Wan 2.1 Works:)

[KintCark]

I tried wan 2.1 and it runs fast at 256x480 2:32 minutes a step I tried using causvid lora but I thought I only needed 4 steps but I think it was my image resolution but it is promising adding openblas will make the generation faster than just opencl and cpu or is cpu already openblas? I can't wait for you to update:) ik 2minutes per step is maybe slow but at least it not 10minutes per step it same speed as comfyui animatediff proccess at 16 batch frames:)

Also add vae on cpu, clip on cpu, and cpu offload features it will help with memory issues

[KintCark]

Wan 2.1 is super fast now I can do 9-12 frames at 56-76 s/it and max resolution at 576x320 on 12 gb Ram I can't wait for your next update also self forcing lora speeds up generation also use it with causvid for great videos

[rene-descartes2021]

@KintCark I'm right there with you, 6 GB RAM here though. I saw some of your your posts on this in leejet/stable-diffusion.cpp.

You got Wan2.1 working in this Local-Diffusion app? Or in Termux with sd.cpp?

Are you using tassd (taehv) instead of wan 2.1's vae? From the linked conversation sounds like you have not tried this yet. Looks like vae's peak memory use will drop from ~5.4GB to ~0.5GB, possibly less with additional config. You may in turn increase other parameters to get better results with Wan2.1.

Also I found a self-forcing LoRA extracted by Kijai, they wrote about it somewhere. I think it may use less memory than causvid or what self-forcing team made, maybe Kijai wrote this but I can't find their comment again, maybe reddit or huggingface somewhere.
I use: Wan2_1_self_forcing_sid_v2_1_3B_lora_rank_32_fp16.safetensors

Well share your results, thanks.

[KintCark]

@KintCark I'm right there with you, 6 GB RAM here though. I saw some of your your posts on this in leejet/stable-diffusion.cpp.

You got Wan2.1 working in this Local-Diffusion app? Or in Termux with sd.cpp?

Are you using tassd (taehv) instead of wan 2.1's vae? From the linked conversation sounds like you have not tried this yet. Looks like vae's peak memory use will drop from ~5.4GB to ~0.5GB, possibly less with additional config. You may in turn increase other parameters to get better results with Wan2.1.

Also I found a self-forcing LoRA extracted by Kijai, they wrote about it somewhere. I think it may use less memory than causvid or what self-forcing team made, maybe Kijai wrote this but I can't find their comment again, maybe reddit or huggingface somewhere.
I use: Wan2_1_self_forcing_sid_v2_1_3B_lora_rank_32_fp16.safetensors

Well share your results, thanks.

No I use gguf but I'm upgraded to a 16 GB ram phone so I haven't tested it on my new phone yet but with12gb ran u gotta delete bloat ware apps and disable system apps like Google Play store etc to get free more memory for ai I'm using comfyui right now and wan 1.3b is super fast at w-192 h-320 or opposite with causvid Lora at 1.5 weight and self forcing Lora at 1.0 weight I couldn't get anything out of it on my old phone you need more ram to run wan 8?12gb only is really 10gb ram on termux so you with 6gb ram u probably only got 4-5gb ram u need new phone as 6gb isn't sefice for wan I really had a hard time with wan in Sdcpp on 10 GB ram

[KintCark]

@KintCark I'm right there with you, 6 GB RAM here though. I saw some of your your posts on this in leejet/stable-diffusion.cpp.

You got Wan2.1 working in this Local-Diffusion app? Or in Termux with sd.cpp?

Are you using tassd (taehv) instead of wan 2.1's vae? From the linked conversation sounds like you have not tried this yet. Looks like vae's peak memory use will drop from ~5.4GB to ~0.5GB, possibly less with additional config. You may in turn increase other parameters to get better results with Wan2.1.

Also I found a self-forcing LoRA extracted by Kijai, they wrote about it somewhere. I think it may use less memory than causvid or what self-forcing team made, maybe Kijai wrote this but I can't find their comment again, maybe reddit or huggingface somewhere.
I use: Wan2_1_self_forcing_sid_v2_1_3B_lora_rank_32_fp16.safetensors

Well share your results, thanks.

Btw yes I use termux on Android

[rene-descartes2021]

Oh gotcha thanks for the clarification. For comfyUI I have a headless minipc I can run the comfyui server on and use the phone's browser to run it, that way I can get lots of RAM and still be mobile. Headless means no monitor or keyboard or anything, it just boots Linux and I ssh into it. If you know Linux, then this is doable to be able to do this stable diffusion stuff locally and while mobile. So I do this instead of buying better phone.

For the phone I found nitrosd-realism (SDXL) works well, I saw the author of Local-Diffusion filed an issue about getting support for it into sdcpp, and it was added, but the author hasn't updated this app to have a sliding scale for the timeshift-step parameter. I use nitrosd-realism_q6_K.gguf and it barely fits into this 6GB RAM Pixel 4. Works well, a 1 step model.

You're right about force stopping the Google Play store. Another tip would be to enable the developer options, and in the developer options click "Standby Apps" and get most of them off the 'Active' setting, which keeps them in RAM so they load fast. I also set a lot of these to 'Restricted' which I don't ever want loaded in the background unless I manually launch the app.