Can't get good quality wan 2.1 videos

[KintCark]

What is the best setting for wan 2.1 I can't get animations to good visual it's just mush and artifacts I even tried causvid but that didn't work either idk how to use wan 2.1 it's the only one that works on my limited ram

[MrSnichovitch]

What are your system specs? CPU/System RAM/GPU +VRAM?

[KintCark]

What are your system specs? CPU/System RAM/GPU +VRAM?

Cpu snapdragon 865

Ram 10602MB

Gpu adreno 650

Vram N/A

[MrSnichovitch]

Can you supply a sample prompt you're attempting to use? If you're not using the --diffusion-fa option, it might be necessary as it seems to be with Vulkan and ROCm, but I can't specifically say for sure since I don't have an ARM-based/Adreno GPU system to test with.

Here's an example prompt I've run on Vulkan using the standard WAN2.1 T2V model & VAE, but with the Q5_K_M text encoder to save RAM. Process used ~500 MiB of system RAM when running, but I'll warn you that the VAE process used up ~ 5.4 GiB of VRAM at peak, so limited memory could still pose a problem.

./sd -M vid_gen 
--diffusion-model models/checkpoints/wan2.1_t2v_1.3B_fp16.safetensors 
--vae models/vae/wan_2.1_vae.safetensors 
--t5xxl models/text_encoders/umt5-xxl-encoder-Q5_K_M.gguf 
--lora-model-dir models/loras --embd-dir models/embeddings 
-s -1 --cfg-scale 5 --steps 20 
--sampling-method euler --scheduler simple 
-W 384 -H 216 
--fps 12 --video-frames 85 
-v 
-p "medium shot, a lovely cat in a carpeted room, turning to face the camera and walking toward it" 
-n "anime, cartoon, drawing, 3d render, cgi render, ai generated, ugly face" 
-o ../VidGenOutputs/wan2.1_t2v_1.3B_fp16_2025-10-20_test02.avi --diffusion-fa

Resulting vid:

wan2.1_t2v_1.3B_fp16_2025-10-20_test02.mp4

[stduhpf]

I mean I don't think you should expect amazing quality out of the 1.3B model

[MrSnichovitch]

@stduhpf You're very right... you shouldn't expect miracles. But you should at least be able to get viable clips from it, which is what my particular example clip did. It followed the prompt and produced what was intended.

[KintCark]

@stduhpf You're very right... you shouldn't expect miracles. But you should at least be able to get viable clips from it, which is what my particular example clip did. It followed the prompt and produced what was intended.

My problem is I don't want to wait for so long just to get a unfinished video my frame rate is too low u did 85 frames and I forgot how many steps but yeah we need fastwan support it only requires 3 steps so we wouldn't need to wait so long

[KintCark]

Can you supply a sample prompt you're attempting to use? If you're not using the --diffusion-fa option, it might be necessary as it seems to be with Vulkan and ROCm, but I can't specifically say for sure since I don't have an ARM-based/Adreno GPU system to test with.

Here's an example prompt I've run on Vulkan using the standard WAN2.1 T2V model & VAE, but with the Q5_K_M text encoder to save RAM. Process used ~500 MiB of system RAM when running, but I'll warn you that the VAE process used up ~ 5.4 GiB of VRAM at peak, so limited memory could still pose a problem.

./sd -M vid_gen 
--diffusion-model models/checkpoints/wan2.1_t2v_1.3B_fp16.safetensors 
--vae models/vae/wan_2.1_vae.safetensors 
--t5xxl models/text_encoders/umt5-xxl-encoder-Q5_K_M.gguf 
--lora-model-dir models/loras --embd-dir models/embeddings 
-s -1 --cfg-scale 5 --steps 20 
--sampling-method euler --scheduler simple 
-W 384 -H 216 
--fps 12 --video-frames 85 
-v 
-p "medium shot, a lovely cat in a carpeted room, turning to face the camera and walking toward it" 
-n "anime, cartoon, drawing, 3d render, cgi render, ai generated, ugly face" 
-o ../VidGenOutputs/wan2.1_t2v_1.3B_fp16_2025-10-20_test02.avi --diffusion-fa

Resulting vid:

wan2.1_t2v_1.3B_fp16_2025-10-20_test02.mp4

How you able to use the wan 2.1 1.3B fp16 model with utm5xxl q5_0 I have 10gb ram with 8gb free u tried wab 2.1 1 3B Q8 and umt5 Q5 but it got oom and termux crashed

[MrSnichovitch]

How you able to use the wan 2.1 1.3B fp16 model with utm5xxl q5_0 I have 10gb ram with 8gb free u tried wab 2.1 1 3B Q8 and umt5 Q5 but it got oom and termux crashed

I mentioned previously that I don't have an ARM-based/Adreno GPU system to test with, so I'm using desktop hardware with more RAM and a GPU with dedicated VRAM to work with. There's no real way for me to give you an apples-to-apples comparison using the Linux system I have vs. the Android system you're using.

When stable-diffusion.cpp runs, it loads/buffers the model, VAE, and text encoder tensors into RAM before processing begins in earnest, meaning that even if you were to use the smallest quants available, such as these:

wan2.1_t2v_1.3b-q2_k.gguf		552.4 MiB
wan_2.1_vae.safetensors			242.1 MiB
umt5-xxl-encoder-Q3_K_M.gguf	2.8 GiB
					            3.62 GiB total

...you'd still be eating up 3.62 GiB of your available 8 GiB RAM before the compute stages start, which consumes RAM/VRAM on top of that.

I've tested those small quants on my system using both the Vulkan and ROCm backends, and I can't get them to produce anything besides blurry shapes. It should be noted that WAN was coded to use CUDA on NVidia hardware, and with it's poor performance in Vulkan and ROCm, it definitely shows. I have no idea if the small quants function properly on NVidia since I have no hardware to test with.

I don't want to be discouraging, but ultimately, you may be tilting at windmills trying to get WAN to generate anything usable with the limited resources you have. Even the full fp16 version of WAN2.1 1.3B doesn't run under Vulkan all that well on my system (ROCm is consistently better), and I still need the fp16 version to get anything usable. I'm afraid there's no information I can give you that might help.

[KintCark]

How you able to use the wan 2.1 1.3B fp16 model with utm5xxl q5_0 I have 10gb ram with 8gb free u tried wab 2.1 1 3B Q8 and umt5 Q5 but it got oom and termux crashed

I mentioned previously that I don't have an ARM-based/Adreno GPU system to test with, so I'm using desktop hardware with more RAM and a GPU with dedicated VRAM to work with. There's no real way for me to give you an apples-to-apples comparison using the Linux system I have vs. the Android system you're using.

When stable-diffusion.cpp runs, it loads/buffers the model, VAE, and text encoder tensors into RAM before processing begins in earnest, meaning that even if you were to use the smallest quants available, such as these:

wan2.1_t2v_1.3b-q2_k.gguf		552.4 MiB
wan_2.1_vae.safetensors			242.1 MiB
umt5-xxl-encoder-Q3_K_M.gguf	2.8 GiB
					            3.62 GiB total

...you'd still be eating up 3.62 GiB of your available 8 GiB RAM before the compute stages start, which consumes RAM/VRAM on top of that.

I've tested those small quants on my system using both the Vulkan and ROCm backends, and I can't get them to produce anything besides blurry shapes. It should be noted that WAN was coded to use CUDA on NVidia hardware, and with it's poor performance in Vulkan and ROCm, it definitely shows. I have no idea if the small quants function properly on NVidia since I have no hardware to test with.

I don't want to be discouraging, but ultimately, you may be tilting at windmills trying to get WAN to generate anything usable with the limited resources you have. Even the full fp16 version of WAN2.1 1.3B doesn't run under Vulkan all that well on my system (ROCm is consistently better), and I still need the fp16 version to get anything usable. I'm afraid there's no information I can give you that might help.

Well I managed to get wan Q5KS umt5-xxl-encoder-Q4_K_S.gguf wan fp8 but i can't manage to get results it only worked once. I try using self forcing causvid and cfg distilled loras but still can't get good results.