://redd.it/1okpsj4
@rStableDiffusion
[https://github.com/tencent-ailab/SongBloom](https://github.com/tencent-ailab/SongBloom)
* **Oct 2025**: Release songbloom\_full\_240s; fix bugs in half-precision inference ; Reduce GPU memory consumption during the VAE stage.
https://redd.it/1okpsj4
@rStableDiffusion
GitHub
GitHub - tencent-ailab/SongBloom: The official code repository for SongBloom: Coherent Song Generation via Interleaved Autoregressive…
The official code repository for SongBloom: Coherent Song Generation via Interleaved Autoregressive Sketching and Diffusion Refinement - tencent-ailab/SongBloom
Which do you think are the best SDXL models for anime? Should I use the newest models when searching, or the highest rated/downloaded ones, or the oldest ones?
https://redd.it/1okrk54
@rStableDiffusion
https://redd.it/1okrk54
@rStableDiffusion
This media is not supported in your browser
VIEW IN TELEGRAM
I'm trying out an amazing open-source video upscaler called FlashVSR
https://redd.it/1oksrsm
@rStableDiffusion
https://redd.it/1oksrsm
@rStableDiffusion